# Oxford Agentic - full site text > Oxford Agentic runs in-person, hands-on AI training in Oxford for business leaders, across three one-day bootcamps, each in a small cohort. Build and run a working agentic AI workflow on a real task from your own work, so the routine runs itself and the hours go back into the work only you can do. Build a second brain: the structured store of your own context that AI tools draw on. Or ship a live website and a working prototype, with Claude Code and Codex doing the typing. No day assumes you arrive able to code, and nothing is hand-waved. Location: Oxford, Oxfordshire, United Kingdom. Affiliation: independent - no affiliation with the University of Oxford, its colleges, or any other institution. Website: https://oxfordagentic.com/ Contact: hello@oxfordagentic.com Generated: 2026-11-21 --- ## Frequently asked questions ### What is Oxford Agentic? Oxford Agentic runs in-person, hands-on AI training in Oxford for business leaders, across three one-day bootcamps. The Oxford Agentic Bootcamp is where you build and run a working agentic AI workflow on a real task from your own work. The Oxford Second Brain Bootcamp is where you build the structured store of your own context that AI tools draw on. The Oxford Agentic Coding Bootcamp is where you ship software you own outright, with Claude Code and Codex doing the typing. All three are hands-on, and none of them expects you to arrive able to code. Nothing is hand-waved either. ### Who are the bootcamps for? Senior people who already use AI daily and have hit the ceiling: good output from a chat window, and still doing the work by hand tomorrow. Founders, directors and C-suite in Oxford and London; senior NHS, university and public-sector staff; consultants and independent operators; product and operations leaders. Engineers and people who build AI for a living are welcome too, and several came to the first cohort. The Oxford Agentic Bootcamp is for building working AI workflows; the Oxford Second Brain Bootcamp is for building the store of your own context those tools draw on; the Oxford Agentic Coding Bootcamp is for shipping software without waiting on a developer. ### Do I need to be technical? No, and that is not the only thing worth saying about it. Nothing on the day requires code, everything is facilitated in the room, and the cohort is kept small so nobody spends the day stuck on setup. Nothing is hand-waved either: part of the day goes on where these workflows break, how you would notice, and what you constrain so a bad run cannot do damage. Engineers are welcome, and some of the room usually builds AI for a living. ### Who is it not for? People who want a conceptual overview of AI - this is a build day, not a briefing. Anyone who cannot bring a real task from their own work, because the whole day runs against one and without it you get a demonstration instead of an outcome. Anyone expecting a vendor or platform pitch, because no tool is being sold. And anyone who cannot install software on their own laptop: corporate and government devices often block tool sign-ups, so it is worth checking before you book rather than at 09:30 on the day. ### What do I need to bring, and do I need to prepare? A laptop, and the raw material the day builds on: at the Oxford Agentic Bootcamp one real task from your own work that currently costs you meaningful time; at the Oxford Second Brain Bootcamp your own documents, notes and examples; at the Oxford Agentic Coding Bootcamp an idea you want to see running. Bringing something genuine matters more than any preparation, and everything is facilitated in the room. The one thing worth checking before you book is whether you can install software on your laptop, since corporate and government devices often block new tool sign-ups. At the Agentic Bootcamp you will get the most from the day with a paid plan such as ChatGPT Plus or Claude Pro, but it is not essential. ### Where in Oxford are the events held? In Oxford colleges and museums. Cohort 1 of the Oxford Agentic Bootcamp ran at St Anne’s College on 21 July 2026, in a seminar room laid out cabaret-style for small-group work. The Agentic Bootcamp’s Cohorts 2, 3 and 4 and the Coding Bootcamp’s first cohort are at Worcester College on Walton Street, in central Oxford; Cohort 5 moves to Jesus College on Turl Street for 28 January 2027, where the room booking is being finalised; the Second Brain Bootcamp’s first cohort was at the Pitt Rivers Museum on South Parks Road. Each event page confirms its own venue and date. ### Which organisations is Oxford Agentic affiliated to? Oxford Agentic currently operates as an independent organisation. ### What will I actually leave with? Something running, built during the day from your own material. At the Oxford Agentic Bootcamp that is a working agentic AI workflow against a real task from your own work, plus the pattern behind it - scout, drafter, reviewer - so you can build the next one yourself. At the Oxford Second Brain Bootcamp it is a populated second brain your AI tools can draw on, with the architecture and maintenance routine to keep it useful. At the Oxford Agentic Coding Bootcamp it is a personal website live on the internet and a working prototype that saves real data, both in your own repository. ### Is one day really enough? It is enough for one thing done properly, which is what the day is scoped to. It is not a survey of the field: no tour of the AI landscape, no model comparisons, no governance policy. You bring one real task, you build against it in facilitated cycles with someone on hand when you get stuck, and you leave with it running plus the pattern underneath it, which is the part that transfers to the next task. The alternative is not a better day. It is a longer course you attend and then never apply. ### Can I bring my team? Yes. Teams get more out of it than individuals do, and if two of you come, bring the same task. Email hello@oxfordagentic.com and we will arrange it. ### What happens after the day? You leave with the group. Everyone in the room has built the same thing, they stay in touch afterwards, and you hear first when the next one is on - reunions and build sessions, editions for particular functions, and the free meetups. Some are paid and some are free. It is an upside rather than a reason to come. ### What if I book but cannot make the date? Tickets are not refundable, but your place transfers to a future cohort at no charge. A date that no longer works should not mean a lost seat - just let us know and we will move you. ### How much does it cost? Pricing is shown on the booking page for each event, along with any early-bird rate. Open the event you are interested in and the booking form sits on that page. If your organisation would rather pay by invoice, email hello@oxfordagentic.com and we will arrange it. --- ## Programme: The Oxford Agentic Bootcamp URL: https://oxfordagentic.com/bootcamps/oxford-agentic-bootcamp/ Last updated: 15 November 2026 The Oxford Agentic Bootcamp is a one-day, in-person bootcamp in Oxford for business leaders. Attendees bring one real task from their own work and leave having built a working agentic AI workflow against it, so the routine runs itself and the hours go back into the work only they can do. Cohorts are capped small and hands-on, not lecture-led. A full day, in a small room of senior people, spent building rather than listening. You arrive with one real task from your own work - the thing that currently eats a morning every week - and you spend the day turning it into a working agentic pipeline you can actually run. What you leave with: - A working agentic pipeline built against your own real task, running by the end of the day - The pattern behind it - scout, drafter, reviewer - and how to apply it to the next task - A clear sense of where the human judgement stays, and where the grunt work goes - A room of senior peers working on the same problem, and their contact details Who it is for: - Founders, directors and C-suite who already believe in AI and want to build with it - Graduates of Oxford AI executive education looking for the hands-on follow-through - Consultants and senior public-sector, NHS and university staff - Anyone comfortable with AI who has never orchestrated several steps into a single workflow - Engineers and people building AI for a living, who come for the pattern and the room ### What is the Oxford Agentic Bootcamp? A one-day, in-person, hands-on course in Oxford where business leaders build and run a working agentic AI workflow on a real task from their own work. Cohorts are small and capped, and no coding is required. ### Do I need to be technical? No, and that is not the whole answer. If you can use a spreadsheet and have used a chatbot you have enough, everything is facilitated in the room, and the point of the small cohort is that nobody gets left stuck on tooling. Nothing is hand-waved either: part of the day goes on where these workflows break, how you would notice and what you constrain so a bad run cannot do damage. Engineers are welcome - several people who build AI for a living came to the first cohort. ### I already use ChatGPT every day. What does the day add? The step up from a chatbot to a workflow. Most people use AI one question at a time - one prompt, one tab, with you carrying the work between the steps. Agentic means several narrow AI steps handed off to each other in a pipeline: a scout that gathers, a drafter that produces, a reviewer that checks. You spend the day building one of those against a real task from your own work, and you leave with it running. Daily chatbot use is the right starting point for that rather than a reason to skip the day, and this is a build day rather than a briefing. ### Who is it not for? People who want a conceptual overview of AI - this is a build day, not a briefing. Anyone who cannot bring a real task from their own work, because the whole day runs against one and without it you get a demonstration instead of an outcome. Anyone expecting a vendor or platform pitch, because no tool is being sold. And anyone whose work laptop blocks new sign-ups, which is worth checking first. ### Which of the three bootcamps should I do first? Start from the ceiling you have hit. If a task you already do by hand eats a morning every week, this is the day: you leave with a workflow running against it. If your AI answers come back generic because the tools know nothing about your business, the Oxford Second Brain Bootcamp builds that context from your own material. If it is an idea you want to see running rather than described, the Oxford Agentic Coding Bootcamp ships a website and a working prototype you own. They compound, because a workflow gets better the moment it can read your second brain, so pick the one that matches the problem in front of you now. ### What does “agentic” actually mean here? Most people use AI one question at a time - one prompt, one tab. Agentic means several narrow AI steps handed off to each other in a pipeline: a scout that gathers, a drafter that produces, a reviewer that checks. The shift is from prompting to orchestrating. ### Does the day cover checking the work an agent produces? Yes. The reviewer is one of the three steps you build - a scout that gathers, a drafter that produces, a reviewer that checks against a standard you write down - so the checking is built into the workflow rather than left until the end. Part of the day goes on where these workflows break, how you would notice and what you constrain so a bad run cannot do damage. You leave with a clear sense of where the human judgement stays, and where the grunt work goes. ### What do I need to bring? One real task from your own work that currently costs you meaningful time, plus a laptop. The task is what you build against all day, so bringing a genuine one matters more than any preparation. ### How do I choose which task to bring? Pick something you repeat rather than something big: a workflow pays back every time it runs, so the thing that currently eats a morning every week beats the hardest problem in your business. The raw material should be words and documents rather than a system only you can reach, you should be able to look at the output and know at a glance whether it is right, and getting it wrong should cost you a correction rather than a client. Some work is worth leaving alone: the writing where doing it is the thinking, anything you only do twice a year, and anything a person has to stand behind. ### Do I need a written-down process before I can automate a task? No. The day is built around three build cycles rather than three talks, so you arrive with the task and the shape of the workflow comes out of building it. Bringing a genuine task matters more than any preparation. One thing does get written down, and it is narrower than a process: the reviewer you build checks against a standard you write down, so that standard is part of what you make on the day. Everything is facilitated in the room, so nobody is left working out the steps on their own. ### What if my task is confidential? You choose the task, so a useful test is to bring one you would be comfortable describing to an external supplier, and leave the rest out. Credentials, keys and other people’s personal data do not belong in an AI workflow in any form. You build on your own machine, and the day is not recorded. That still leaves most of the tasks that cost you a morning every week. ### Can I install software on my work laptop? Worth checking before you book. Corporate and government devices often block new tool sign-ups, and the day is spent building on your own machine. This caught people out in Cohort 1, so it is better found now than at 09:30 on the day. If your laptop is locked down, email hello@oxfordagentic.com and we will work out what you need beforehand. ### Where in Oxford is it held? In central Oxford, and which college depends on the cohort. Cohorts 2, 3 and 4 are all at Worcester College on Walton Street - 16 September, 19 November and 3 December 2026 - and Cohort 5 moves to Jesus College on Turl Street for Thursday 28 January 2027, where the room booking is being finalised. Cohort 1 ran at St Anne’s College in July 2026, in a seminar room laid out cabaret-style for small-group work, so an older listing may still show that address. Each event page confirms its own venue and date, and that is the page to go by. ### What time does the day start and finish? Registration and coffee are from 09:00 and the day finishes at 17:00. The last block is the wrap and networking rather than new material, so check both ends against your travel before you book. Each event page confirms its own date and times, and that is the page to go by. ### How many people are in the room? Not many, and it is capped. The size is deliberate: the day is facilitated, not lectured, which stops working in a large room. Cohort 1 ran with sixteen, and the places left in the next one are shown on its booking page. ### Do I need a paid AI subscription? Bring a laptop. You will get the most from the day with a paid plan such as ChatGPT Plus or Claude Pro, but it is not essential - we make sure everyone can take part either way. ### How do I justify this to my manager? Some people pay for themselves, some expense it, and occasionally a company buys two or three seats. If you need to put it in an email, this is the version that survives being forwarded: a day out to build one working agentic workflow on a task we already do by hand every week, and come back able to build the next one without help. Pricing sits on each event’s booking page, so the numbers are easy to attach. ### Can I bring my team, or book for several people? Yes. Teams get more out of it than individuals do, and if two of you come, bring the same task so the same workflow goes back to work with both of you. Email hello@oxfordagentic.com and we will arrange it. ### Can my organisation pay by invoice? Yes. If your organisation would rather pay by invoice than book on a card form, email hello@oxfordagentic.com and we will arrange it. ### What if I book but cannot make the date? Tickets are not refundable, but your place transfers to a future cohort at no charge. A date that no longer works should not mean a lost seat - just let us know and we will move you. ### Is the day recorded? No. The value is in being in the room and building, not in a video to watch later. You leave with the workflow you built rather than a recording you might get back to. ### What happens after the day? You leave with the workflow you built rather than a recording, and you leave with the group. Everyone in the room has built the same thing, you go away with their contact details, and people stay in touch afterwards. You also hear first when the next one is on - reunions and build sessions, editions for particular functions, and the free meetups. Some of those are paid and some are free. It is an upside rather than a reason to come. ### Is lunch included? Yes. Lunch and refreshments are provided, so you can spend the day building rather than sorting out food. We ask attendees about dietary and access needs before the day. ### Who runs it? Drummond Gilbert, who tutors in AI in Management at Saïd Business School and runs Oxford’s Practical AI meetup, with Nicolai Thomson (founder, Jenesys AI), Emine Gokce Phillips (co-founder and CEO, Novaria) and Jonathan Waddingham (product consultant). Each runs part of the day, and the day is facilitated rather than lectured, so there are people in the room with you while you build. --- ## Programme: The Oxford Second Brain Bootcamp URL: https://oxfordagentic.com/bootcamps/second-brain-bootcamp/ Last updated: 4 November 2026 The Oxford Second Brain Bootcamp is a one-day, in-person course in Oxford for business leaders who use AI daily and have hit the ceiling of what prompting alone can do. Attendees build a second brain - a structured store of their own context, voice and standards that AI assistants can draw on - and leave with it populated, organised and set up to maintain. Every AI tool you use starts each conversation knowing nothing about your business. The judgement you have accumulated - how you write, what good looks like, which clients matter and why - lives in your head, so every answer comes back generic. A second brain fixes the supply problem: a structured, maintained store of your own context that your AI tools can draw on, built in a day from your own material. What you leave with: - A working second brain, populated with your own material rather than a template - An architecture for organising your context, so it stays findable as it grows - A maintenance routine, so the system improves with use instead of decaying - The connections that let your AI tools actually draw on it, across tools Who it is for: - Business leaders, founders, partners and senior operators who already use AI most days - People who have hit the ceiling of what prompting alone can do - Anyone tired of re-explaining their business to a chatbot that forgets by Tuesday ### What is the Oxford Second Brain Bootcamp? A one-day, in-person course in Oxford where business leaders build a second brain - a structured store of their own context, voice and standards that AI assistants can draw on. You build it from your own material during the day and leave with it working, organised and set up to maintain. ### How is this different from the Oxford Agentic Bootcamp? The Agentic Bootcamp is about building workflows - pipelines of AI steps that do a task. The Second Brain Bootcamp is about building context - the store of your own knowledge those tools draw on. They compound: a workflow gets better the moment it can read your second brain. Start with whichever ceiling you have hit first. ### Do I need to be technical? No. If you use AI tools most days you have everything the day requires. The work is deciding what your context is and structuring it - thinking clearly about your own business, not writing code. ### Who is it not for? People who want a conceptual overview of AI - this is a build day, not a briefing. Anyone who cannot bring their own material, because the day is spent populating a second brain with real documents and notes, and without them you leave with a template rather than a second brain. Anyone expecting a vendor or platform pitch: no tool is being sold, and the day is deliberately not a tutorial in one app. And anyone whose work laptop blocks new sign-ups, which is worth checking first. ### Which AI tools does this work with? The ones you already use. A second brain is your own material, structured and stored so a model can read it. It is not a feature of any one product, and the day is not a tutorial in a particular app: it is spent deciding what your context is, organising it, and then connecting it to the tools you work in. That is deliberate. Tools change, and a second brain built inside one of them has to be rebuilt when you move. ### What do I need to bring? A laptop, and your own material - the documents, notes and examples that carry how you work and what you know. The day is spent populating a second brain with real content rather than a template, so bringing genuine material matters more than any preparation. ### Can I install software on my work laptop? Worth checking before you book. Corporate and government devices often block new tool sign-ups, and the day involves connecting a second brain to the tools you work in, on your own machine. This caught people out at the first Oxford Agentic Bootcamp, so it is better found now than at 09:30 on the day. If your laptop is locked down, email hello@oxfordagentic.com and we will work out what you need beforehand. ### What if my material is confidential? A useful test is to bring what you would be comfortable putting in an email to an external supplier, and leave out the rest. Credentials, keys and other people’s personal data do not belong in a second brain in any form. That still leaves most of what makes AI answers specific to you: how you write, what good looks like, and the decisions you have already made and why. ### Where in Oxford is it held? In central Oxford. Cohort 1 was at the Pitt Rivers Museum on South Parks Road. The exact venue and date are confirmed on each event’s own page. ### What time does the day start and finish? The day runs 09:00 to 17:00, so it is worth checking both ends against your travel before you book. The exact date and times are confirmed on each event’s own page. ### Can I bring my team, or book for several people? Yes. Teams are welcome, and bringing a colleague means the same approach goes back to work with two of you rather than one. Email hello@oxfordagentic.com and we will arrange it. ### Can my organisation pay by invoice? Yes. If your organisation would rather pay by invoice than book on a card form, email hello@oxfordagentic.com and we will arrange it. ### How big is the group? Small, and capped. The exact number of places left is on the booking page. ### What if I book but cannot make the date? Tickets are not refundable, but your place transfers to a future cohort at no charge. A date that no longer works should not mean a lost seat - just let us know and we will move you. ### Who runs it? Drummond Gilbert, who tutors in AI in Management at Saïd Business School and runs Oxford’s Practical AI meetup, with Adrian Shedden (founder, Accelerated Impact Ops), Glenn Smith (founder, GenieOS) and Adam Martin (founding CTO). Each runs part of the day, and all are practitioners who build with these tools daily rather than commentators on them, so it is a build day with people in the room rather than a briefing. --- ## Programme: The Oxford Agentic Coding Bootcamp URL: https://oxfordagentic.com/bootcamps/agentic-coding-bootcamp/ Last updated: 19 November 2026 The Oxford Agentic Coding Bootcamp is a one-day, in-person course in Oxford for people who understand the concepts but have not yet built and shipped something of their own. Claude Code and Codex do the typing; you learn the workflow around them, from the brief to version control, deployment and testing, and leave with two things live that you own outright. Most people who use AI every day have never shipped software with it. The tools can now write the code, but the workflow around them - the brief the agent reads first, the repository, the undo button, the deploy, the check that what it wrote is right - is what stands between a demo and something live. This is a day of sitting next to a senior engineer who does the typing, not a course that turns you into a junior one. What you leave with: - A personal website live on the internet, in your own GitHub repository - A working product prototype that saves real data and has basic tests - A take-home kit of the prompts and commands used on the day - A way to keep costs under control once you are building on your own Who it is for: - People who understand the concepts but have not yet built and shipped something of their own - Founders and operators who want to prototype an idea without waiting on a developer - Anyone who has watched a coding agent work and wants to know what to do around it ### What is the Oxford Agentic Coding Bootcamp? A one-day, in-person course in Oxford on how software actually gets built and shipped with coding agents such as Claude Code and Codex. You work through the whole path, from writing the brief the agent reads to deploying, testing and reviewing what it wrote, and leave with a live website and a working prototype. ### Do I need to be able to code? No, and that is not the whole answer. The day does not assume you can code: the agents do the typing. What it does teach is everything around the code - the terminal, repositories, version control, deployment, databases and tests - and how to check what the AI wrote rather than trusting it. Expect to leave understanding how the pieces fit, not just having watched them. ### I cannot write code. What will I actually have finished by five o’clock? Two things you own outright: a personal website live on the internet, in your own GitHub repository, and a working product prototype that saves real data and has basic tests. You also leave with the prompts and commands used on the day, and a way to keep costs under control once you are building on your own. What the day does not do is turn you into a working developer - the agents do the typing, and what it teaches is the workflow around the code: the brief the agent reads first, the repository, the undo button, the deploy, and the check that what it wrote is right. ### Who is it not for? Anyone who wants to come out of the day a working developer: the agents do the typing, and what the day teaches is the workflow around the code rather than how to write it. Anyone whose problem is a recurring task they already do by hand, because the Oxford Agentic Bootcamp is the day for that one. And anyone who cannot bring a laptop they can install software on, since the day sets up a code editor, a terminal and a GitHub account on your own machine, which is worth checking before you book. ### How is this different from the Oxford Agentic Bootcamp? The Agentic Bootcamp builds workflows: pipelines of AI steps that do a task you already do by hand. The Coding Bootcamp builds software: a website and a product prototype, deployed and tested, in your own repository. If your problem is a recurring task, start with the Agentic Bootcamp. If it is an idea you want to see running, start here. ### How many people are in the room? A small cohort, so the day stays facilitated rather than lectured. Every session is show, then do: a facilitator builds each step in front of the room, then you build it on your own laptop, with facilitators moving between tables so nobody spends the day stuck on setup. A large room would stop that working. ### What should I bring? A laptop you can install software on. The day involves setting up a code editor, a terminal and a GitHub account on your own machine, and corporate or government devices often block that. If your laptop is locked down, email hello@oxfordagentic.com before you book and we will work out what you need. ### Where in Oxford is it held? At Worcester College on Walton Street, in central Oxford, which is also where the Oxford Agentic Bootcamp runs. Cohort 1 is there on Thursday 26 November 2026. Each event page confirms its own venue and date, and that is the page to go by. ### What time does the day start and finish? The day runs 09:00 to 17:00, so it is worth checking both ends against your travel before you book. Each event page confirms its own date and times, and that is the page to go by. ### How do I justify this to my manager? If you need to put it in an email, this is the version that survives being forwarded: a day out to ship two working things - a website live on the internet and a prototype that saves real data - both in your own repository, and to come back able to build the next one with a coding agent rather than waiting on a developer. You also take away the prompts and commands used on the day, and a way to keep costs under control once you are building on your own. Pricing sits on each event’s booking page, so the numbers are easy to attach. ### Can I bring my team, or book for several people? Yes. Bringing a colleague means two of you go back able to build rather than one, and it helps if you arrive with the same idea in mind. Email hello@oxfordagentic.com and we will arrange it. ### Can my organisation pay by invoice? Yes. If your organisation would rather pay by invoice than book on a card form, email hello@oxfordagentic.com and we will arrange it. ### What if I book but cannot make the date? Tickets are not refundable, but your place transfers to a future cohort at no charge. A date that no longer works should not mean a lost seat - just let us know and we will move you. ### Who runs it? Drummond Gilbert, who tutors in AI in Management at Saïd Business School and runs Oxford’s Practical AI meetup, with Jonathan Waddingham (CPO, ex-JustGiving) and Mike Bailey (Coppice Labs). Every session is show, then do, with facilitators in the room. --- ## Event: The Oxford Agentic Coding Bootcamp - Cohort 1 URL: https://oxfordagentic.com/events/oxford-agentic-coding-bootcamp-cohort-1/ Programme: The Oxford Agentic Coding Bootcamp Status: upcoming Date: Thursday, 26 November 2026 Time: 09:00–17:00 (Europe/London) Venue: Worcester College, Walton Street, Oxford, OX1 2HB Cohort 1 of the Oxford Agentic Coding Bootcamp runs on Thursday 26 November 2026 at Worcester College, Oxford. It is a hands-on day for people who understand the concepts but have not yet shipped something of their own: Claude Code and Codex do the typing, and you leave with a live website and a working prototype in your own repository. Who it is for: - People who understand the concepts but have not yet built and shipped something of their own - Founders and operators who want to prototype an idea without waiting on a developer - Anyone who has watched a coding agent work and wants to know what to do around it What you will build: - A personal website live on the internet, in your own GitHub repository - A working product prototype that saves real data, has basic tests and sits on your site - A take-home kit of the prompts and commands used on the day, and how to keep costs down Agenda: - Session 1 Context before code - Writing the brief the agent reads first, and asking for a plan before it builds. - Session 2 Your workbench - The editor, the terminal and repositories, and why they are less scary than they look. - Session 3 Version control as your undo button - Commits, branches and pull requests. - Session 4 Shipping - Deploying to the web, preview links, and going live many times a day. - Session 5 Giving it a memory - A real database, API keys, and what never goes in the repository. - Session 6 Trust but verify - Tests, fact-checking what the AI wrote, and a second model reviewing the code. - Session 7 Testing on a real person - Swapping prototypes and capturing feedback. ### When and where is it? Thursday 26 November 2026, 09:00 to 17:00, at Worcester College on Walton Street in central Oxford. Registration is open - you can book on this page. ### Do I need to be able to code? No, and that is not the whole answer. The agents do the typing, so the day does not assume you can code. It does teach everything around the code - the terminal, version control, deployment, databases and tests - and how to check what the AI wrote rather than trusting it. ### Who is it not for? Anyone who wants to come out of the day a working developer: the agents do the typing, and what the day teaches is the workflow around the code rather than how to write it. Anyone whose problem is a recurring task they already do by hand, because the Oxford Agentic Bootcamp is the day for that one. And anyone who cannot bring a laptop they can install software on, since the day sets up a code editor, a terminal and a GitHub account on your own machine, which is worth checking before you book. ### How does the day run? Every session is show, then do. A facilitator builds each step in front of the room, then you build it on your own laptop with facilitators moving between tables, so nobody spends the day stuck on setup. ### What should I bring? A laptop you can install software on. The day involves setting up a code editor, a terminal and a GitHub account on your own machine, and corporate or government devices often block that. If your laptop is locked down, email hello@oxfordagentic.com before you book and we will work out what you need. ### How do I justify this to my manager? If you need to put it in an email, this is the version that survives being forwarded: a day out to ship two working things - a website live on the internet and a prototype that saves real data - both in your own repository, and to come back able to build the next one with a coding agent rather than waiting on a developer. You also take away the prompts and commands used on the day, and a way to keep costs under control once you are building on your own. Pricing is on this page, so the numbers are easy to attach. ### Can I bring my team? Yes. Bringing a colleague means two of you go back able to build rather than one, and it helps if you arrive with the same idea in mind. Email hello@oxfordagentic.com and we will arrange it. ### What if I cannot make the date? Tickets are not refundable, but your place transfers to a future cohort at no charge. A date that no longer works should not mean a lost seat - just let us know and we will move you. ### Can my organisation pay by invoice? Yes. Pricing and the booking form are on this page. If your organisation would rather pay by invoice than book on the form, email hello@oxfordagentic.com and we will arrange it. ### Who runs it? Drummond Gilbert, who tutors in AI in Management at Saïd Business School and runs Oxford’s Practical AI meetup, with Jonathan Waddingham (CPO, ex-JustGiving) and Mike Bailey (Coppice Labs). --- ## Event: The Oxford Agentic Bootcamp - Cohort 4 URL: https://oxfordagentic.com/events/oxford-agentic-bootcamp-cohort-4/ Programme: The Oxford Agentic Bootcamp Status: upcoming Date: Thursday, 3 December 2026 Time: 09:00–17:00 (Europe/London) Venue: Worcester College, Walton Street, Oxford, OX1 2HB Cohort 4 of the Oxford Agentic Bootcamp runs on Thursday 3 December 2026 at Worcester College, Oxford. You bring one real task from your own work and leave having built a working agentic AI workflow against it, so the routine runs itself and the hours go back into the work only you can do. A small cohort, facilitated rather than lectured. Who it is for: - Founders, directors and C-suite building with AI, not reading about it - Graduates of Oxford AI executive education who want the hands-on follow-through - Consultants, and senior NHS, university and public-sector staff - Anyone who uses AI daily but has never chained several steps together - Engineers who build AI for a living - several came to Cohort 1 What you will build: - A working agentic pipeline running on a real task from your own work - A review step that checks its own output, so it compounds rather than single-shots - A written recipe, so you can rebuild and extend it on Monday Agenda: - 09:00 Registration and coffee - 09:15 Welcome and framing - Why prompting and orchestrating are different skills, and what today is actually for. - 09:35 Anatomy of an agent pipeline - Built live in front of the room, on a real example - scout, drafter, reviewer. - 10:15 Build one - your own task, first pass - Facilitated. Everyone gets to a rough working pass, however untidy. - 11:30 Guest session - A working practitioner on how they use this daily. - 12:15 Build two - add a review step - Push the pipeline past single-shot by giving it something that checks its own work. - 13:00 Lunch - 13:45 Guest session - AI coding and agentic tooling in practice. - 14:15 Build three - the real thing - Apply it properly to the task you came with. This is the one you take home. - 15:30 Showcase and debrief - Everyone runs what they built, and says what they will use it for next week. - 16:00 Making it stick - 16:20 Wrap and networking ### What did people in Cohort 1 actually build? Real tasks from their own work rather than a set exercise, so they varied. The ones that surfaced in the room were lead generation, internal reporting and marketing content. The common shape was a step that gathers, a step that produces and a step that checks, handing off to each other. If your task has that shape, it will fit the day. ### Do I need to be technical? No, and that is not the whole answer. No coding is required and the day is facilitated in the room, so nobody spends it stuck on setup. Nothing is hand-waved either: part of the day goes on where these workflows break, how you would notice and what you constrain. Engineers are welcome, and several people who build AI for a living came to Cohort 1. ### What should I bring? A laptop, and one real task from your own work that currently costs you meaningful time each week. That task is what you build against all day. A paid plan such as ChatGPT Plus or Claude Pro gets you more out of it, but it is not essential and we make sure everyone can take part either way. ### What if my task is confidential? You choose the task, so a useful test is to bring one you would be comfortable describing to an external supplier, and leave the rest out. Credentials, keys and other people’s personal data do not belong in an AI workflow in any form. You build on your own machine, and the day is not recorded. That still leaves most of the tasks that cost you a morning every week. ### Can I install software on my work laptop? Worth checking before you book. Corporate and government devices often block new tool sign-ups, and the day involves setting things up on your own machine. This caught people out in Cohort 1, so it is better found now than at 09:30 on the day. If your laptop is locked down, email hello@oxfordagentic.com and we will work out what you need beforehand. ### How do I justify this to my manager? If you need to put it in an email, this is the version that survives being forwarded: a day out to build one working agentic workflow on a task we already do by hand every week, and come back able to build the next one without help. Pricing is on the booking page, so the numbers are easy to attach. ### Can my organisation pay by invoice? Yes. Pricing and the booking form are on this page. If your organisation would rather pay by invoice than book on the form, email hello@oxfordagentic.com and we will arrange it. ### How big is the cohort? Small, and capped. Cohort 1 ran with sixteen. The day is facilitated rather than lectured, which stops working in a large room. The exact number of places left is on the booking page. ### What happens after the day? You leave with the workflow running, plus the written recipe for rebuilding and extending it on Monday, rather than a recording. You also leave with the group: everyone in the room has built the same thing, people stay in touch afterwards, and you hear first when the next one is on - reunions and build sessions, editions for particular functions, and the free meetups. Some of those are paid and some are free. It is an upside rather than a reason to come. ### Who is it not for? People who want a conceptual overview of AI - this is a build day, not a briefing. Anyone who cannot bring a real task from their own work, because the whole day runs against one and without it you get a demonstration instead of an outcome. And anyone expecting a vendor or platform pitch, because no tool is being sold. ### Who runs it? Drummond Gilbert, who tutors in AI in Management at Saïd Business School and runs Oxford’s Practical AI meetup, with Nicolai Thomson (founder, Jenesys AI), Emine Gokce Phillips (co-founder and CEO, Novaria) and Jonathan Waddingham (product consultant). Each runs part of the day, and the day is facilitated rather than lectured, so there are people in the room with you while you build. --- ## Event: The Oxford Agentic Bootcamp - Cohort 5 URL: https://oxfordagentic.com/events/oxford-agentic-bootcamp-cohort-5/ Programme: The Oxford Agentic Bootcamp Status: upcoming Date: Thursday, 28 January 2027 Time: 09:00–17:00 (Europe/London) Venue: Jesus College, Turl Street, Oxford, OX1 3DW (room booking being finalised) Cohort 5 of the Oxford Agentic Bootcamp runs on Thursday 28 January 2027 at Jesus College, Oxford. You bring one real task from your own work and leave having built a working agentic AI workflow against it, so the routine runs itself and the hours go back into the work only you can do. A small cohort, facilitated rather than lectured. Who it is for: - Founders, directors and C-suite building with AI, not reading about it - Graduates of Oxford AI executive education who want the hands-on follow-through - Consultants, and senior NHS, university and public-sector staff - Anyone who uses AI daily but has never chained several steps together - Engineers who build AI for a living - several came to Cohort 1 What you will build: - A working agentic pipeline running on a real task from your own work - A review step that checks its own output, so it compounds rather than single-shots - A written recipe, so you can rebuild and extend it on Monday Agenda: - 09:00 Registration and coffee - 09:15 Welcome and framing - Why prompting and orchestrating are different skills, and what today is actually for. - 09:35 Anatomy of an agent pipeline - Built live in front of the room, on a real example - scout, drafter, reviewer. - 10:15 Build one - your own task, first pass - Facilitated. Everyone gets to a rough working pass, however untidy. - 11:30 Guest session - A working practitioner on how they use this daily. - 12:15 Build two - add a review step - Push the pipeline past single-shot by giving it something that checks its own work. - 13:00 Lunch - 13:45 Guest session - AI coding and agentic tooling in practice. - 14:15 Build three - the real thing - Apply it properly to the task you came with. This is the one you take home. - 15:30 Showcase and debrief - Everyone runs what they built, and says what they will use it for next week. - 16:00 Making it stick - 16:20 Wrap and networking ### What did people in Cohort 1 actually build? Real tasks from their own work rather than a set exercise, so they varied. The ones that surfaced in the room were lead generation, internal reporting and marketing content. The common shape was a step that gathers, a step that produces and a step that checks, handing off to each other. If your task has that shape, it will fit the day. ### Do I need to be technical? No, and that is not the whole answer. No coding is required and the day is facilitated in the room, so nobody spends it stuck on setup. Nothing is hand-waved either: part of the day goes on where these workflows break, how you would notice and what you constrain. Engineers are welcome, and several people who build AI for a living came to Cohort 1. ### What should I bring? A laptop, and one real task from your own work that currently costs you meaningful time each week. That task is what you build against all day. A paid plan such as ChatGPT Plus or Claude Pro gets you more out of it, but it is not essential and we make sure everyone can take part either way. ### What if my task is confidential? You choose the task, so a useful test is to bring one you would be comfortable describing to an external supplier, and leave the rest out. Credentials, keys and other people’s personal data do not belong in an AI workflow in any form. You build on your own machine, and the day is not recorded. That still leaves most of the tasks that cost you a morning every week. ### Can I install software on my work laptop? Worth checking before you book. Corporate and government devices often block new tool sign-ups, and the day involves setting things up on your own machine. This caught people out in Cohort 1, so it is better found now than at 09:30 on the day. If your laptop is locked down, email hello@oxfordagentic.com and we will work out what you need beforehand. ### How do I justify this to my manager? If you need to put it in an email, this is the version that survives being forwarded: a day out to build one working agentic workflow on a task we already do by hand every week, and come back able to build the next one without help. Pricing is on the booking page, so the numbers are easy to attach. ### Can my organisation pay by invoice? Yes. Pricing and the booking form are on this page. If your organisation would rather pay by invoice than book on the form, email hello@oxfordagentic.com and we will arrange it. ### How big is the cohort? Small, and capped. Cohort 1 ran with sixteen. The day is facilitated rather than lectured, which stops working in a large room. The exact number of places left is on the booking page. ### What happens after the day? You leave with the workflow running, plus the written recipe for rebuilding and extending it on Monday, rather than a recording. You also leave with the group: everyone in the room has built the same thing, people stay in touch afterwards, and you hear first when the next one is on - reunions and build sessions, editions for particular functions, and the free meetups. Some of those are paid and some are free. It is an upside rather than a reason to come. ### Who is it not for? People who want a conceptual overview of AI - this is a build day, not a briefing. Anyone who cannot bring a real task from their own work, because the whole day runs against one and without it you get a demonstration instead of an outcome. And anyone expecting a vendor or platform pitch, because no tool is being sold. ### Who runs it? Drummond Gilbert, who tutors in AI in Management at Saïd Business School and runs Oxford’s Practical AI meetup, with Nicolai Thomson (founder, Jenesys AI), Emine Gokce Phillips (co-founder and CEO, Novaria) and Jonathan Waddingham (product consultant). Each runs part of the day, and the day is facilitated rather than lectured, so there are people in the room with you while you build. --- ## Event: The Oxford Agentic Bootcamp - Cohort 3 URL: https://oxfordagentic.com/events/oxford-agentic-bootcamp-cohort-3/ Programme: The Oxford Agentic Bootcamp Status: past Date: Thursday, 19 November 2026 Time: 09:00–17:00 (Europe/London) Venue: Worcester College, Walton Street, Oxford, OX1 2HB Cohort 3 of the Oxford Agentic Bootcamp ran on Thursday 19 November 2026 at Worcester College, Oxford. The day asked each person to bring one real task from their own work and to build a working agentic AI workflow against it. A small cohort, facilitated rather than lectured. Who it was for: - Founders, directors and C-suite building with AI, not reading about it - Graduates of Oxford AI executive education who want the hands-on follow-through - Consultants, and senior NHS, university and public-sector staff - Anyone who uses AI daily but has never chained several steps together - Engineers who build AI for a living - several came to Cohort 1 What the day was built around: - A working agentic pipeline running on a real task from your own work - A review step that checks its own output, so it compounds rather than single-shots - A written recipe, so you can rebuild and extend it on Monday Agenda: - 09:00 Registration and coffee - 09:15 Welcome and framing - Why prompting and orchestrating are different skills, and what today is actually for. - 09:35 Anatomy of an agent pipeline - Built live in front of the room, on a real example - scout, drafter, reviewer. - 10:15 Build one - your own task, first pass - Facilitated. Everyone gets to a rough working pass, however untidy. - 11:30 Guest session - A working practitioner on how they use this daily. - 12:15 Build two - add a review step - Push the pipeline past single-shot by giving it something that checks its own work. - 13:00 Lunch - 13:45 Guest session - AI coding and agentic tooling in practice. - 14:15 Build three - the real thing - Apply it properly to the task you came with. This is the one you take home. - 15:30 Showcase and debrief - Everyone runs what they built, and says what they will use it for next week. - 16:00 Making it stick - 16:20 Wrap and networking ### What did people in Cohort 1 actually build? Real tasks from their own work rather than a set exercise, so they varied. The ones that surfaced in the room were lead generation, internal reporting and marketing content. The common shape was a step that gathers, a step that produces and a step that checks, handing off to each other. If your task has that shape, it will fit the day. ### Do I need to be technical? No, and that is not the whole answer. No coding is required and the day is facilitated in the room, so nobody spends it stuck on setup. Nothing is hand-waved either: part of the day goes on where these workflows break, how you would notice and what you constrain. Engineers are welcome, and several people who build AI for a living came to Cohort 1. ### What should I bring? A laptop, and one real task from your own work that currently costs you meaningful time each week. That task is what you build against all day. A paid plan such as ChatGPT Plus or Claude Pro gets you more out of it, but it is not essential and we make sure everyone can take part either way. ### What if my task is confidential? You choose the task, so a useful test is to bring one you would be comfortable describing to an external supplier, and leave the rest out. Credentials, keys and other people’s personal data do not belong in an AI workflow in any form. You build on your own machine, and the day is not recorded. That still leaves most of the tasks that cost you a morning every week. ### Can I install software on my work laptop? Worth checking before you book. Corporate and government devices often block new tool sign-ups, and the day involves setting things up on your own machine. This caught people out in Cohort 1, so it is better found now than at 09:30 on the day. If your laptop is locked down, email hello@oxfordagentic.com and we will work out what you need beforehand. ### How do I justify this to my manager? If you need to put it in an email, this is the version that survives being forwarded: a day out to build one working agentic workflow on a task we already do by hand every week, and come back able to build the next one without help. Pricing is on the booking page, so the numbers are easy to attach. ### Can my organisation pay by invoice? Yes. Pricing and the booking form are on this page. If your organisation would rather pay by invoice than book on the form, email hello@oxfordagentic.com and we will arrange it. ### How big is the cohort? Small, and capped. Cohort 1 ran with sixteen. The day is facilitated rather than lectured, which stops working in a large room. The exact number of places left is on the booking page. ### What happens after the day? You leave with the workflow running, plus the written recipe for rebuilding and extending it on Monday, rather than a recording. You also leave with the group: everyone in the room has built the same thing, people stay in touch afterwards, and you hear first when the next one is on - reunions and build sessions, editions for particular functions, and the free meetups. Some of those are paid and some are free. It is an upside rather than a reason to come. ### Who is it not for? People who want a conceptual overview of AI - this is a build day, not a briefing. Anyone who cannot bring a real task from their own work, because the whole day runs against one and without it you get a demonstration instead of an outcome. And anyone expecting a vendor or platform pitch, because no tool is being sold. ### Who runs it? Drummond Gilbert, who tutors in AI in Management at Saïd Business School and runs Oxford’s Practical AI meetup, with Nicolai Thomson (founder, Jenesys AI), Emine Gokce Phillips (co-founder and CEO, Novaria) and Jonathan Waddingham (product consultant). Each runs part of the day, and the day is facilitated rather than lectured, so there are people in the room with you while you build. --- ## Event: The Oxford Second Brain Bootcamp - Cohort 1 URL: https://oxfordagentic.com/events/oxford-second-brain-bootcamp-cohort-1/ Programme: The Oxford Second Brain Bootcamp Status: past Date: Wednesday, 21 October 2026 Time: 09:00–17:00 (Europe/London) Venue: Pitt Rivers Museum, South Parks Road, Oxford, OX1 3PP Cohort 1 of the Oxford Second Brain Bootcamp ran on Wednesday 21 October 2026 at the Pitt Rivers Museum, Oxford. It was a one-day, in-person course for business leaders who use AI daily and have hit the ceiling of prompting alone: everyone built a structured second brain from their own material and left with it working and set up to maintain. No coding was required. Who it was for: - Business leaders, founders, partners and senior operators who use AI most days - People who have hit the ceiling of what prompting alone can do - Anyone tired of re-explaining their business to a tool that forgets by Tuesday What the day was built around: - A working second brain, populated with your own material during the day - An architecture for your context, so it stays findable as it grows - A maintenance routine, so the system improves with use rather than decaying - The connections that let your AI tools draw on it, across tools Agenda: - Morning Why context beats prompting - What a second brain is, its anatomy, and designing yours. - Midday The build - Populating it with your own material, in formats that stay sustainable. - Afternoon Making it last - Maintenance, memory management, and connecting it to your tools. - Close What it makes possible - Where this goes next, and time with the room. ### When and where was it? Wednesday 21 October 2026, 09:00 to 17:00, at the Pitt Rivers Museum on South Parks Road in central Oxford. That date has passed and dates for the next cohort are being confirmed. ### Can I still join this cohort? No - this date has passed and registration for it has closed. Dates for the next cohort are being confirmed. ### What is a second brain? A structured, maintained store of your own context - how you write, what good looks like, what your business knows - that your AI assistants can draw on. With it, answers stop being generic; without it, every conversation starts from zero. --- ## Event: The Oxford Agentic Bootcamp - Cohort 2 URL: https://oxfordagentic.com/events/oxford-agentic-bootcamp-cohort-2/ Programme: The Oxford Agentic Bootcamp Status: past Date: Wednesday, 16 September 2026 Time: 09:00–17:00 (Europe/London) Venue: Worcester College, Walton Street, Oxford, OX1 2HB Cohort 2 of the Oxford Agentic Bootcamp ran on Wednesday 16 September 2026 at Worcester College, Oxford, and sold out beforehand. Everyone brought one real task from their own work and spent the day building a working agentic AI workflow against it, facilitated rather than lectured. Who it was for: - Founders, directors and C-suite building with AI, not reading about it - Graduates of Oxford AI executive education who want the hands-on follow-through - Consultants, and senior NHS, university and public-sector staff - Anyone who uses AI daily but has never chained several steps together - Engineers who build AI for a living - several came to Cohort 1 What the day was built around: - A working agentic pipeline running on a real task from your own work - A review step that checks its own output, so it compounds rather than single-shots - A written recipe, so you can rebuild and extend it on Monday Agenda: - 09:00 Registration and coffee - 09:15 Welcome and framing - Why prompting and orchestrating are different skills, and what the day was actually for. - 09:35 Anatomy of an agent pipeline - Built live in front of the room, on a real example - scout, drafter, reviewer. - 10:15 Build one - your own task, first pass - Facilitated. Everyone got to a rough working pass, however untidy. - 11:30 Guest session - A working practitioner on how they use this daily. - 12:15 Build two - add a review step - Pushing the pipeline past single-shot by giving it something that checked its own work. - 13:00 Lunch - 13:45 Guest session - AI coding and agentic tooling in practice. - 14:15 Build three - the real thing - Applying it properly to the task each person came with - the one they took home. - 15:30 Showcase and debrief - Everyone ran what they built, and said what they would use it for next week. - 16:00 Making it stick - 16:20 Wrap and networking ### Has the Oxford Agentic Bootcamp run before? Yes. Cohort 1 ran on Tuesday 21 July 2026 at St Anne’s College, Oxford, with sixteen attendees, and Cohort 2 on Wednesday 16 September 2026 at Worcester College. Later cohorts have run the same one-day format in an Oxford college, and the events page lists the dates on sale. ### Can I still join this cohort? No - this date has passed. The next cohort is the one to book; it runs the same format, in the same kind of small, capped room. ### What did people in Cohort 1 actually build? Real tasks from their own work rather than a set exercise, so they varied. The ones that surfaced in the room were lead generation, internal reporting and marketing content. The common shape was a step that gathers, a step that produces and a step that checks, handing off to each other. If your task has that shape, it will fit the day. --- ## Event: The Oxford Agentic Bootcamp - Cohort 1 URL: https://oxfordagentic.com/events/oxford-agentic-bootcamp-cohort-1/ Programme: The Oxford Agentic Bootcamp Status: past Date: Tuesday, 21 July 2026 Time: 09:00–17:00 (Europe/London) Venue: St Anne’s College, 56 Woodstock Road, Oxford, OX2 6HS Cohort 1 of the Oxford Agentic Bootcamp ran on Tuesday 21 July 2026 at St Anne’s College, Oxford. Sixteen founders, directors and senior professionals spent the day taking a real workflow apart, building agents and skills against their own work, and working through what it takes to bring that back into an organisation. No coding was involved. Who it was for: - Founders, directors and C-suite - Consultants and senior public-sector professionals - Oxford AI executive-education graduates What the day was built around: - A working agentic pipeline running against a real task - A review step, so the pipeline compounds rather than single-shots - A written recipe for rebuilding and extending it afterwards Agenda: - 09:00 Registration and coffee - 09:30 Deconstructing a real workflow - Breaking a workflow into steps before deciding where AI belongs, worked through on webinar summaries and lead generation, with a worksheet. - 10:00 The main block - context, skills and agents - Connecting your own email, files and meeting notes as context; building skills for recurring tasks; agents, subagents and workflows; Claude Code and shared repositories; small self-contained tools. Broken for lunch. - 14:30 Adoption and governance - The five-level adoption journey, from personal assistant to autonomous systems, with compliance review and marketing analytics as worked examples, and where data integrity and risk sit. - 16:00 Close and networking Outcomes: - 16 attendees, Cohort 1, 21 July 2026 - 3 teaching blocks: deconstruction, build, adoption - 1 real task each, brought from their own work ### Has the Oxford Agentic Bootcamp run before? Yes. Cohort 1 ran on Tuesday 21 July 2026 at St Anne’s College, Oxford, with sixteen attendees, and Cohort 2 on Wednesday 16 September 2026 at Worcester College. Later cohorts have run the same one-day format in an Oxford college, and the events page lists the dates on sale. ### Can I still join this cohort? No - this date has passed. The next cohort is the one to book; it runs the same format, in the same kind of small, capped room. ### What did people actually work on? Real tasks from their own work rather than a set exercise, so they varied. The ones that surfaced in the room were lead generation, internal reporting and marketing content. The common shape was a step that gathers, a step that produces and a step that checks, handing off to each other. --- ## Outcomes and testimonials "I came into the July 2026 AI Agentic Bootcamp as a basic AI user who had never used Claude, and I left with the ability to start building professional-grade agents for my own workflow." - Fei Gan, Cohort 1 attendee, Serambi Alam (Cohort 1, 14 August 2026) "Best, most practical AI session I have attended to date - highly recommend it!" - Kirsten Samuel, CEO, Kamwell (Cohort 1, 13 August 2026) "I left with more than ideas. I left with something I could apply the following week." - Dr Zeynep Hizir, Transformation Executive (Cohort 1, 2 August 2026) "It was a really good session… Looking forward to the next one." - Shrestha Mullick, Founder, The Brand Mirror (Cohort 1, 23 July 2026) --- ## People ### Drummond Gilbert URL: https://oxfordagentic.com/team/drummond-gilbert/ Role: Founder and host, Oxford Agentic Drummond Gilbert runs the Oxford Agentic Bootcamp. He teaches on AI at the University of Oxford’s Saïd Business School, organises Oxford’s Practical AI meetup, and has spent his career as a marketing leader - most recently as a fractional CMO - building and shipping with these tools rather than writing about them. ### Nicolai Thomson URL: https://oxfordagentic.com/team/nicolai-thomson/ Role: Founder, Jenesys AI Nicolai Thomson is the founder of Jenesys AI, a multimodal AI platform for accounting practices. He leads the technical strand of the bootcamp, and builds with these tools daily across a portfolio of accounting and AI projects. ### Emine Gokce Phillips URL: https://oxfordagentic.com/team/emine-gokce-phillips/ Role: Co-founder and CEO, Novaria Dr Emine Gokce Phillips is co-founder and CEO of Novaria and runs the AI Centre for Business. She delivers AI workshops for organisations, and brings the side of adoption that tooling alone never solves: getting an organisation to actually change how it works. ### Jonathan Waddingham URL: https://oxfordagentic.com/team/jonathan-waddingham/ Role: Product consultant Jonathan Waddingham is an independent product consultant and the former Chief Product Officer at Lightful. He works on the step most people skip: defining the problem precisely enough that it is worth handing to a machine. ### Adrian Shedden URL: https://oxfordagentic.com/team/adrian-shedden/ Role: Founder, Accelerated Impact Ops Adrian Shedden runs Accelerated Impact Ops and his life on the system the Second Brain Bootcamp teaches - one human, teams of AI agents. A fintech lawyer turned serial founder, he also hosts Mindstone Practical AI Bristol. ### Glenn Smith URL: https://oxfordagentic.com/team/glenn-smith/ Role: Founder, GenieOS Glenn Smith has put AI into production for twenty years, from FX market-making at UBS and Barclays to advising HSBC on AI strategy, with two startup exits along the way. He now runs two AI companies built on proprietary memory systems. ### Adam Martin URL: https://oxfordagentic.com/team/adam-martin/ Role: Founding CTO Adam Martin is a founding CTO with exits behind him and a master’s in computing from Cambridge. He built a second brain memory system for teams, where humans and agents share one memory. ### Mike Bailey URL: https://oxfordagentic.com/team/mike-bailey/ Role: Founder, Coppice Labs Mike Bailey runs Coppice Labs, a product consultancy in Oxford that works with founders and product teams on where a digital product goes next and what the evidence behind that decision supports. He is one of the three people who run the Oxford Agentic Coding Bootcamp. --- ## Guide: You’ve finished an AI executive programme. Now what? URL: https://oxfordagentic.com/guides/after-the-ai-executive-programme/ Updated: 18 October 2026 AI executive programmes are good at what they set out to do: the strategy, the governance, the frame. What most leave open is the build - you finish able to talk about agentic AI credibly and still without a working example in your own work. The next step is not another course; it is one built workflow. A large share of the people in our rooms have completed a serious AI programme - an executive education course, a diploma, a strategy module. They are not beginners, and they are not confused about the technology. They tend to say the same sentence: I understand this well - I have just never built anything. That sentence is not a criticism of the programme. It is a description of where the programme's job ends. ## What the programme gave you A good executive programme leaves you with things that are genuinely hard to acquire alone: the economic frame, the governance questions, the vocabulary to hold your own in a board discussion, and a considered view of where this technology fits your organisation. Those are strategic assets, and they age well. What it almost never leaves you with is a working example - an AI workflow, built by you, running against a task from your own week. Courses assess understanding; they rarely assess artefacts. So the understanding is real, and the hands are still empty. ## Why the gap matters more than it looks Without a built example, AI strategy stays delegated. You can commission work, but you cannot judge it from the inside: what is easy, what is genuinely hard, where the failure modes live, how long things should take. Leaders who have built one working pipeline - even a modest one - ask noticeably different questions of their teams and their vendors, because they have felt where the effort actually goes. There is also a personal cost to the gap. The longer the distance between finishing the course and building anything, the more the knowledge settles into something you once learned rather than something you use. ## What closing it actually takes Less than the course did. Building a first agentic workflow is not a semester of work; it is a focused day, if three conditions are met: - **A real task from your own work** - not an exercise, and [worth choosing carefully](/guides/how-to-choose-your-first-agentic-task/). The transfer problem disappears when there is nothing to transfer. - **A working pattern to follow** - [a few narrow steps handing off to each other](/guides/scout-drafter-reviewer-pattern/), with the judgement kept where it belongs. The pattern is learnable in a morning. - **Someone to unstick you** - because the thing that ends most self-directed attempts is not the concept but the first snag, alone at a desk. None of this requires code, and none of it repeats what the programme taught. It sits on top of it. ## The honest sequencing If you have not done any structured AI learning, [an executive programme or a good course](/guides/ai-bootcamp-vs-course-vs-exec-education/) is a defensible place to start - the frame matters. If you have finished one, the highest-return next step is not a second course. It is turning what you know into one working thing, and taking the pattern home. If your question is when rather than whether, [how the cohort calendar usually falls](/guides/when-ai-executive-cohorts-run/) is worth a look before you plan around it. That is the day the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is built to be: one day in Oxford, a small room of senior peers, your own real task, and a working agentic workflow by the time you leave. --- ## Guide: Agentic AI for business leaders: where to start URL: https://oxfordagentic.com/guides/agentic-ai-for-business-leaders-where-to-start/ Updated: 23 July 2026 If you already use AI daily and want to go further, the place to start is one real task from your own work, not another tool. Pick a job you repeat, break it into a few narrow steps, and hand each step to the model in turn. That shift, from prompting to orchestrating, is where agentic AI begins. Most senior people we meet are past the "what is AI" stage. They use it every day, they have views on where it goes, and they are quietly frustrated that it still feels like a cleverer search box. The gap is not knowledge. It is having built something that runs on its own. ## Start with a task, not a tool The instinct is to pick a tool and look for uses. Reverse it. Choose one job you do again and again that eats real time: the weekly board update, the first draft of a proposal, sifting inbound, turning a transcript into actions. A repeated job with a clear input and a clear output is the right size to build against. If you cannot describe the job to a new hire in two minutes, it is too vague to hand to a machine. Sharpen the task first. ## Break it into narrow steps A single prompt asking for everything gives you a mediocre everything. The agentic move is to split the job into a few steps, each narrow enough to be reliable, and let them hand off: - a step that gathers the raw material, - a step that produces a first draft, - a step that checks the draft against what good looks like. None of this is magic, and none of it removes your judgement. The machine does the scaffolding; the decision at the end stays yours. What changes is that the part which used to take a morning now takes a fraction of it. ## Keep a human at the decision, not the grunt work The failure mode is either trusting the output blindly or doing everything yourself. The useful middle is to automate the gathering and drafting, and spend your attention on the one judgement that matters. Build the workflow so that judgement is the last step, not the whole thing. ## Where this gets practical Reading about this only gets you so far. The point at which it clicks is when you have built one working pipeline against your own task and watched it run. That is what a day at the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is for: you bring the task, you leave with the workflow, and you take home the pattern to build the next one. --- ## Guide: Agentic AI for consultants: sell the judgement, not the assembly URL: https://oxfordagentic.com/guides/agentic-ai-for-consultants/ Updated: 18 November 2026 Consultants get the most from agentic AI on the assembly work inside an engagement: synthesising discovery interviews, drafting proposals from past proposals, turning meeting notes into structured findings. The advice, the judgement and the client relationship stay human. Unlike hiring another pair of hands, a workflow built for one engagement is still there for the next one. Consulting has an odd relationship with AI. The work is mostly words, research and structure - exactly what the tools are good at - yet the thing clients pay for is judgement, which the tools do not have. The way through is to be precise about which half of the job you are automating. ## The work that suits it Most engagements contain a layer of assembly: gathering, structuring and drafting that has to happen before the thinking can. That layer is where agentic workflows earn their keep: - **Discovery synthesis.** Ten interviews produce ten sets of notes in ten shapes. A workflow that pulls them into one structure, groups what was said by theme, and flags where people disagreed turns a day of collation into a review. - **Proposal assembly.** Most proposals are built from previous proposals. A pipeline that drafts from your own past work, in your own structure, against the client's actual brief, gets you to the version worth editing much faster. - **Findings from meetings.** Notes and transcripts into structured findings, each tagged with where it came from - so the trail from evidence to recommendation is visible when a client asks. - **Checking the deliverable against the brief.** A reviewer step that reads the draft and the brief side by side and lists what was promised but not answered. Dull, mechanical, and exactly the thing that gets missed at 11pm before a deadline. ## The work that does not The recommendation itself. The prioritisation. Anything where the client is paying for your view rather than your assembly. Hand those over and you are not more efficient - you are selling a worse product with your name on it. Confidentiality draws its own line. Client material goes only into tools your engagement terms and their policies actually allow. That is not a footnote for a consultant; it is the condition for using any of this at all. Check first, build second. ## Why this compounds for consultants in particular A consultancy sells time, which means every efficiency either becomes margin or capacity. An associate you hire helps on the engagements they staff. A workflow you build is different in kind: the discovery-synthesis pipeline from this engagement is still there for the next one, and the one after that. Firms that treat workflows as reusable assets - named, kept, improved between engagements - are building something the timesheet never shows. ## Where this gets practical The hard part is not agreeing with any of this. It is building the first workflow against a real engagement rather than a demo. That is what a day at the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is for: bring one real task - a synthesis, a proposal, a review step - build a working pipeline against it in a room where someone walks over when you get stuck, and leave with it running. No coding required, and nothing hand-waved. --- ## Guide: Agentic AI for finance teams: where it actually helps URL: https://oxfordagentic.com/guides/agentic-ai-for-finance-teams/ Updated: 1 November 2026 Finance teams get the most from agentic AI on the repeated, evidence-heavy work around the numbers rather than the numbers themselves: drafting commentary, triaging reconciliations, assembling board packs, first-pass contract review. The judgement, the sign-off and the controls stay with the team. The machine removes the assembly, not the accountability. Finance is one of the better places to apply agentic workflows, and one of the easier places to apply them badly. The distinction is worth getting right before you build anything. ## The work that suits it The pattern that works is repeated, evidence-heavy, and ends in something a human signs off. That describes a lot of a finance month: - **Commentary and variance narrative.** The numbers come from the ledger; the explanation gets rewritten every month in much the same shape. A pipeline that gathers the movements, drafts the narrative and checks it against last month's wording removes most of the typing without touching the figures. - **Reconciliation triage.** Not doing the reconciliation, but sorting the breaks: grouping them, proposing the likely cause, and putting the handful that need a person at the top. - **Board and investor pack assembly.** Pulling the same inputs into the same structure, every time, with a step that checks nothing is missing before it reaches you. - **First-pass document review.** Contracts, supplier terms, policy documents - a step that flags what departs from your standard, so a person reads the exceptions rather than everything. ## The work that does not Anything where the output is the number itself, or where being subtly wrong is expensive and hard to spot. Do not hand over the calculation, the judgement, or the control. If you cannot easily check whether an output is right, it is a poor candidate - not because the tool cannot attempt it, but because you have no cheap way to catch it when it drifts. ## Build the check into the workflow The single most useful habit in a finance context is making the reviewer a step in the pipeline rather than an afterthought. A workflow that drafts and then explicitly checks its own draft against a stated standard catches far more than one that drafts and stops. It also leaves you something to show: a record of what was checked and against what. Keep the audit trail in mind from the start. Where the inputs came from, what the workflow did, and who approved the output are not bureaucratic extras in finance - they are the thing that makes the output usable. ## Is there a one-day, hands-on agentic AI course for finance teams in the UK? There is a one-day, in-person build day in Oxford - the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) - and it is worth being straight about what it is not. It is not a finance-only course, and no part of the day is pitched at the finance function in particular. What is finance-specific is the task you bring: you spend the day building against your own variance commentary, your own reconciliation triage, your own board pack, rather than a worked example from somebody else's business. No coding is required, the cohort is capped small and mixed by function, and you leave with the workflow running. If the prior question is whether a cohort seat is the right unit of training at all, [an in-house day or a cohort seat](/guides/in-house-ai-day-or-send-people-to-a-cohort/) is the comparison to read first. ## Can we build against real finance data on a day like that? You bring a task, not a data set, and that distinction does most of the work. A useful test when choosing one is whether you would be comfortable describing the task to an external supplier; if you would not, pick a different one. Credentials, keys and other people's personal data do not belong in an agentic workflow in any form. That still leaves plenty. A commentary or reconciliation-triage workflow can be built against a redacted or representative extract and pointed at the real ledger once it is back inside your own controls, because what you are building is the pipeline rather than a one-off output. The day is not recorded, and you build on your own machine. ## Does this need sign-off before anyone books? Usually two different ones, and they are easier taken separately. The spend and the day out are a manager's decision, and [getting an AI bootcamp approved](/guides/getting-an-ai-bootcamp-approved/) is the version of that case which survives being forwarded. Putting the resulting workflow into a finance process is the other, and it is the one finance has to answer properly: where the inputs came from, what the workflow did, and who approved the output. [Getting an AI workflow past IT and legal](/guides/getting-an-ai-workflow-past-it-and-legal/) sets out what each side tends to ask for. Neither conversation needs the workflow to exist first, but both go better once it does. ## Where this gets practical The step people find hard is not understanding this, it is building the first one. That is the point of a day at the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/): you bring one real task from your own work, build a workflow against it, and leave with it running. No coding, and the judgement stays where it belongs. --- ## Guide: Agentic AI for public-sector and NHS leaders: start where governance says yes URL: https://oxfordagentic.com/guides/agentic-ai-for-public-sector-and-nhs-leaders/ Updated: 21 November 2026 Public-sector and NHS leaders can apply agentic AI today by choosing tasks their governance already permits: assembling papers from public documents, synthesising consultation responses, first drafts of reports in house style. Decisions about individuals, clinical judgement and personal data stay out of scope. The constraint is not a blocker - it is the selection criterion for what to build first. Nobody in the public sector needs telling that AI could help. The question that actually decides things is narrower: what can you use it on, given information governance, procurement and the fact that you cannot adopt tools unilaterally? The honest answer is: more than most people think, provided you choose the work with the constraint in mind rather than against it. ## Governance is the design brief The private-sector framing - pick your most painful task and automate it - fails in a trust or a council, because the most painful task usually involves personal data, an individual's circumstances, or a decision someone must be accountable for. Start instead from what your organisation's rules already permit, and the field is still surprisingly large: - **Papers and reports assembled from public material.** Board papers, committee briefings and strategy documents are mostly synthesis of documents that are already public or internal-but-unrestricted. A workflow that gathers, structures and drafts against your house style removes the assembly without touching a single record. - **Consultation and engagement synthesis.** Hundreds of free-text responses, grouped by theme with disagreements flagged and every claim traceable to a response - the kind of work that eats a fortnight and suits a pipeline exactly. - **Meeting notes into minutes and actions.** In your organisation's own approved tools, a drafter-and-checker pair turns notes into consistent minutes with an action list, ready for a human to correct rather than compose. - **First drafts against a template.** Job descriptions, service specifications, policy summaries - anywhere the structure is fixed and the content is unrestricted, a draft-then-review workflow earns its keep. ## The work that stays out of scope Decisions about individuals: eligibility, safeguarding, anything clinical. Not because a workflow could not produce an output, but because accountability for those decisions cannot be delegated, and being subtly wrong about a person is the failure mode your governance exists to prevent. The same goes for feeding personal or patient data into tools that have not been through your organisation's approval process. If the task needs data you would not put in an email to an external supplier, it is not the task to learn on. None of this is a limitation unique to AI. It is the same line your organisation already draws, applied consistently. ## Budgets are an argument for method, not against it Public-sector budgets do not stretch to speculative platforms, and they do not need to. An agentic workflow is a method - a sequence of narrow steps with a check at the end - built in tools your organisation may already license. The investment is a day of learning the pattern, not a procurement exercise. And the wins, while smaller-scoped than a vendor deck would claim, are real: hours back each week on assembly work, applied by the person who understands the service. ## Where this gets practical The gap between agreeing with all this and having a working example is the part that rarely closes on its own. That is what a day at the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is for: bring one real task - a paper, a synthesis, a template draft, chosen with your own governance in mind - and build a working agentic workflow against it, with someone on hand when you get stuck. No coding, and nothing about the day requires putting restricted data anywhere it should not be. --- ## Guide: A plain-English glossary of agentic AI URL: https://oxfordagentic.com/guides/agentic-ai-glossary/ Updated: 16 October 2026 Agentic AI comes wrapped in vocabulary that makes a simple idea sound harder than it is. This glossary defines the terms that actually come up - agent, workflow, orchestration, prompt, context, tool use and a handful more - in plain English, for business leaders who want to build rather than translate. Most of the vocabulary around agentic AI describes simple ideas. Here are the terms worth knowing, defined the way we use them in the room. ## Agent A piece of AI given a goal and the means to work towards it - rather than a single question and a single answer. An agent can take several steps, check its own progress and use tools along the way. The word sounds grander than the reality: a well-built agent is usually narrow, doing one job dependably. ## Agentic workflow Several narrow AI steps connected so they hand off to each other: one gathers, one drafts, one checks. The workflow is the unit that does useful work - most of the value people attribute to "agents" comes from a well-shaped workflow, not from any single clever step. ## Orchestration Deciding how the steps connect: what runs first, what each step passes to the next, and where a human sits. Orchestration is the skill that separates using AI from building with it - it is closer to delegation than to programming. ## Prompt The instruction you give a model. Prompting well matters, but it is a single-step skill; the shift that changes your working week is going from writing one good prompt to orchestrating several narrow ones. ## Context Everything the model can see when it works: your instructions, the documents you give it, the conversation so far. Most disappointing AI output is a context problem, not an intelligence problem - the model was never shown what good looks like. ## Second brain A structured, maintained store of your own context - how you write, what good looks like, what your business knows - kept where your AI tools can draw on it, so you stop re-explaining yourself in every conversation. It is the fix for context being the bottleneck: build the supply once instead of retyping it each time. Building one from your own material is a day's work - [we run a bootcamp for exactly that](/bootcamps/second-brain-bootcamp/). ## Tool use An agent calling something outside itself - searching the web, reading a file, updating a spreadsheet. Tool use is what turns a text generator into something that can act on your systems rather than just describe them. ## Pipeline Another word for a workflow, borrowed from engineering: work flowing through stages in order. If you can sketch your task as three boxes with arrows, you have designed a pipeline. ## Scout, drafter, reviewer The shape we teach first, because it fits an enormous amount of knowledge work: a step that gathers the raw material, a step that produces a first version, and a step that checks it against a stated standard before a human sees it. ## Human in the loop The deliberate decision to keep a person at the point of judgement - approving, correcting, deciding - while the machine handles the assembly around them. Good agentic design is mostly about placing this point well. ## Model The underlying AI - ChatGPT, Claude, Gemini and their peers. For business use the differences matter less than people expect; the shape of your workflow matters more than which model runs it. ## MCP (Model Context Protocol) A standard that lets AI tools connect to your systems - calendars, files, databases - in a common way, so an agent can reach the things your work actually lives in. You do not need to know how it works; you will increasingly benefit from tools that use it. ## Coding agent An AI that writes and changes code itself, working in your files and your terminal rather than in a chat window. Claude Code and Codex are the two most people meet first. The typing stops being the hard part; what matters is the brief the agent reads before it starts, asking for a plan before it builds, and checking what it wrote. That workflow is the subject of the [Oxford Agentic Coding Bootcamp](/bootcamps/agentic-coding-bootcamp/). ## Repository The folder a project's files live in, kept with a full history of every change. It is the undo button for work an agent did: if a change makes things worse, you go back to the version before it. Agents write quickly, which is exactly why the history is worth keeping. ## No-code Building all of the above through plain-language instructions and existing tools, without programming. The entire [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) runs this way: the constraint is clear thinking about your task, not code. The Coding Bootcamp is the exception in subject rather than in prerequisite - it builds real software, and the agents still do the typing. None of our days expects you to arrive able to code. --- ## Guide: AI bootcamp, AI course or executive education: which do you need? URL: https://oxfordagentic.com/guides/ai-bootcamp-vs-course-vs-exec-education/ Updated: 7 November 2026 An executive education programme gives you the strategic frame, an online course gives you knowledge at your own pace, and a bootcamp gives you a working result. The right choice depends on what you want to hold at the end: a view, an understanding, or something built and running against your own work. These labels get used interchangeably, but they describe genuinely different things. The cleanest way to choose is to ask what you want to be holding when it ends. ## Executive education: a frame University executive programmes - the diplomas, the strategy modules - are good at what they are for: the landscape, the economics, the governance questions, a peer network, a credential. If you need to set direction, allocate budget or speak credibly about AI at board level, this is the right instrument. Some now include a build exercise as well, and the better ones are worth doing. What they rarely leave you with is your own work running afterwards: a guided build on a sample case is a different thing from the task that was actually on your desk on Monday. That is not a flaw; it is not the job. Many of the people in our rooms have completed exactly this kind of programme and arrive saying some version of the same thing: I understand it - I have just never built anything. [Closing that particular gap](/guides/after-the-ai-executive-programme/) is a different decision from the one this page is about. ## An online course: knowledge, at your pace Online courses are the most flexible and the most affordable route, and the good ones teach real material. They suit people who want depth on their own schedule, and they are unbeatable for reference - you can return to a module the week you need it. Their weakness is completion, and more precisely application. Watching someone build is not building. The friction that stops people is rarely the concept; it is the first stuck moment at their own desk, with no one to unstick them. That is where most courses quietly end. ## A bootcamp: a result A bootcamp - at least as we mean it - is a short, facilitated stretch of time with one measure of success: you leave with something working. The format exists to remove the two things that stall self-directed learning: the setup friction, because someone walks the room, and the deferral, because a date in the diary with other people in the room is a commitment kept. The trade-off is scope. A day cannot cover a landscape; it covers one pattern, properly, applied to one real task from your own work. Our view is that this trade is the right one for operators - the pattern transfers, and the working result is what turns understanding into habit. That is a narrower promise than most training pages make, and it is worth being precise about it: [what you can actually build in one day](/guides/what-you-can-build-in-one-day-with-ai/) sets out what finishes inside a day and the four things that do not. ## Choosing between them - **You need to set strategy or govern AI use** - executive education. - **You want broad knowledge on your own schedule** - an online course. - **You want something built and running against your own work** - a bootcamp. - **You have already done the first two** - that is the most common person in our room, and the gap they describe is exactly what a bootcamp closes. They compound rather than compete: the frame from a programme, the depth from a course and a working result from a day of building are worth more together than any one alone. ## The fourth option: doing it yourself, for nothing All three formats above cost money, and an honest comparison has a fourth column, because most of the material is public. Documentation, tutorials and recorded talks cover a great deal of what any of them teach. It is worth being direct about what that does and does not settle. ### Is there anything in the day you could not find for free? Very little, and we would rather say so. The pattern we teach is [scout, drafter, reviewer](/guides/scout-drafter-reviewer-pattern/), and it is written down in public, on this site included - these guides give the argument away rather than trail it. What a day buys is not withheld information. It is a finished thing: your own task, built and running by the time you leave, with someone walking the room at the moment it stalls. ### You are disciplined, so why not just do it yourself? You might, and some people should. The useful test is not whether you could but whether you already have. Many of the people in our rooms are not short of material or of ability - they arrive saying some version of "I understand it, I have just never built anything", and the thing that stopped them was never the concept. It was the first stuck moment at their own desk, on an ordinary working day, with something more urgent waiting. A date in the diary and other people in the room is what removes the deferral. If self-directed study has already left you with something you use every week, you do not need a bootcamp. ### When is a paid day not worth it? When you have nothing real to bring. The day is built around one task from your own work, so arriving without one gets you a demonstration instead of an outcome - and a demonstration is the one thing free material already does well. [What you can build in one day](/guides/what-you-can-build-in-one-day-with-ai/) sets out what finishes inside a day and what does not; if your task sits outside that, the money is better spent elsewhere. One axis this leaves out. If you are choosing for a team rather than for yourself, the question is not only which of the three formats but whether to send two or three people at all: [an in-house day or an open cohort](/guides/in-house-ai-day-or-send-people-to-a-cohort/) is the comparison, and it usually turns on whether you are buying alignment across a team or capability in a few people. The [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is our version of the third: one day in Oxford, a small room of senior people, one real task each, no coding, and a working agentic workflow by the time you leave. --- ## Guide: Claude, Cowork, Claude Code or Codex: which one for which job URL: https://oxfordagentic.com/guides/claude-cowork-claude-code-or-codex/ Updated: 18 November 2026 The quickest way to choose between Claude, Cowork, Claude Code and Codex is to ask what your work ends in. A conversation or a draft suits the Claude chat app. Finished documents and files on your own computer suit Cowork. Software suits Claude Code or Codex. The skill that carries across all four is designing the workflow, not learning the app. Four tools, two companies, and names that seem designed to be confused. Most comparisons of them are written by developers for developers, which is not much use if your working day is documents, spreadsheets and email rather than code. This is the plain version, as the tools stand in September 2026. They change monthly, so treat anything about specific features as a snapshot and check each product's own page before you commit. ## The question that decides it Ask what the job ends in. - **A conversation, a draft, an answer.** The Claude chat app is enough. Most people already use it this way, and most people stop here. - **A finished document, spreadsheet or file on your own computer.** That is what Cowork is for. - **Working software.** That is Claude Code or Codex. Almost every confusion between these tools comes from skipping that question and comparing feature lists instead. ## What each one is for **Claude** is Anthropic's chat assistant, in a browser, a desktop app or on your phone. You ask, it answers, and you carry the result wherever it needs to go. Good for thinking, drafting and one-off tasks. It is not where multi-step work gets done on your behalf. **Cowork** is Anthropic's agent for people who do not write code. It lives in the Claude desktop app and works on the files and folders on your own computer: give it a task, and it goes off and produces the finished document, deck or spreadsheet rather than a reply in a chat window. If your work is documents rather than code, this is the one to learn first. **Claude Code** is Anthropic's agentic coding tool. It works across a whole codebase, reading files, making changes and running commands, and it began in the terminal before spreading into code editors and the desktop app. It is the developer's tool. Some non-developers use it anyway, because the same approach that builds software also builds quite serious automation, but it assumes you are comfortable working in a codebase. **Codex** is OpenAI's agent. It started as a coding agent - in the terminal, in code editors, and in a cloud version you hand tasks to and check back on later - and OpenAI has since widened it beyond code. If your organisation lives in ChatGPT, it is the natural counterpart to the Claude tools. ## If you do not write code Start with whichever agent comes with the subscription you already pay for. That is Cowork if you are on a paid Claude plan, and Codex if you are on ChatGPT. Switching subscriptions to chase a feature is rarely worth it: the differences between the leading tools matter far less than whether you have a workflow worth running in any of them. Skip Claude Code until you have a reason. It is excellent, and it is the wrong first step for someone whose work does not end in software. The reason, when it arrives, is usually that you want to build something rather than automate something: an idea you would like running on the internet rather than a task you would like to stop doing by hand. That is a different day's work, and [what stands between the code an agent writes and something live](/guides/why-the-code-an-agent-wrote-is-not-live/) is mostly not code. ## The part that does not change The tools will look different by next spring. What carries across all four is the part underneath: choosing a task worth automating, breaking it into steps that each do one thing, and deciding where [a human needs to stay in the loop](/guides/where-the-human-goes-in-an-agentic-workflow/). Someone who can do that picks up any new tool in an afternoon. Someone who learned one app's buttons starts again every time it changes. ## Where this gets practical The [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) covers where each of these tools fits as part of the day, and then spends most of it building: you bring one real task and leave with a working agentic workflow running against it. Cohort 4 runs on [3 December 2026 at Worcester College](/events/oxford-agentic-bootcamp-cohort-4/). No coding required, and nothing hand-waved. If the thing you want is software rather than a workflow, the [Oxford Agentic Coding Bootcamp](/bootcamps/agentic-coding-bootcamp/) is the one for that: Claude Code and Codex do the typing, and the day is spent on the workbench around them. It does not assume you can code. What it teaches is everything around the code - the terminal, version control, deployment, databases and tests - and how to check what the AI wrote rather than trusting it. Cohort 1 runs on [26 November 2026 at Worcester College](/events/oxford-agentic-coding-bootcamp-cohort-1/). --- ## Guide: Getting a one-day AI bootcamp approved, and what to promise URL: https://oxfordagentic.com/guides/getting-an-ai-bootcamp-approved/ Updated: 16 November 2026 Approval turns on naming the task, the deliverable and the date. A build day supplies all three: you bring one real task you already do by hand, and you leave with a working agentic workflow running against it. So the request can promise a specific thing running by the end of the day, rather than a better understanding of AI. Most of what is written about getting training signed off was written for a multi-week course, and it gives the same advice every time: tie it to a priority your manager already owns, name a concrete deliverable, agree a date to report back. The advice is right. It is also where the generic version stops, because a course that teaches a subject cannot tell you what your deliverable is going to be. A build day can. [The Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) runs against one real task you bring from your own work, and you leave with a working agentic pipeline running against it. The three things the request needs - the task, the deliverable and the date - are not things you have to invent for the email. They are the format. ## What is actually being decided Rarely the fee on its own. Someone signing this off is deciding three things at once: whether a day out of your diary is worth it this month, whether anything comes back that they can see, and whether saying yes makes work for them afterwards. The fee is usually the smallest of the three and the easiest to attach, because pricing sits on each event booking page. Write to those three and the request gets shorter, not longer. ## Name the task, not the subject "I would like to go on an AI course" is a request about a subject. "I would like to spend Thursday turning the weekly report into something that runs itself" is a request about a task, and it is the version that survives being forwarded. The task is the one part nobody can supply for you. The published condition is that it has to be real: one task from your own work that currently costs you meaningful time, because the whole day runs against it. The programme's own "Who is it not for?" answer is blunt about the alternative - without a real task you get a demonstration instead of an outcome. In [Cohort 1](/events/oxford-agentic-bootcamp-cohort-1/) the tasks that surfaced in the room were lead generation, internal reporting and marketing content, one each, brought from sixteen people's own work. [Choosing yours](/guides/how-to-choose-your-first-agentic-task/) is worth ten minutes before you write anything, because the rest of the request rests on it. ## Promise the thing running, and no more than that What you can honestly say you will come back with is published rather than implied: - The workflow itself, built and running against the task you named. - The pattern behind it - [a step that gathers, a step that produces, a step that checks](/guides/scout-drafter-reviewer-pattern/) - which is what lets you build the second one without help. - A clear sense of where the human judgement stays and where the grunt work goes. Then offer to demonstrate it rather than write it up. "I will show you the thing running in the week after" is a promise that can be kept or visibly not kept, which is the kind that is easy to agree to. What not to promise is a number. The temptation is to put a figure on the hours it saves, because that is the sentence that looks most like a business case. You will not have measured it yet, and [effort moved around is not the same as time saved](/guides/is-your-ai-workflow-saving-time-or-just-effort/). Promise the artefact, measure afterwards, and bring the number to the next conversation once it is actually yours. Nor should the request promise more than a day produces. One day produces one workflow on one task. What scales is the pattern, and saying that plainly is more convincing than the version that overreaches. ## The four questions that come back **"Can you not watch something online instead?"** The day is deliberately not recorded, which is an answer rather than a dodge: [watching someone build is not building](/guides/ai-bootcamp-vs-course-vs-exec-education/), and the friction that stops people is rarely the concept. It is the first stuck moment at their own desk with nobody there to unstick them. [Why that is harder online](/guides/why-in-person-beats-online-for-learning-ai/) is a separate argument, and already written. **"Is this a pitch for some tool?"** No tool is being sold. The published "not for" list says so directly: anyone expecting a vendor or platform pitch should not come, and part of the day goes on where these workflows break and what you constrain so a bad run cannot do damage. **"What if the date stops working?"** Places are not refundable, but a place transfers to a future cohort at no charge. A diary clash is not a lost fee, which is usually the risk inside the question. **"Can we pay properly rather than on a card?"** An organisation can pay by invoice - email hello@oxfordagentic.com and it is arranged. That matters more than it sounds, because a request that needs a personal card and an expense claim is a request with a second approval hidden in it. ## If you are the one who signs it off Then the question is not approval but who goes, and [the seat belongs to whoever owns the task](/guides/who-should-take-the-seat/) rather than to the most senior person who would like a day in Oxford. If two of you go, bring the same task, so the same workflow goes back to work with both of you. And if it turns out several people want the day, [whether to send them to a cohort or bring a day in-house](/guides/in-house-ai-day-or-send-people-to-a-cohort/) is the question to settle first. ## The short version Four sentences, in this order: the task, the deliverable, the date, and the offer to demonstrate it. If the request runs longer than that, something in it is about the subject rather than about the task. Cohort 4 is on [Thursday 3 December 2026 at Worcester College](/events/oxford-agentic-bootcamp-cohort-4/), 09:00 to 17:00, and its booking page carries the figure to attach. --- ## Guide: Getting an agentic workflow past IT, legal and your manager URL: https://oxfordagentic.com/guides/getting-an-ai-workflow-past-it-and-legal/ Updated: 18 November 2026 An agentic workflow that works on your laptop still has to be allowed to run inside your organisation. IT will ask what it touches and what it can do unasked. Legal will ask whose data it is and which terms govern it. Your manager will ask what it replaces and who owns it. Answer all three on one page before you ask. Building an agentic workflow that works is the part people expect to be hard. Then comes the next conversation: the one where it has to be allowed to run somewhere that is not your laptop. Most of what is written about that conversation is written for the person on the other side of the table - the security lead with a seven-gate approval checklist. This is for the person asking. The questions are predictable, and turning up with the answers written down changes the meeting from a request into a review. ## What IT and security will ask ### What does it touch? Every system, folder, inbox and tool the workflow reads from or writes to. Be specific. "Our shared drive" is a worry; "one folder of supplier reports, read only" is an answer. ### Whose access does it use? A workflow running on your personal login can do everything you can do, which is far more than it needs. Expect to be asked for access scoped to the task alone. ### What can it do without asking? Reading and drafting are one conversation. Sending, deleting, paying or changing records are another. If anything irreversible sits behind [a human checkpoint](/guides/where-the-human-goes-in-an-agentic-workflow/), say so up front. ### Which tool, and is it approved? Check your organisation's list before the meeting, not during it. A workflow built in an unapproved tool is a rebuild, however good it is. ### How would you stop it? Who can switch it off, and how quickly. The honest answer "I would turn off this one connection" is more reassuring than any description of how carefully it was built. ## What legal and data protection will ask ### Whose information goes through it? Your own drafts and internal material are one thing. Named customers, staff records or anything about identifiable people is another, and data protection law applies to it however clever the workflow is. ### Which terms govern the tool? Whether the provider can use your inputs for training and how long material is kept depends on the exact plan your organisation pays for. Read those terms first, and bring the relevant sentences rather than a summary of them. The same questions come up for [anything you store for AI to read](/guides/is-it-safe-to-put-company-documents-in-a-second-brain/). ### Does a client contract say anything? Material a client gave you is governed by your agreement with them, not by your AI plan. If it is under a confidentiality clause, that clause decides. ### Who is accountable for what it produces? Someone has to own the output. "The workflow" is not an answer anybody in legal will accept. ## What your manager will ask ### What does it replace? The task, how long it takes now, and how often. Hours a week is the clearest currency, and it is the same number you need to [judge whether a pilot worked](/guides/how-to-run-an-ai-pilot-that-survives/). ### What counts as better? Faster, fewer errors, or simply no longer the job nobody wants. Decide before it runs, so the judgement afterwards is not a matter of opinion. ### Who owns it, and what happens when they are away? A workflow only one person understands is a single point of failure with a nicer interface. Name a second person who could run it, or at least switch it off. ### How would you know it had gone wrong? Agentic workflows [fail quietly](/guides/how-to-notice-when-an-ai-workflow-is-wrong/), so "I would notice" is not enough. Name the check. ## Put it on one page All of the above fits on one page, and writing it is most of the work. A single note covering what it touches, whose access it uses, what it can do unasked, whose data flows through it, what it replaces and who owns it lets all three conversations happen at once rather than in sequence. Start with the smallest version that still earns its keep. A workflow that only reads and drafts, with a person sending the result, is far easier to approve than one that acts on its own. You can widen what it is allowed to do once it has a track record. ## Where this gets practical The security and stakeholder questions that decide whether a workflow sticks are part of the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) curriculum, alongside building one. You bring one real task, build the workflow against it, and leave with a concrete plan for introducing it where you work. Cohort 4 runs on [3 December 2026 at Worcester College](/events/oxford-agentic-bootcamp-cohort-4/). No coding required, and nothing hand-waved. --- ## Guide: How long does it take to build a second brain for AI? URL: https://oxfordagentic.com/guides/how-long-to-build-a-second-brain/ Updated: 18 November 2026 A working second brain for AI takes about a day of focused work, provided you build it from material you already have. The usual answer of weeks or months describes a note-taking system, which grows by capturing new material over time. A second brain for AI is written from what is already in your head and your documents, then kept current in minutes at a time. Ask how long a second brain takes and the standard answer is weeks. A week or so to set it up, a month for the habit to settle, several months before it pays back. The same guides tend to warn you not to build one in a sitting, because a system rushed in an afternoon gets abandoned by the weekend. That answer is right for what it describes, and it describes a different thing. ## Why the usual answer is weeks The standard second brain is a note-taking system. Its value comes from accumulation: you capture what you read and hear, file it, distil it, and months later you have a body of material worth drawing on. You cannot shortcut that, because the material does not exist yet. You are growing it. So of course it takes months. The time is the point. ## Why a second brain for AI is quicker A second brain for AI is not grown. It is written. What your AI tools need from you already exists: [how you write, what good looks like, the decisions you have made and why, the recurring facts of your business, and the shape of the work you repeat](/guides/what-goes-in-a-second-brain/). It is in your head, your sent email, your old proposals and your best documents. None of it has to be captured over months. It has to be extracted, written down so it stands on its own, and organised so a model can find it. That is bounded work. It is also concentrated work, which is why it goes faster in one focused day than in fifteen-minute slots spread over a month. Most of the effort is deciding what your context actually is, and that is much easier to do in one sitting than in fragments. ## What a day gets you, and what it does not A day, with your own material in front of you, is enough to leave with a second brain that is populated, organised and connected to the tools you use. It is not a template to fill in later. It does not get you everything you will ever want in it. That is fine, because it is meant to be curated rather than complete: a small, current, well-written store beats a large and stale one, and [the way these systems usually die](/guides/why-your-second-brain-stopped-getting-used/) is volume nobody reads. The day also only works if you bring the material. A second brain built from nothing produces a template, however well it is structured. ## The part that takes minutes, indefinitely The warning about rushed systems being abandoned is fair, and it points at the real long-term cost. It is not the build. It is keeping it true. That does not need a weekly review ritual. It needs an update at the moment something changes: a new price, a person leaving, a decision reversed, a standard raised. A minute each time, while you still have the context, is what stops [a second brain going stale](/guides/keeping-a-second-brain-from-going-stale/). So the honest total is one day to build, then a few minutes whenever the facts move. ## Where this gets practical The [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) is that day. You build your second brain from your own material, organise it so it stays findable as it grows, connect it to your tools, and finish with the routine that keeps it current. [Cohort 1 ran on 21 October 2026 at the Pitt Rivers Museum](/events/oxford-second-brain-bootcamp-cohort-1/), and dates for the next cohort are being confirmed. No coding required, and no months of capture first. --- ## Guide: How to choose the task you bring to an AI bootcamp URL: https://oxfordagentic.com/guides/how-to-choose-your-first-agentic-task/ Updated: 4 November 2026 The best first task for an agentic workflow is one you repeat, that runs on words and documents, whose output you can check at a glance, and that annoys you enough to care. Some work is worth leaving alone permanently: the writing where doing it is the thinking, and anything a person has to stand behind. "Bring one real task from your own work" is easy to say and quietly difficult to do. Asked cold, most people either reach for the hardest problem in their business - too big to build in a day - or something so trivial that automating it proves nothing. The task you choose decides what you walk away with, so it is worth choosing deliberately. If the format itself is new to you, [what an agentic bootcamp actually is](/guides/what-is-an-agentic-bootcamp/) covers the shape of the day the task is for. Five tests, in the order they matter. ## 1. You do it repeatedly A workflow pays back every time it runs, so frequency beats size. The monthly board summary, the weekly pipeline review, the supplier update you assemble every Friday - each is worth more than a one-off, however painful the one-off is. If you do it once a quarter or less, save it for your second build. ## 2. It runs on words and documents Agentic workflows are strongest where the raw material is text: notes, emails, reports, transcripts, spreadsheets exported to something readable. If your candidate task starts with "first I read through..." or "I pull together the...", it is probably right. If it starts with a system only you can access from one locked-down screen, or with a conversation only you can have, it will fight you all afternoon. ## 3. You can check the output at a glance The first build succeeds when you can look at what came out and know immediately whether it is right. Summaries, drafts, structured lists and comparisons all have that property - you have done the task yourself a hundred times, so your eye is already calibrated. A task whose output you would need an afternoon to verify is a poor place to learn, because you cannot tell whether the workflow is working. This matters because the pattern you are learning has a review step built in: a stage that checks the draft against a stated standard before you see it. To write that standard down, you have to know what good looks like. On a task you know well, you do. ## 4. It is annoying enough to care about You will hit a snag at some point - everyone does - and the difference between pushing through and drifting off is whether you actually want the result. The task that makes you sigh when it comes round again is the right raw material. Mild irritation is a better fuel than ambition here. ## 5. Getting it wrong is affordable The first task should be one where an imperfect output costs you a correction, not a client. Decisions about people, anything with personal data your organisation has not approved for the tools, and outputs that go straight to someone important without review are all wrong for a learning build - not forever, just first. You are learning to build the check; do not pick the one task where a missed check is expensive. ## The tasks to leave alone, permanently The five tests above are about picking a safe first build. There is a separate question that never goes away, and it is the one that separates people who get value from this from people who automate their way into a mess: not every task is worth automating, and some should not be. Working out the difference is most of the skill. Three kinds are worth protecting. **The ones where doing it is the thinking.** Some writing exists to work out what you believe. A strategy note, a difficult piece of feedback, the argument you are still forming. Hand that to a pipeline and you get a document without the thinking that was the actual output. You have not saved an hour; you have skipped it. **The ones you do rarely.** A workflow takes real time to build and a little to maintain. Something you do twice a year almost never repays that, and a half-remembered pipeline you have to debug before each use costs more than doing the job by hand. **The ones where a person has to stand behind the output.** Not because a workflow could not produce it, but because when someone asks "why does it say this", the answer has to be a person who decided, not a process that ran. None of this is squeamishness about the technology. It is the same judgement that decides what you delegate to people, applied consistently - and the reason the reclaimed hours are worth having is that they go back into exactly this kind of work. ## Can I bring confidential work to a build day? The advice the rest of the field gives is to bring sample or public data and clear the tool with your employer first. That is sound, and it also leaves you with nothing of your own to build against, which is most of the point of the day. The test worth applying is about the task rather than the data: would you be comfortable describing this task to an external supplier? If you would, bring it. If you would not, pick another one, and most people have several that qualify. Credentials, keys and other people’s personal data do not belong in an agentic workflow in any form, whoever is running the day. Two things decide the rest, and they are worth asking any provider before you book: whose machine you build on, and whether the day is recorded. At the Oxford Agentic Bootcamp you build on your own laptop and the day is not recorded. What your own side will ask once the workflow exists is a separate conversation, and [getting an AI workflow past IT and legal](/guides/getting-an-ai-workflow-past-it-and-legal/) sets out the questions each of them tends to bring. ## What if the task is fine but the data is sensitive? You bring a task, not a data set, and that distinction does most of the work. A workflow can be built against a redacted or representative extract and pointed at the real source afterwards, once it is back inside your own controls, because what you are building on the day is the pipeline rather than any one output. It also keeps the build easy to check while you are still learning, which is test 3 above. If what is sensitive is a body of documents rather than a process, [is it safe to put company documents in a second brain](/guides/is-it-safe-to-put-company-documents-in-a-second-brain/) answers the same question for the storage side. ## What this looks like in practice Put together, the tests point at a familiar shape: the recurring assembly job. The report built from the same sources every time. The meeting notes that become minutes and actions. The inbound enquiries that need triaging into three piles. The draft that always starts from the last version. Dull, frequent, text-based, checkable - and collectively hours of your week. Bring one of those, and by the end of a build day you have it running: a pipeline against your own task, a review step that checks the output, and a written recipe so you can rebuild and extend it afterwards. That is the deal at the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/), and the choice of task is the half of it you can get right before you arrive. --- ## Guide: Agentic workflows fail quietly. Here is how you would notice URL: https://oxfordagentic.com/guides/how-to-notice-when-an-ai-workflow-is-wrong/ Updated: 17 August 2026 An agentic workflow rarely breaks in a way you can see. It returns something plausible, in the right format, at the right time, and quietly stops being correct. Noticing takes deliberate checks rather than attention: a written standard, a known answer you re-run, a spot-check habit, and knowing which step failed. Ordinary software has the decency to fail loudly. It throws an error, the page breaks, something turns red, and you know. An agentic workflow does not do that. It hands you a document in the right format, at the right time, of the right length, that is confidently and quietly wrong. That difference is the whole problem. Nothing alerts you, because from the machine's point of view nothing went wrong: every step ran and produced an output. The failure is in the content, and content has no exit code. ## Why the odds get worse as you add steps A workflow of narrow steps is more reliable than one clever prompt, which is the reason the pattern is worth learning. But it comes with arithmetic worth understanding: the steps multiply rather than average. Each stage is mostly right, and "mostly" compounds down the chain, so a longer pipeline is less trustworthy end to end than any single step in it looks. This is not an argument for fewer steps. It is an argument for a check at the end, and for keeping the chain no longer than the job needs. ## Four things that actually catch it **Write the standard down before you build.** The most useful artefact in the whole workflow is a paragraph describing what a good output looks like: what it must contain, what it must never do, what would make you send it back. Without that written down, the reviewer step has nothing to check against, and "does this look right" quietly becomes "does this look like the last one". **Keep one input whose answer you already know.** Run it through occasionally. If the output drifts, something upstream has changed - a model update, a source that reorganised itself, a document that is no longer where it was. This is the cheapest early warning available, and almost nobody sets one up. **Spot-check on a schedule, not on a feeling.** Confidence in a workflow grows faster than its reliability does. After a fortnight of good output the temptation is to stop looking, which is exactly when a silent change starts costing you. One in ten, every week, opened properly. **Know which step failed, not just that it did.** When something is wrong, the useful question is which stage produced it: did the scout gather the wrong material, did the drafter ignore good material, or did the reviewer wave through something it should have caught. A workflow you cannot decompose is one you can only throw away and rebuild. ## What good failure looks like A workflow that fails well is one that stops rather than guesses. If the source document is missing, the honest output is "I could not find the input", not a confident summary of nothing. Most of the design work in a reliable pipeline is deciding, in advance, what each step should do when its input is not what it expected. That is unglamorous and it is most of the difference between something you use on Monday and something you abandon by Thursday. ## Where this gets practical None of this needs code, and none of it is hypothetical - it is the part of a build that people skip when they are excited, and the part they wish they had not skipped a month later. At the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) the review step is built alongside the workflow rather than after it, because a pipeline without one is a draft generator that nobody is checking. No coding required, and nothing hand-waved. --- ## Guide: How to run an AI pilot that survives contact with your team URL: https://oxfordagentic.com/guides/how-to-run-an-ai-pilot-that-survives/ Updated: 18 November 2026 AI pilots rarely fail on the technology. They fail because the pilot was designed to evaluate a tool rather than change a specific task, and because nobody who actually does the work owned it. Pick one task, give it one owner who does that task, decide in advance what better looks like, and plan for week two. Most AI pilots do not collapse because the model was not good enough. They collapse quietly, a fortnight in, when the person who was enthusiastic in the kick-off goes back to doing the job the old way because the old way is what fits their Tuesday. That is an organisational failure, not a technical one, and it is avoidable. ## Pilot a task, not a tool The most common design mistake is scoping the pilot around software: three months to evaluate a platform, a committee, a report at the end. Nothing about that produces a changed way of working, and everyone can feel it. Scope it around one task instead. Not "explore AI in operations" but "the weekly supplier summary that takes someone half of every Wednesday". A task has an owner, a frequency, a current cost and an obvious before-and-after. A tool has a vendor. ## Give it to someone who does the work The pilot should be built and owned by the person whose task it is, not by an innovation team building on their behalf. Two reasons. They know the exceptions and the judgement calls that never made it into any process document, and those are exactly what a workflow gets wrong. And a workflow handed to someone as a finished object gets abandoned the first time it misbehaves, because they have no idea what is inside it. If that person cannot build it, that is worth knowing early. It is a stronger argument for teaching them than for building it for them. ## Define "better" before you start Write down, in advance, what the task costs today and what would count as an improvement. Hours, error rate, turnaround, or simply "it stops being the thing they dread on a Wednesday". Vague success criteria always resolve to "it was interesting" and then to nothing. Be honest that the comparison is against how the job is done now, not against perfection. A workflow that gets a summary eighty per cent right in four minutes may beat a person doing it perfectly in three hours, or may not - but you cannot tell without having written down which one matters here. Be precise about which kind of better you are claiming, too. On small tasks AI has been measured to reduce effort without reducing clock time, and people cannot reliably tell the two apart: [saving time, or just feeling faster](/guides/is-your-ai-workflow-saving-time-or-just-effort/). ## Build the check into the workflow The reason pilots lose trust is a bad output that nobody caught. A pipeline that drafts and then explicitly reviews its own draft against a stated standard catches far more than one that drafts and stops, and it leaves you a record of what was checked. Where the output is wrong in a way that would be expensive and hard to spot, that is a signal to keep the human step, not to try harder. ## Plan for week two Week one has attention. Week two has the rest of the job. This is where most pilots quietly die, so decide up front what happens there: who notices if it stops being used, when you look at it again, and what small fix is allowed without another approval round. A pilot with a review date survives. One that ends with "let us see how it goes" does not. And keep the pilot small enough that it can succeed. One task, done properly, changes more minds in an organisation than a broad programme that produces a slide deck. ## Where this gets practical All of this is easier to agree with than to do, and the hardest part is the first working version. That is what a day at the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is for: bring the real task, build the workflow yourself with someone on hand when you get stuck, and leave with it running rather than with a plan to run it. No coding required, and nothing hand-waved. --- ## Guide: Should you run an in-house AI day, or send people to a cohort? URL: https://oxfordagentic.com/guides/in-house-ai-day-or-send-people-to-a-cohort/ Updated: 10 November 2026 An in-house day suits a whole team that shares systems and needs one agreed way of working. A cohort suits a few senior people who need capability rather than consensus, and who gain from hearing how others solved the same problem. The deciding question is whether you are buying alignment across a team or judgement in a handful of people. If you have been asked to sort out AI training, you will have noticed that almost every provider sells the same shape: they come to you, for a day, for your team. It is a good format. It is also the only one most of them offer, which means the honest comparison with the alternative rarely gets written. The alternative is sending two or three people to an open cohort and letting what they learn spread from there. Neither is better. They are good at different things, and the choice is usually decided by something other than price. ## What an in-house day is good at - **Everyone ends up with the same way of working.** If the problem is that six people each invented their own approach, a shared room on a shared day fixes it in a way that sending one person never will. - **Your own systems, in the room.** Real data, real permissions, real constraints, and whoever administers your tools sitting two seats away when something needs approving. - **Confidential work stays inside.** Some tasks cannot be discussed in front of strangers, and that settles it on its own. - **No travel, and one date to agree.** For a team of twelve, that is not a small thing. ## What a cohort is good at - **Nobody has to organise it.** No minimum headcount, no venue, no scoping call. Someone books a seat. - **The other people in the room.** This is the part that does not survive being brought in-house. You hear five other ways to approach your problem, from people with no stake in your politics, and some of them are better than yours. - **It suits small numbers honestly.** Below about four people, a private day is usually poor value, and most providers will still happily sell you one. - **Being out of the building.** A day at your own desk is a day where people get pulled into their normal work at eleven o'clock. ## The question that settles it Ask what you are actually buying: **alignment, or capability?** If the goal is that a team stops doing something inconsistently, you are buying alignment, and alignment is made in a room together. If the goal is that two or three capable people come back able to build things and show others, you are buying capability, and capability travels home in individuals. A second test, if the first is close: would the day still be worth it if only half the invited list turned up? For a cohort the answer is yes, because each person's value is their own. For an in-house day it is usually no, which tells you how much of its value depends on getting everyone in the room on one date. ## Three questions that come up when the choice is close ### How many people does it take before an in-house day is the better buy? There is no clean threshold, and the honest version of the answer is a floor rather than a number: below about four people a private day is usually poor value, and most providers will still happily sell you one. Above that, headcount does not decide it on its own. Six people who have each invented their own way of working are a case for one room on one date. Six people who would each build something different are six separate pieces of capability, and that is what a cohort seat buys. Work out which of the two you are paying for, then count heads. ### Should we send two people to a cohort first, and decide on an in-house day afterwards? It is a reasonable order to do it in, and it answers the one thing a scoping call cannot: whether the material is worth putting in front of your whole team. Two things make it work. Send people who own a task rather than whoever has the clearest diary, because every seat is spent on a real task the attendee brought - [who goes is a question of its own](/guides/who-should-take-the-seat/). And if two of you come, bring the same task, so one workflow goes back with both of you. What the trial will not tell you is whether your team would end up agreeing on one way of working, because that is the thing only a shared room produces. ### Should the whole team come, or only the people who will build? On a cohort day, only the people who will build. Every seat is spent on a task that person brought, so anyone sent along to watch spends the day on a hypothetical, and the published list of who the day is not for puts it plainly: this is a build day, not a briefing. If the reason you want everybody in the room is that everybody has to agree how the work gets done, that is the alignment case above, and an in-house day is the right instrument for it. Otherwise send the builders, and book an hour in the same week to be shown the thing running rather than told about it. ## Where this gets practical The [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is the cohort version: one day in Oxford, a small capped room, and everyone builds a working agentic AI workflow against a real task they brought. Teams are welcome, and if two of you come, bring the same task so the same workflow goes back to work with both of you. Email hello@oxfordagentic.com and we will arrange it. If what you need is alignment across a larger team, be honest about that and buy the in-house version from someone who does it well. A cohort seat is the wrong instrument for that job, and you would be back in six months. --- ## Guide: Is it safe to put company documents in an AI second brain? URL: https://oxfordagentic.com/guides/is-it-safe-to-put-company-documents-in-a-second-brain/ Updated: 18 November 2026 Usually yes, for most of what a second brain actually needs - your own writing, standards, decisions and recurring facts. It depends on three things: which account the material goes into, what category of information it is, and whose information it is. Personal data about other people and anything under a client contract need their own answer, and credentials never go in. Most advice about AI and company documents is about a single moment: should I paste this contract into a chat? A second brain raises the stakes slightly. You are not pasting something once. You are storing it somewhere your AI tools will read again and again, across tools and over months. That makes the decision worth taking deliberately, once, rather than at four o'clock on a Thursday with a deadline in front of you. The good news is that most of what a useful second brain needs is the easy kind. ## Question one: which account is it going into? The account matters more than the tool. The terms that govern your material differ between personal and business plans, and they change, so nobody's summary of them - including this one - is a substitute for reading the terms for the exact plan your organisation pays for. Do that once and write it down: whether your inputs are used for training, how long material is kept, and who can access it. Copy the actual sentences, with the date you read them. It takes ten minutes and settles most of the argument. The rule that follows is simple. Work material goes into the work account your organisation has approved. Never a personal one, however convenient. ## Question two: what kind of information is it? Sort anything you might add into four groups. - **Public.** Already published or publishable: your website, public prices, articles, industry material. No restriction. - **Internal, not personal.** Your own drafts and notes, how you write, what good looks like, decisions and the reasons for them, the recurring facts of the business. Fine on an approved account. - **Confidential or personal.** Named customers, anything about identifiable people, salaries, unreleased figures. Redact it, or get an explicit decision before it goes in. - **Regulated or contractually fenced.** Health data, payment details, national identifiers, anything a law or a contract says how to handle. Not without a written decision from whoever owns that risk. The second group is where almost all the value of a second brain sits, and it is the group people wrongly treat as forbidden. [The five kinds of context worth storing](/guides/what-goes-in-a-second-brain/) are overwhelmingly internal and not personal: your voice, your standards and your reasoning, none of which is anyone else's personal data. ## Question three: whose information is it? A document from a client is governed by your agreement with that client, not by your AI plan. An enterprise plan can answer the question of how the provider treats your data and still leave this one open. If a client's material is under a confidentiality clause, the clause decides, and "I put it in my second brain" is rarely something it anticipated. The practical answer is usually to store the lesson rather than the document. What you learned about how that client works, written in your own words and without identifying detail, is often more useful to your AI tools than the original document would have been. ## Two habits that carry most of the load **Scope access narrowly.** When a tool connects to your files or drive, grant it the folder your second brain lives in, not everything you have. A connector reading your whole drive will find things you forgot were there. **Keep credentials out, always.** Passwords, keys and access codes do not go in a second brain in any form, on any account. There is no version of that which is worth the risk. ## Where this gets practical The [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) is built from your own material, so it is worth running this sort before you arrive and bringing what passes it. The day is spent turning that material into a second brain your AI tools can draw on, organised so it stays findable and set up to maintain. [Cohort 1 ran on 21 October 2026 at the Pitt Rivers Museum](/events/oxford-second-brain-bootcamp-cohort-1/), and dates for the next cohort are being confirmed. No coding required, and nothing in it you did not choose to bring. --- ## Guide: Is your AI workflow saving time, or just feeling faster? URL: https://oxfordagentic.com/guides/is-your-ai-workflow-saving-time-or-just-effort/ Updated: 13 October 2026 On simple tasks, AI assistance saves effort rather than time. In a preregistered study of 1,237 people, assisted and unassisted completion took the same time, yet people predicted AI would be much faster and reported it felt easier. A multi-step workflow is a different proposition, which is why you measure the whole job rather than the step that impressed you. Everyone who has built something with AI has had the same moment: the thing runs, the output appears, and it feels like you have got an hour back. The uncomfortable question is whether you have. ## The illusion has been measured In May 2026, Sunny Yu, Myra Cheng and five co-authors published two preregistered studies on exactly this. The first, [Cognitive offloading and the speedup illusion in human-AI interaction](https://arxiv.org/abs/2605.23177), ran with 1,237 participants on simple cognitive tasks and appears in the proceedings of the Cognitive Science Society's 2026 meeting. Its central result is blunt: actual completion times with and without AI assistance did not differ, while participants predicted that AI would be significantly faster. The interesting part is where the error sits. People estimated their own unassisted times accurately. What they got wrong was the assisted time, which they consistently underestimated. The authors call this a speedup illusion, and note that the same bias did not appear when participants imagined being helped by another person. Something about handing work to a machine, specifically, makes us expect the clock to move. The companion paper, [The efficiency-gain illusion](https://arxiv.org/abs/2605.22687), ran three preregistered studies across 2,691 people and found that we reach for AI on small tasks even when it saves neither time nor effort - and that using it once makes us more likely to use it again, entrenching the miscalibration rather than correcting it. ## Effort and time are not the same saving The finding that should change how you talk about this is the second one in that first paper: people reported lower subjective effort with AI despite equivalent completion times. The authors' conclusion is that completion time alone "is not sufficient to characterize efficiency gains". That is not a debunking. Effort is a real cost. A job that used to take forty minutes of concentration and now takes forty minutes of supervision is genuinely easier, and at the end of a long week that matters. But it is a different claim from "it saves me half a day", and only one of the two will survive a conversation with your finance director. So say which one you got. Time, effort and quality are all legitimate returns. They are not interchangeable, and conflating them is how an AI project ends up with a number nobody can defend. ## What the research does not say Both papers are explicitly about simple tasks - arithmetic, spell-check, short questions, the things you finish in a couple of minutes. Neither tests a multi-step workflow running against a real job over weeks. So this is not evidence that AI does not save time. It is something narrower and more useful: the smaller the task, the more likely the saving you feel is effort rather than clock time, and the less you should trust your own estimate. That bound matters, because the shape of the work changes the answer. A single prompt replacing a two-minute task has very little room to win once you count reading the output. A workflow that gathers, drafts and checks, running on something you do every week, is a different proposition entirely - and the honest way to find out is to measure it rather than to assume either way. ## What to measure instead Four things, none of which take long. - **Time the whole job by hand, once, before you build.** Not the step you intend to automate. The job, from the thing existing to the thing being finished properly. You need this number and you will not be able to reconstruct it afterwards. - **Measure the third run, not the first.** The first includes your learning. The third is the one you will actually live with. - **Count the checking.** Reviewing, correcting and the occasional run you throw away are part of the cost. If verifying the output needs the same expertise as producing it, be honest that the saving may be small. - **Measure early, before it becomes habit.** This is the practical consequence of the carryover finding: confidence in a saving grows with use whether or not the saving is real. A workflow with a review step in it will usually look slower than a single clever prompt, and be worth considerably more. That is the point of [the scout, drafter, reviewer shape](/guides/scout-drafter-reviewer-pattern/) - and it is also why measuring only the drafting step flatters almost any build. ## Where this gets practical If you are choosing something to build against, the tests in [how to choose your first agentic task](/guides/how-to-choose-your-first-agentic-task/) point at the work where a real saving is most likely: frequent, text-based, and checkable at a glance. Frequency is what turns a modest per-run saving into a number worth having. The [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is a day spent building one of these against a task you brought, which means you leave with the thing running and the before-and-after still fresh enough to measure. Take the baseline before you come. --- ## Guide: A stale second brain is worse than none at all URL: https://oxfordagentic.com/guides/keeping-a-second-brain-from-going-stale/ Updated: 24 August 2026 A second brain decays whether or not you notice: standards change, people leave, decisions get superseded. In a note system that is clutter you skim past. In one your AI reads it is worse, because the model treats a two-year-old position as current and answers with it. The fix is a review routine, not a bigger vault. Search for why second brains fail and you get the same answer: they become graveyards. People capture more than they ever read back, the vault swells, and the habit dies somewhere around note four hundred. That is a real problem, and it is not the one that matters here. It describes a system with a human reader, where the cost of a stale note is that you scroll past it. Once a model is reading your context instead, staleness stops being clutter and starts being wrong answers. ## The model cannot tell what is out of date You know your old pricing is old. You remember changing it, you would never quote it, and if you saw it in a folder you would ignore it without thinking. A model has none of that. It reads what it is given, treats a superseded document with exactly the confidence it treats a current one, and produces an answer built on the wrong thing. Worse, the more carefully you wrote that old document, the more authoritative it looks. This is the same failure shape as [a workflow that breaks quietly](/guides/how-to-notice-when-an-ai-workflow-is-wrong/): nothing errors, output keeps arriving, and the only signal is that it is subtly wrong in a way you have to already know the answer to catch. ## Four things that go stale, in the order they bite **Positions you have changed.** Pricing, scope, what you will and will not take on. These are the most damaging because they are the ones a model most confidently repeats. **People and roles.** Who does what, who to ask, who owns a relationship. Teams move faster than anyone updates a document about them. **Standards.** What good looks like drifts as you get better. A quality bar written eighteen months ago quietly holds your output to a lower line than you would accept today. **Anything with a date in it.** Last year's plan, the current quarter's targets, the roadmap. These do not decay gradually - they flip from useful to actively misleading on a known date, which is what makes them the easiest to handle. ## The routine, which is smaller than it sounds **Review on a trigger, not a calendar.** A monthly audit is a chore nobody keeps. The things that actually work are event-shaped: when you change a price, when someone leaves, when you lose or win work for a reason worth recording. The trigger is the decision, and the update takes a minute at the moment you already have the context. **Date anything that expires, and say what supersedes it.** A document that opens "as of March 2026, superseded by the Q3 note" is safe to leave in place. One that just stops being true is not. This is the single highest-value habit here and it costs one line. **Delete rather than archive.** An archive folder your model can still read is not an archive. If it is wrong, it goes - the version history has it if you ever need it back. **Re-ask a question you know the answer to.** Once a month, ask your assistant something about your own business where you already know what good looks like. If the answer has drifted, something underneath it has gone stale. It is the cheapest test available and almost nobody runs it. ## Why this is the half people skip Building a second brain is satisfying. Maintaining one is not, which is why most of the writing about them stops at capture and organisation, and why most of them are quietly untrustworthy within a year. The routine is what turns it from a project into something that compounds - and it is why the [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) ends with how you will keep it rather than with what is in it. You leave with the thing populated from your own material, and with the habit that stops it decaying. No coding required, and nothing hand-waved. --- ## Guide: Ten questions to ask before booking AI training URL: https://oxfordagentic.com/guides/questions-to-ask-before-booking-ai-training/ Updated: 28 October 2026 Ask what you will personally have running at the end, whose work you build against, how many people are in the room, what happens when you get stuck, and what the day deliberately does not cover. A provider who answers those five plainly is usually worth the day. Vagueness on the first two is the warning sign. Most AI training now describes itself in the same words: practical, hands-on, applied, built for leaders. The descriptions have converged, so they no longer help you choose. The questions below do, because the answers still differ enormously between providers. None of them are trick questions. A provider who has thought about their format will answer each one in a sentence. ## 1. What will I personally have running at the end? The single most useful question, and the one that separates formats faster than any other. Answers range from "notes and a framework" through "a demo you watched someone build" to "a workflow of your own, running". All three are legitimate products. Only you know which one you are paying for, so make them say which it is. If the answer is a list of things you will "understand" or "be able to", that is a knowledge product, not a build one. The word itself is the problem. "Hands-on" covers [four quite different formats](/guides/what-hands-on-ai-training-actually-means/), and this question is usually what settles which one you are being sold. ## 2. Whose work do I build against - mine, or a case study? A prepared case study is easier to run and much easier to get right. It is also the point at which most training stops transferring, because the hard part of applying AI is the mess in your own work, not the tidy version in the exercise. Ask whether you bring a real task, and what happens to it during the day. If you do bring one, ask what makes a good one, and take the answer seriously - it usually determines how much you get out of the day. ## 3. How many people are in the room, and who is helping them? Cohort size decides what the format can actually be. Above roughly thirty people, a day cannot be facilitated in any meaningful sense; it becomes a lecture with exercises, however it is billed. Ask for the cap, not the typical number, and ask how many people are circulating to help. ## 4. What is the ratio of talking to building? Ask for it in minutes. A day described as hands-on can still be five hours of presentation with a workshop after lunch. A straight answer sounds like a running order: this long building, this long between builds, longest block of anyone talking is this. ## 5. What happens when I get stuck? Everyone gets stuck, usually on something dull - a login, a permission, a tool that behaves differently on their machine. Whether that costs you ten minutes or the afternoon is a property of the format, not of you. Ask who walks the room. ## 6. What do I need to bring or set up beforehand? Two things worth knowing early: whether you need a paid AI subscription, and whether any setup is expected before you arrive. Neither is unreasonable. Both are annoying to discover the night before. ## 7. Who else will be in the room? Peer level matters more in a small room than the syllabus does. A day of founders and directors working on their own operations is a different experience from a mixed room of engineers and analysts, even with identical content. Ask who the day is designed for, and who actually came last time. ## 8. What does the day deliberately not cover? The most revealing question on this list. Any honest one-day format has a scope decision behind it, and the provider should be able to name what they left out and why. A day that claims to cover the landscape, the tooling, the governance and a build is describing four days. ## 9. What if I book and then cannot make the date? Diaries move, and this is the question most buyers only think to ask afterwards. The cancellation policy is usually the last line on a course page and the first thing that matters once a client meeting lands on the same morning. "Transferable" on its own does not tell you much. Three things decide whether it means anything. Does the place move to another date at all, or is the fee simply gone? Is there a charge for moving it, and is that a flat admin fee or a percentage of what you paid? And is the move gated by a notice period - often counted in working days rather than calendar days, which can close the window a week earlier than you would expect. Ask for the answers in writing before you pay rather than after. If someone else is approving the spend, [they will ask the same thing](/guides/getting-an-ai-bootcamp-approved/), and a clear answer removes the easiest reason to say not yet. It is a small question that tells you how a provider treats the people who have already paid. ## 10. What do I get after the day, and for how long? The question almost nobody asks until the following week. One-day formats differ more here than anywhere else on this list, and the difference is rarely on the booking page. Some end at 17:00, and that is the whole offer. Others send a written pack afterwards, keep a channel open for a week or two, or book a short call once you have tried the thing on your own work. Those are different products at a similar price. Two halves of the answer are worth separating. The first is what you take away: the thing you built, in your own hands rather than inside the provider's account, plus whatever is written down so you can rebuild it. The second is what happens when it breaks a fortnight later, which is usually when people actually need the help. Ask for both in writing, because "ongoing support" covers everything from a mailing list to a named person who answers. A provider who says the day ends on the day has given you a fair answer, and an easier one to plan around than a vague promise. What you want to avoid is finding the boundary at the moment you have crossed it. ## Using the checklist You do not need perfect answers to all ten. You need answers that are consistent with each other: a room capped small enough for the help promised, a build ratio consistent with the result claimed, a scope narrow enough to be delivered in the time. Most mismatches show up when you put two answers side by side. For the record, here are ours. We run three one-day bootcamps in Oxford, and several of the ten answers differ between them. The [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is one day in Oxford, capped at a small room, built around build cycles rather than talks, with no block longer than twenty-five minutes of anyone talking at you. You bring one real task from your own work and a laptop, and you leave with a working agentic workflow running against that task, plus the pattern behind it. No coding is required. A paid AI plan helps but is not essential. The day is not recorded, because the value is in the room. If you cannot make the date, your place transfers to a future cohort at no charge: let us know and we will move you. After the day you keep the workflow and a written recipe for rebuilding and extending it, and you leave with the group: everyone in the room has built the same thing, you go away with their contact details, and you hear first when the next one is on. What that day deliberately does not cover: the AI landscape, model comparisons, and governance policy. One pattern, applied properly to one real task, is what fits in a day. The [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) runs in a small, capped group for people who already use AI most days and have hit the ceiling of what prompting alone can do. You bring a laptop and your own material - the documents, notes and examples that carry how you work and what you know - and build a second brain out of that rather than from a template, so you leave with it populated, organised and set up to maintain. Nothing technical is required: the work is thinking clearly about your own business rather than writing code. What it is not is a conceptual overview of AI, a vendor pitch or a tutorial in one app. The [Oxford Agentic Coding Bootcamp](/bootcamps/agentic-coding-bootcamp/) is for people who understand the concepts but have not yet built and shipped something of their own. It does not assume you can code - Claude Code and Codex do the typing - and what it teaches is everything around the code: the brief the agent reads first, the terminal, version control, deployment, databases and testing, and how to check what the agent wrote rather than trusting it. You leave with a personal website live on the internet and a working prototype that saves real data, both in your own GitHub repository. Question 6 is the one that matters most here: bring a laptop you can install software on, because it involves setting up a code editor, a terminal and a GitHub account on your own machine, and corporate or government devices often block that. If yours is locked down, email hello@oxfordagentic.com before you book and we will work out what you need. All three run 09:00 to 17:00, in person in Oxford, and all three answer question 9 the same way: places are not refundable, but if the date stops working your place transfers to a future cohort at no charge. --- ## Guide: Scout, drafter, reviewer: the workflow pattern that does real work URL: https://oxfordagentic.com/guides/scout-drafter-reviewer-pattern/ Updated: 18 November 2026 Scout, drafter, reviewer is the workflow shape behind most useful agentic AI: one step gathers the raw material, one produces a first version, and one checks it against a stated standard. Three narrow steps beat one clever prompt because each can be tested, improved and trusted separately, and the pattern transfers to almost any repeated knowledge task. Ask one AI prompt to research, write and polish a report and you get something plausible, mediocre and unaccountable. Ask three narrow steps to do it - one gathering, one drafting, one checking - and you get something you can actually use, improve and trust. That three-step shape is the pattern we teach first, because it fits an enormous amount of knowledge work. It is worth understanding properly. ## The scout: gather before you generate The scout's only job is to collect the raw material: pull the figures, find the documents, extract the relevant passages, list what changed since last time. It produces no prose and makes no judgements - it assembles the evidence the next step will work from. Separating this out matters because most bad AI output fails at the input, not the writing. A model asked to draft from nothing invents; a model asked to draft from a scout's evidence pack has something true to say. When a workflow goes wrong, the scout's output is the first place to look - and because it is a separate step, you can look. ## The drafter: one version, from the evidence The drafter turns the scout's material into a first version of the thing you actually need: the summary, the report section, the reply, the comparison. It works only from what the scout gathered, in the shape and voice you have specified. Keeping the drafter narrow is what makes it good. It is not deciding what matters or checking itself - it is doing the one thing models do best, turning structured material into clear prose, with the ingredients already on the bench. ## The reviewer: check against a standard, in writing The reviewer reads the draft against a stated standard before any human sees it: is every claim supported by the scout's material, does it follow the format, did it answer what was asked, does it break any of your rules. It flags or fixes what fails. This is the step people skip, and it is the step that makes the whole thing trustworthy. A pipeline that drafts and stops produces work you must check line by line, which quietly costs the time you saved. A pipeline that checks its own draft - and shows you what it checked - produces work you can review at a glance. Writing the standard down is also clarifying in itself: it forces you to say what good actually looks like, which most of us have never put into words. ## Why three narrow steps beat one clever prompt Three reasons, all practical: - **You can see where it went wrong.** One giant prompt fails as a blob; a pipeline fails at a step, and you fix that step. - **Each step can be improved alone.** A better scout raises the quality of everything downstream without touching the drafter. Improvements compound. - **Trust is earned per step.** You will trust the scout's gathering long before you trust anything's judgement. The pattern lets you hand over the parts you are ready to hand over and keep the rest. The pattern also places the human deliberately: after the reviewer, at the point of judgement - approving, correcting, deciding - rather than in the middle, retyping context between tools. ## Where it fits, and where it does not Anything with the shape *gather, produce, check* is a candidate: the weekly report, the meeting minutes, inbound triage, first-draft anything, the comparison you rebuild every month. That is most repeated knowledge work. It is the wrong shape where the output is a judgement about a person, where being subtly wrong is expensive and hard to spot, or where the task runs on information you cannot give the scout. Those are the tasks to keep, not to force into the pattern. ## Will this still be true when the tools change? The tools move every few months, so the question is a fair one: how much of this has a shelf life. ### Does what I learn go out of date in six months? Some of it does, and it is the part that matters least. Which agent you opened, where the button sat, what the menu was called in April - all of that moves, and quickly. The shape does not, because gather, produce, check describes the work rather than a product. A report that needs its evidence assembled before anyone writes prose needed that before these tools existed and will need it after they have been replaced. What the pattern actually makes you do is write three things down - what the scout may use, what the drafter is for, and the standard the reviewer checks against - and not one of them is an instruction for a particular tool. It is also why no tool is being sold on the day itself. A day built around a product dates when the product does. ### If the models keep getting better, will one prompt not just do it? The models have got better and the split has not stopped earning its keep, because the reasons for it were never mainly about the model being weak. You keep the steps apart so you can see which one went wrong, improve one without disturbing the others, and hand over the parts you have come to trust while keeping the rest. A better model makes each of the three steps better and leaves all three reasons standing. The honest version: for a small job nobody else depends on, a strong model can do the lot in one pass, and you should let it. From the moment someone else relies on the output, you will want to know which step produced the sentence you are being asked about. ### Should I wait until the tools settle down? They are not going to settle, so waiting is a decision to carry on without anything running rather than a decision to get something better later. It also improves the half that was never the constraint. What decides whether one of these is worth the effort is the task you point it at, which is your own and is not changing, and [choosing that task well](/guides/how-to-choose-your-first-agentic-task/) is the part most people get wrong. There is a longer version of the same argument about [waiting for the right cohort](/guides/when-ai-executive-cohorts-run/), and it ends in the same place. ## Where this gets practical Reading about the pattern and having one running against your own work are different things, and the second one is a day's work with someone on hand. That is what the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is for: you bring one real task, build the scout, the drafter and the reviewer against it, and leave with the pipeline running and a written recipe to rebuild it. No coding required, and nothing hand-waved. --- ## Guide: A second brain for AI is not the note-taking kind URL: https://oxfordagentic.com/guides/second-brain-for-ai-vs-note-taking/ Updated: 4 November 2026 “Second brain” usually means a personal note system, with a human doing the reading. A second brain for AI serves a different reader: a model with no memory of you and no way to ask a follow-up. Notes made for yourself rarely work unedited, because they assume context the model does not have. If you have met the phrase "second brain" before, you probably met it as a note-taking method. The term was popularised by Tiago Forte's book of that name, and it describes a personal knowledge system: capture what you come across, organise it so you can find it again, and draw on it when you write or decide something. PARA, Obsidian, Notion, a decade of YouTube videos about folder structures. That is a real and useful practice. It is not what this is. A second brain for AI has the same ambition - your knowledge, stored outside your head, working for you - but a different reader. And the reader turns out to change almost everything about how you build it. ## The difference is who does the retrieving A personal note system is written for you. You already know what you meant. You can skim, skip, recognise a half-finished thought, and remember the conversation a note came out of. The system only has to get you back to the right note; your own memory does the rest. An AI has none of that. It reads what you give it, cold, with no history and no ability to ask what you meant. Everything your notes were quietly relying on has to be on the page. Four consequences follow, and they are the practical difference between a second brain that improves your AI's answers and one that does nothing. ### Notes have to stand on their own Personal notes are full of shorthand: "same as the Acme approach", "the usual caveats", "per Tuesday's call". Perfectly clear to you. To a model, "the Acme approach" is a phrase with no referent, and the honest thing it can do is guess. Anything you want the model to use has to carry its own context. Not longer, just self-contained. ### Structure stops carrying meaning In a personal system, where a note lives is part of what it says. A note in `Clients/Acme/Pricing` does not need to say it is about Acme's pricing, because the path says so. AI tools tend to read a passage, not a hierarchy. Meaning that lives in the folder name usually does not travel. It has to be in the text. ### Volume works against you Skimming is free for you. You open a folder of forty notes, ignore thirty-eight, and lose nothing. A model does not ignore things. Feed it three contradictory drafts of your pricing policy and it will average them, or pick one, and you will not know which. In a personal system, an archive is harmless clutter. In an AI-facing one, every stale document is an active source of wrong answers. This is why a second brain for AI is a curated supply rather than a complete record - the value is genuinely in what you leave out. ### Nothing tells the model a note is out of date You know your old positioning is old. You remember changing it. The model has no way to tell, so a superseded document reads as current, and the more confidently it is written the worse the problem. So supersession has to be explicit: dates on things that change, old versions removed rather than left lying around, and a routine that actually gets run. This is the part most systems skip and the reason most quietly stop being trusted. ## Do you need both? They serve different jobs and can share material, but neither replaces the other. If your problem is that you read a lot and cannot find things again, a personal note system solves that and this day is not what you need. If your problem is that you use AI daily and every conversation starts from zero - re-explaining your business, correcting the same wrong assumptions, rewriting output that sounds like nobody - that is a supply problem, and better note-taking will not fix it. Plenty of people arrive with years of notes already. That is a good starting position and it is not a finished second brain, because the notes were written for a reader who already knew everything. The work is deciding what a model actually needs and rewriting the useful parts so they survive being read cold. ## Where this gets built Knowing the distinction is not the same as having the thing. That gap is what the [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) closes: one day in Oxford, your own material, and you leave with a second brain populated, structured so it stays findable as it grows, and connected to the tools you use. No coding, and no requirement to have kept good notes beforehand. --- ## Guide: What actually goes in a second brain for AI? URL: https://oxfordagentic.com/guides/what-goes-in-a-second-brain/ Updated: 18 October 2026 A second brain for AI holds five kinds of context: how you write, what good looks like, the decisions you have already made and why, the recurring facts of your business, and the shape of the work you repeat. It is a curated supply, not an archive - and credentials and other people’s personal data stay out of it. "Build a second brain" is easy to nod at and hard to start, because the obvious next question has no obvious answer: what do you actually put in it? Dump everything and you have a worse search engine. Put in too little and the answers stay generic. The useful middle is narrower and more specific than most people expect. If the term itself is doing too much work, [what a second brain actually means](/guides/what-second-brain-actually-means/) settles that first. Five kinds of context earn their place. ## 1. How you write Not a style guide - actual examples. Three or four pieces of your own writing that you were happy with, in the forms you produce most: the client update, the board note, the difficult email. A model shown four real examples of your voice will match it better than one given a page of adjectives describing it. Include the ones with edits visible if you have them. Seeing what you cut is more instructive than seeing the final version. ## 2. What good looks like The standard you judge work against, written down. This is the one people skip, and it is the one that makes a reviewer step possible: a workflow can only check a draft against a standard if the standard exists somewhere other than your head. It rarely takes long. What must a proposal always contain? What makes a summary useless to you? What would you send back? Half a page of that outperforms any amount of general instruction. ## 3. The decisions you have already made, and why The reasoning behind your positions - why you price the way you do, why you stopped offering that service, which kinds of client work out and which do not. Most AI advice is generic because the model has no access to the thinking that makes your situation yours. The "why" matters more than the "what". A decision without its reason cannot be applied to a new case; with it, the same judgement can be reused. ## 4. The recurring facts The things you would otherwise retype every week: who your people are and what they do, your products and what distinguishes them, your constraints, your calendar shape, the acronyms your industry uses that a general model will read wrongly. Dull to write, disproportionately useful. This is the layer that stops you spending the first third of every conversation explaining yourself. ## 5. The shape of the work you repeat For each task you do regularly: what triggers it, what inputs it needs, what the output looks like, and where it usually goes wrong. This is also the raw material for any workflow you build later - the two things compound, which is why they suit two days rather than one. ## What should not go in - **Credentials, keys and passwords.** Never, in any form. - **Personal data about other people**, unless your organisation has cleared the tool for it. Someone else's information is not yours to relocate. - **Anything you would not put in an email to an external supplier.** A useful test, and a fast one. - **Everything else you have ever written.** A second brain is a curated supply, not an archive. Volume without curation is how these systems get quietly abandoned - the value is in what you left out. ## The part that decides whether it survives A second brain decays. Standards change, people move, decisions get superseded, and a store that silently goes stale is worse than none because you stop being able to trust it. So the maintenance routine is not an afterthought to the build: what gets reviewed, how often, and what triggers an update. That is why building one is [a day's work rather than an afternoon's](/guides/how-long-to-build-a-second-brain/), and why the day ends with the routine rather than the contents. The [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) does exactly that: you build yours from your own material, organise it so it stays findable as it grows, and leave with the routine that keeps it worth having. --- ## Guide: What does ‘hands-on’ actually mean in AI training? URL: https://oxfordagentic.com/guides/what-hands-on-ai-training-actually-means/ Updated: 4 November 2026 “Hands-on” in AI training covers four different things: watching a live demo, following a guided exercise, building against a set brief, and building against your own real task. Only the last one reliably changes how you work afterwards. The test is simple - ask what you will personally have running at the end. Nearly every AI course now describes itself as hands-on, practical or applied. The word has stretched to cover four quite different experiences, and the difference between them is the difference between a good day out and a changed working week. ## The four things "hands-on" can mean **1. Watching a live demonstration.** Someone builds something in front of you. Genuinely useful for seeing what is possible - and it is still a talk. Your hands are not on anything. **2. Following a guided exercise.** You type along with an instructor towards a known destination, usually on prepared material. You finish with the instructor's example, built through your keyboard. Better; the friction is real but the problem is borrowed. **3. Building against a set brief.** You work on a case or a sample dataset with support in the room. Real building, real decisions - but when you get home, the transfer is still yours to do: your work does not look like the brief. **4. Building against your own task.** You bring a piece of your actual work - the report, the sift, the weekly assembly job - and build the workflow against that. When the day ends there is nothing to transfer, because the thing you built is already about your job. Each step up costs more facilitation and caps the room size, which is why most large-format training stops at levels one or two. That is an economic choice, not a dishonest one - but it is worth knowing which one you are buying. Delivery mode is tangled up in it as well, because the upper two levels are harder to run at a distance: when someone is stuck, the fix is a person walking over. That is [the case for doing this in a room](/guides/why-in-person-beats-online-for-learning-ai/). ## Three questions that settle it Course pages rarely say which level they mean, so ask directly: - **What will I personally have running at the end?** The honest answers range from "notes and ideas" to "a working workflow against your own task". Both are legitimate products; they are different products. - **Whose problem do I work on?** A demo, a script, a case study, or my own work. - **What happens when I get stuck?** A forum, office hours, or a person who walks over. The answer predicts whether you finish. ## The checklists you will find elsewhere are written for online courses Search for how to tell whether a course is genuinely hands-on and most of the advice you will find is built for self-paced video: look for executable notebooks, a GitHub repository, graded projects rather than quizzes, a capstone, a syllabus that lists deliverables, a recent content-update date. All sensible. None of it tells you anything about a day in a room, because a day in a room has no syllabus page to inspect and no repository to open. Two signals replace them, and both can be checked before you pay. **What you are asked to bring.** A day built around your own work has to ask you for something in advance - a task, a document, an account you can sign into. If nothing is required of you beforehand, your material will not be on the screen, whatever the page says. Level four cannot be arranged on the morning. **Whether anyone has mentioned your laptop.** Building means installing or signing into something, and corporate and government devices often block tool sign-ups. A provider who expects you to build has usually hit this before and will warn you; a provider who has not raised it is not planning on you installing anything. For the rest - how many people are in the room, the ratio of talking to building, what happens when you get stuck - there is a fuller list in [ten questions to ask before booking AI training](/guides/questions-to-ask-before-booking-ai-training/). ## Why this matters more for senior people The more senior you are, the scarcer the day and the higher the cost of a format that ends in intentions. An afternoon of demos can be worthwhile for orientation - once. After that, the only training worth a diary slot is the kind measured by what runs afterwards. For what it is worth, our own format is built deliberately at level four: the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) caps the room small, asks everyone to bring one real task, and treats "what will you use this for next week?" as the closing question of the day. No coding, and no stretch of anyone talking at you for more than twenty-five minutes. The [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) sits at the same level for the same reason. You are asked to bring your own documents and notes, and the day is spent building a second brain out of them rather than out of a supplied example, so what exists at the end is yours and already populated. Level four is not a property of the subject. It is a property of whose material is on the screen. --- ## Guide: What is an agentic AI bootcamp? URL: https://oxfordagentic.com/guides/what-is-an-agentic-bootcamp/ Updated: 12 November 2026 An agentic AI bootcamp is a short, hands-on course where you build and run a working agentic workflow rather than learn about one. You bring a real task from your own work, chain several narrow AI steps together to handle it, and leave with the workflow running. Most need no coding. The word "bootcamp" has been attached to almost every kind of AI training by now, so it is worth being precise about what makes one agentic. ## The difference is orchestration, not tooling Most AI training teaches you to use a model well: how to prompt it, where it fails, what it is for. That is genuinely useful, and it is a single-step skill. You ask, it answers, you judge the answer. Agentic work is what happens when you stop doing that one step at a time. Instead of one prompt doing everything, you break a job into narrow steps and hand them off: something gathers the source material, something drafts against it, something else reviews the draft before you see it. Each step is small enough to be reliable. The chain does the work that used to eat your morning. That shift - from prompting to orchestrating - is the thing an agentic bootcamp exists to get you through. It is not conceptually hard. It is just hard to do for the first time on your own, at your desk, between meetings. ## What actually happens on the day The format varies, but the ones worth attending share a shape: - **You bring a real task.** Not a case study, not a sample dataset - something from your own work that currently costs you real time. This matters more than any other detail, because it is what stops the day being theoretical. - **You build in cycles, not lectures.** A rough working version early, then something that reviews its own output, then a proper pass at the real thing. Each cycle ends with the room comparing notes. - **Someone walks the room.** The failure mode of hands-on AI training is people getting stuck on setup rather than on the idea. A facilitated room fixes that in a minute rather than an afternoon. - **You leave with it running.** The test of the day is whether the thing works on Monday, not whether the slides were good. What that adds up to by the end of the day is narrower than most training pages imply, and worth knowing before you book: [what you can realistically build in one day](/guides/what-you-can-build-in-one-day-with-ai/). ## Who they are for Overwhelmingly, operators rather than engineers: founders, directors, consultants, senior public-sector and university staff. People who already use AI daily, have formed real views about it, and have still never chained three steps together on something they own. Technical depth is rarely the constraint. Most agentic bootcamps require no coding at all - the tools have moved far enough that the interesting problem is decomposition, not syntax. Knowing which part of your job to hand to a machine, and which judgement to keep, is the actual skill. If you are earlier than that, and still working out where any of this fits your own work, [there is a slower starting point](/guides/agentic-ai-for-business-leaders-where-to-start/). ## How to tell a good one from a bad one Three questions worth asking before you book: 1. **Will I build something, or watch something being built?** A live demo is a talk. Ask what you will personally have running by the end. 2. **How many people are in the room?** Facilitation stops working somewhere around twenty-five. Above that, you are at a lecture with a laptop open. 3. **Whose task do I work on?** If the answer is a prepared exercise rather than your own work, the transfer to your actual job is your problem to solve later. ## What does the day look like, hour by hour? Many providers describe the day without breaking it into sessions, so a full timetable is worth asking for before you book, whoever you book with. The Oxford Agentic Bootcamp publishes its own in full on [each cohort's event page](/events/). The shape is this: registration and coffee from 09:00, then twenty minutes on why prompting and orchestrating are different skills, then forty minutes taking an agent pipeline apart live in front of the room. The first build on your own task runs from 10:15 to 11:30, facilitated, and everyone gets to a rough working pass however untidy. A guest session follows, then a second build that pushes the pipeline past single-shot by adding a step that checks its own work, then lunch at 13:00. The afternoon opens with a second guest session, gives the real build an hour and a quarter from 14:15, and closes with a showcase at 15:30 where everyone runs what they made and says what they will use it for, a last stretch on making it stick, then wrap and networking to a 17:00 finish. ## How much of the day is building, and how much is listening? That is the first question above, and a published timetable is what lets you answer it rather than take a provider's word for it. On this one the three facilitated build blocks come to three hours and a quarter - an hour and a quarter on the first rough pass, forty-five minutes on adding the review step, an hour and a quarter on the task you actually came with - with a further half hour at the end where the room runs what it built. Set against that: about an hour of framing and live demonstration at the start, and an hour and a quarter of guest sessions across the day. Any timetable you cannot break down that way is worth a question before you book. The point of a day like this is not to leave better informed. It is to leave having done it once, so that the second time is something you can do alone. --- ## Guide: Three different things get sold as a ‘second brain’ URL: https://oxfordagentic.com/guides/what-second-brain-actually-means/ Updated: 18 November 2026 The phrase now covers at least three different products. One is a task and attention system. One is a store of context that AI tools read. One is a gut-health talk, which is a genuine and separate field. They are sold with the same two words, so the useful question before booking anything is which problem it solves. Search for a second brain workshop and you will get results that have almost nothing to do with each other. Some are about managing your workload. Some are about AI. At least one is about your gut, which is a real and unrelated field of medicine. There are also consultancies and software firms trading under the name. This is not a small labelling problem. These are all things you can book, on the same two words, and the pages do not always make the difference obvious until you are some way in. ## The three you are most likely to meet **A task and attention system.** The oldest business meaning, and the one behind most full-day productivity workshops. The problem it solves is that you are carrying too much in your head: commitments, follow-ups, half-finished threads. You build a trusted place to put all of it, plus the habits of capturing and reviewing, so you stop using your memory as a to-do list. The reader of this system is you. **A store of context that AI tools read.** The newer meaning, and the one this site uses. The problem it solves is different: you use AI every day and every conversation starts from nothing, so you re-explain your business, correct the same wrong assumptions, and rewrite output that sounds like nobody in particular. What you build is a supply of your own context, voice and standards, structured so a model can draw on it. The reader is the model, which [changes almost everything about how you build it](/guides/second-brain-for-ai-vs-note-taking/). **The gut.** The enteric nervous system is nicknamed the second brain, and workshops on gut health use the phrase in that sense. Entirely legitimate, entirely unrelated, and easy to land on by accident. ## The question that separates them Before booking anything with this name, ask what the finished thing is for. If the answer is *so I stop forgetting commitments*, it is a task system. If the answer is *so my AI tools stop giving generic answers*, it is a context store. Those are different problems, and neither one is a lesser version of the other. The confusing cases are the ones that mention AI without changing the underlying job. A task system with an AI assistant bolted on is still a task system, and it is a good thing to own. It will not fix generic AI output, because generic output is not caused by a disorganised to-do list. ## Which one you actually need Work backwards from the sentence you would use to complain about your own week. "I am drowning and things are falling through the cracks" is a task-system problem. Go and solve it, and do not let anyone sell you an AI day instead. "I keep explaining the same things to a machine that should already know them" is a supply problem. More discipline with folders will not touch it, because [the thing that failed was never the filing](/guides/why-your-second-brain-stopped-getting-used/). "Both" is common, and they are compatible. They can even share material. They are just not the same day, and buying one expecting the other is the mistake this page exists to prevent. ## Where this one sits The [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) is squarely the second kind. One day in Oxford, built from your own material, and you leave with a second brain populated and connected to the tools you already use. No coding, and no attempt to be a task system as well. If, having read the above, your problem is actually the first kind, this is not your day, and it is better that you know that now than at eleven o'clock on the morning of it. --- ## Guide: What can you actually build in one day with AI? URL: https://oxfordagentic.com/guides/what-you-can-build-in-one-day-with-ai/ Updated: 21 November 2026 In one day, with no coding, you can get a multi-step AI workflow running on a real task you already repeat - something that gathers the material, produces a first version, and checks it against a standard you wrote down. What you will not have is a polished product, an approved integration, or anything you can leave running unwatched. The honest answer to this question is narrower than the marketing and wider than the scepticism. A day is not long enough to build a product. It is more than long enough to get something real running, provided you pick the right thing and accept what "running" means at that stage. ## Three passes, not one build A day that produces something working is almost never one long build. It is three passes at the same task, each one shorter than you expect. The first pass gets to a rough working version, however untidy - usually before lunch, and usually uglier than the person who built it would like. The second adds a step that checks the output against a standard, which is what moves the thing past a single clever prompt. The third applies it properly to the task you actually came with, now that you know where it breaks. That is the shape [the Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) runs to, and the reason it works is that the first pass is deliberately allowed to be bad. What this means in practice: budget your day in thirds, not in one heroic sitting. If you have not got something rough running by midday, the problem is the task, not the time. ## What fits: work shaped like gather, produce, check The jobs that finish inside a day all have the same silhouette. Something collects the raw material, something writes a first version from it, and something checks that version against a stated standard. Three narrow steps, each testable on its own - [the scout, drafter, reviewer pattern](/guides/scout-drafter-reviewer-pattern/) in full. In [Cohort 1’s room](/events/oxford-agentic-bootcamp-cohort-1/), the tasks that surfaced were lead generation, internal reporting and marketing content. Different industries, same shape: gather, produce, check. If the thing you want to build has that shape and runs on words and documents rather than on a system you cannot reach from a laptop, a day is enough. If it does not, no amount of time in a room will help, and [choosing the task](/guides/how-to-choose-your-first-agentic-task/) matters more than choosing the tool. ## Four things that will not fit - **Anything that needs access you do not have on the day.** If the workflow depends on a system whose credentials sit with someone who is not in the room, you will spend the day building around a hole. - **Anything that needs sign-off before it can touch real data.** That is a separate conversation with a separate timeline, and worth [starting before the build day](/guides/getting-an-ai-workflow-past-it-and-legal/) rather than after it. - **Anything that has to be right without a human reading it.** A review step raises the floor; it does not remove the need for someone to look. Workflows that must be unattended are a longer project. - **Software other people are meant to use.** Working code on your laptop and something live for colleagues are separated by six unglamorous jobs - [which is why the thing an agent wrote is not live](/guides/why-the-code-an-agent-wrote-is-not-live/). ## The "no coding" part is mostly true Most of what is described above is built by chaining tools that already exist, which is why most people who do this never write a line of code. The skill being learned is orchestration - deciding what each step is for, what it hands to the next one, and how you would know it had gone wrong. Code starts to matter when the thing has to run on a schedule, live somewhere other than your own machine, or talk to a system that offers an API and no interface. That is a genuinely different day, and the reason [the Oxford Agentic Coding Bootcamp](/bootcamps/agentic-coding-bootcamp/) exists as its own format rather than as an advanced afternoon. ## Is one day enough? Four versions of this question arrive, and they want different answers. ### Would a longer course not give me more? It would give you more time, and time is not usually what is missing. A day is enough to get one workflow running on a task you already repeat. It is not enough to make you fluent at building them, and it would be a poor claim to make. The people in our rooms tend to arrive saying some version of "I understand it, I have just never built anything" - they are not short of explanation, material or ability, and a few more weeks of any of those would not move the thing that is stuck. What they are short of is one finished example with their own work inside it. If what you want is the frame, the governance view or the breadth, [a programme or an online course is the better buy](/guides/ai-bootcamp-vs-course-vs-exec-education/), and the formats compound rather than compete. ### Will I build something, or watch someone else build it? You build. You bring one real task from your own work and a laptop, and the day is spent building rather than watching. On the Oxford Agentic Bootcamp that is structural: the day runs on build cycles rather than talks, with no block longer than twenty-five minutes of anyone talking at you, and someone walking the room so the first stuck moment costs you a few minutes rather than an evening. The day is not recorded, which is the other half of the same decision - there is nothing to take home except the thing you built and the recipe for rebuilding it. The one condition is the task. Arrive without a real one and you get a demonstration instead of an outcome, which is the one thing a day in a room does no better than a video. ### Is a day enough if nothing happens afterwards? No, and that is worth saying plainly rather than waiting to be asked. What you leave with is a working draft, and a draft nobody opens again stops being worth the day inside a fortnight. It is why the recipe matters more than the workflow it produced: the recipe is the part that lets you rebuild and extend the thing at your own desk, on an ordinary Tuesday, with the task in front of you and nobody walking the room. If other people are meant to use what you built, [planning for week two](/guides/how-to-run-an-ai-pilot-that-survives/) is the same discipline applied to a team. And before concluding that it worked, [measure the whole job rather than the step that impressed you](/guides/is-your-ai-workflow-saving-time-or-just-effort/). ### What if the task turns out to be too big for the day? You find that out before lunch, which is what the three passes are for. The first pass is allowed to be rough precisely so the signal arrives early: if nothing is running by midday, the task is the problem and not the time. The fix from there is a smaller task rather than a longer day - the same gather, produce, check shape, cut back until each step is testable on its own. If what you brought is one of the four things above that will not fit, more hours in the room would not have finished it either. What you leave with is the recipe as much as the workflow, so a narrower build still leaves the part you rebuild from at your own desk. ## What you should expect to leave with Three things, if the day has gone well: a workflow running against a real task from your own work, a review step inside it that checks its own output, and a written recipe so you can rebuild and extend it on Monday. The recipe is the part people underrate. A workflow you cannot rebuild is a demo you happened to attend. That is a smaller claim than "transform your business in a day", and a much more useful one. Cohort 4 runs on [3 December 2026 at Worcester College](/events/oxford-agentic-bootcamp-cohort-4/), and the task you bring decides most of what you get. --- ## Guide: When do AI executive cohorts run, and when should you go? URL: https://oxfordagentic.com/guides/when-ai-executive-cohorts-run/ Updated: 20 November 2026 Executive AI programmes used to follow the academic calendar. Most now run online several times a year, so the timing question has changed: it is no longer when a cohort is available, but where you are in your own progression. The useful sequence is understand, then build, then bring it back. People ask what time of year executive AI cohorts begin, and the honest answer is that the question has aged out of usefulness. It made sense when executive education followed the university calendar and you waited for an autumn intake. Most AI programmes no longer work that way. ## What actually varies now Online programmes repeat. Oxford's own generative and agentic AI programme runs online across four weeks and listed two more start dates for 2026, on 16 September and 11 November, rather than a single autumn intake (checked on its own page, 31 August 2026). That is how most online executive AI education behaves now, because the subject moves too quickly to teach annually - a curriculum written for last September is visibly dated by spring. In-person days are the exception, and they are scheduled one date at a time rather than in a season. Ours were set that way, individually, which is why the 2026 sequence is as uneven as it is. The Oxford Agentic Bootcamp ran on [16 September](/events/oxford-agentic-bootcamp-cohort-2/) and again on [19 November](/events/oxford-agentic-bootcamp-cohort-3/), and the Second Brain Bootcamp ran on [21 October](/events/oxford-second-brain-bootcamp-cohort-1/) at the Pitt Rivers Museum. The days still on sale are the first Coding Bootcamp, which runs on [26 November](/events/oxford-agentic-coding-bootcamp-cohort-1/), a fourth agentic day on [3 December](/events/oxford-agentic-bootcamp-cohort-4/), and the first of the 2027 dates, which is on [28 January](/events/oxford-agentic-bootcamp-cohort-5/) at Jesus College rather than Worcester, with the room booking there still being finalised. Worcester College holds all of the others, and [the events page](/events/) is the list that stays current. Our September day and that online programme happened to begin on the same date, which made the choice a clean one: four weeks online, or one day in a room with your own work in front of you. If the agentic day and the coding day both look plausible, they are different questions rather than different levels. The agentic day builds a workflow around a task you already repeat. The coding day is for people who understand the concepts and have not yet shipped something of their own, and the agents do the typing. So the practical advice is dull but true: check the provider's page rather than waiting for a season that no longer exists. If you are looking at an online programme and the next cohort is two months out, there is very likely another one after it. ## The timing question worth asking instead Since availability has stopped being the constraint, the useful question is where you are rather than what month it is. Most people arrive at this having done the reading. They can hold their own on agentic AI in a board discussion and could not build one if you asked them to tomorrow. That gap is not fixed by finding a better course to understand it with. It is fixed by building something. The sequence that works is unglamorous: - **Understand the frame first, if you do not have it.** The economics, the governance questions, where this fits your organisation. A [programme or a structured course](/guides/ai-bootcamp-vs-course-vs-exec-education/) is a reasonable way to get it, and it ages well. - **Then build one thing.** Not a second course. One workflow, against a task from your own week, that still runs on Monday. - **Then bring it back.** The step almost everyone skips, and the one that decides whether any of it survives contact with the organisation. ## The mistake in waiting There is a version of this where you wait for the right cohort, the quieter quarter, the moment the diary clears. It rarely arrives, and the cost of waiting is not neutral. The tools improve every month. The ability to build with them does not improve on its own, and the distance between the two is what widens. Someone who understood agentic AI well in March and has still built nothing understands it slightly less well now, because the thing they understood has moved. If you have already finished a programme, the next step is not another one - [that gap has its own page](/guides/after-the-ai-executive-programme/). If you have not, you can honestly do either first. What does not work is doing the understanding twice. ## What to do between cohorts If the date you want is weeks away, the gap is not dead time. Pick the task you would bring. It should be something you already do by hand most weeks, boring enough that nobody would miss it, with a recognisable before and after - [choosing it well](/guides/how-to-choose-your-first-agentic-task/) does more for the day than any amount of preparatory reading. Arriving with that decided is the single biggest difference between people who leave with something running and people who leave with notes. --- ## Guide: Where the human actually goes in an agentic workflow URL: https://oxfordagentic.com/guides/where-the-human-goes-in-an-agentic-workflow/ Updated: 18 November 2026 A human belongs in an agentic workflow at three points: before anything irreversible or external happens, wherever the output is a judgement about a person, and at the standard the work is checked against. Anywhere else, a human in the loop is usually labour rather than oversight, and it costs the time the workflow was meant to save. "Keep a human in the loop" is the most repeated advice in agentic AI, and it rarely says where. So people put themselves in the obvious place: the middle. The first tool produces something, they read it, tidy it, paste it into the next tool, and carry on. That feels like oversight. Mostly it is labour. You are acting as the cable between two steps, retyping context a workflow could have passed along itself, and the time you meant to save goes on that. ## Three places a human genuinely belongs **Before anything irreversible or external.** Sending, publishing, paying, deleting, or anything a client will see. These are the actions where a mistake cannot be quietly corrected afterwards, so the workflow should stop and ask. Everything upstream of them - gathering, drafting, checking - can usually run on its own. **Wherever the output is a judgement about a person.** Who to hire, who to chase, whose account to flag, what to say to someone having a bad month. A workflow can gather the evidence and draft the options. The decision stays with someone who can be accountable for it. **At the standard.** This is the one people miss. The most useful human contribution to a workflow is often not approving its output but writing down what good looks like, so a review step can check against it. In the [scout, drafter, reviewer pattern](/guides/scout-drafter-reviewer-pattern/) that written standard is what the reviewer reads, and it is yours to write and to keep current. ## The test for any step Ask one question of each step: if this were wrong, who would find out, and how late? If the answer is "a client, next week", a human belongs before that step. If the answer is "the next step would catch it within seconds", the human is probably redundant there. If the answer is "nobody, ever", that is the dangerous case, and it is a monitoring problem rather than an approval one: [a workflow that fails quietly](/guides/how-to-notice-when-an-ai-workflow-is-wrong/) needs a check you run on purpose, not a person skimming everything. ## Approving everything is not the safe option It feels cautious to review every output. In practice it decays fast. By the fortieth draft that was fine, you are clicking approve without reading, which gives you the look of oversight with none of the substance, and a false sense of safety on the day it matters. Fewer checkpoints, placed where the risk actually is, get read properly. That is worth more than many checkpoints nobody reads. ## Start tight, then move the human On day one, it is reasonable to check more than you will eventually need to. You have no evidence yet about where the workflow goes wrong. As the evidence comes in - the steps that never fail, the ones that fail in the same way every time - move the human out of the first kind and towards the second. The order matters. Gathering usually earns trust first, drafting later, and judgement about people arguably never. A workflow you trust in stages is one you can hand over in stages, which is far easier than handing over all of it at once. ## Where this gets practical Deciding where the checkpoints go is much easier with a real task in front of you than in the abstract, which is why where to keep a human in the loop is part of the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) curriculum. You bring one real task, build the workflow against it, and decide in the room where a human needs to stay. Cohort 4 runs on [3 December 2026 at Worcester College](/events/oxford-agentic-bootcamp-cohort-4/). No coding required, and nothing hand-waved. --- ## Guide: Who should take the seat - you, or someone in your team? URL: https://oxfordagentic.com/guides/who-should-take-the-seat/ Updated: 16 November 2026 Send whoever owns the task the day will be spent on. The whole day runs against one real task the attendee brings, so seniority is the wrong axis: the right person is the one whose morning the task eats, who can change how it is done on Monday, and who will build the next one. If all three point at you, go yourself. You have one seat and two candidates: you, or the person who actually does the work. Most of the material written about choosing AI training answers a different question - whether to bring a day in-house, or which format to pick - and then describes who the day is *for* in job titles. That does not settle it, because the job titles usually fit both of you. The question is normally asked as a seniority question. Seniority is the wrong axis. ## A build day sorts by task ownership, not rank On a day where you build something, the thing being built has to belong to someone. [The Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) runs against one real task that the attendee brings from their own work, and its own "Who is it not for?" answer is blunt about the consequence: anyone who cannot bring a real task should not come, because without one you get a demonstration instead of an outcome. That single constraint does most of the deciding for you. The seat should go to whoever arrives with a genuine task, not to whoever would normally be sent to Oxford. ## Three questions that settle most cases ### Whose morning does the task eat? Not whose budget, whose morning. The person who personally loses the time pays a different kind of attention to getting it back, and they already know where the task is awkward - the exceptions, the bit that always needs checking, the file that is never where it should be. Someone describing a colleague's workflow from memory finds that out at 11am, in the room, with the day running. ### Who can change how it is done on Monday? If the person who goes has to ask permission to alter the process afterwards, the day produces a recommendation rather than a change. Recommendations go in folders. This is the question that most often sends the seat upwards, and sometimes it genuinely should. ### Who will build the next one? What transfers from the day is not the workflow, it is the pattern behind it - [a step that gathers, a step that produces, a step that checks](/guides/scout-drafter-reviewer-pattern/). The first workflow only ever covers the task you brought. The pattern is what makes the second and third ones happen without help, so the seat is worth more to someone who will keep using it than to someone who wanted to see what this is about. When all three point at the same person, that is who goes, whatever the org chart says. ## When it should be you Go yourself if the decision in front of you is not "can this task be faster" but "is this how we are going to work". Those are different decisions, and the second one is hard to make from a write-up. In [Cohort 1's room](/events/oxford-agentic-bootcamp-cohort-1/) the afternoon was adoption and governance: the five-level journey from a personal assistant up to autonomous systems, worked through on compliance review and marketing analytics, and where data integrity and risk sit. That material is aimed at whoever will have to decide how far the organisation goes, and how it will know when a workflow has gone wrong. If that is you, send yourself. The other honest reason to go yourself: if you have approved AI spending without ever having built anything, a day spent building recalibrates what you ask for. It is much harder to be sold something afterwards. ## When it should be them Go the other way if the answer to "whose morning" is plainly someone else's. The tasks that surfaced in Cohort 1 were lead generation, internal reporting and marketing content - in most organisations, none of those three sit with the chief executive. If you do delegate the seat, delegate the decision with it. Agree before they go what they are allowed to change when they come back, and who they need to tell rather than ask. A workflow built by someone who then has to campaign for permission to use it is a day half spent. ## Two of you, and what to do when there is only one seat If you can send two people, the programme page's own advice is to bring the same task, so the same workflow goes back to work with both of you - and teams do get more out of the day than individuals. Email hello@oxfordagentic.com and that can be arranged. It is also a better answer than sending one person this year and another next: two people who have built the same thing can keep it working between them. With one seat, the owner of the task goes, and you book an hour in the same week to be *shown the thing running*. Not a summary, not slides. If it cannot be demonstrated, it was not finished, and you want to know that in week one rather than in the quarter when you assumed it had been adopted. ## The two ways this goes wrong **Sending someone with no task of their own.** Usually well meant - the person with the most available diary, or whoever has capacity that week. They spend the day on a hypothetical, and nothing comes home. The day is not a briefing, and the published "not for" list says so. **Not checking the laptop.** A device that blocks new tool sign-ups is a real problem on a build day, and corporate and government machines often do. This is worth checking about your candidate before you decide who goes, not at 09:30 on the day - and if theirs is locked down, that is worth raising when you book rather than discovering in the room. ## If the honest answer is "both, for different reasons" That usually means there are two tasks and one ambition, and it is worth saying out loud which one the day is being bought for. Deciding whether to adopt something and getting a specific morning back are both good reasons to spend a day. They are just not the same purchase, and a seat bought for the first while hoping for the second tends to deliver neither. Cohort 4 runs on [3 December 2026 at Worcester College](/events/oxford-agentic-bootcamp-cohort-4/). If you are still deciding between sending a few people and bringing a day in-house, [that comparison is a separate one](/guides/in-house-ai-day-or-send-people-to-a-cohort/) and worth settling first. --- ## Guide: Why in-person beats online for learning AI URL: https://oxfordagentic.com/guides/why-in-person-beats-online-for-learning-ai/ Updated: 27 October 2026 For senior leaders who want to build with AI rather than learn about it, a day in the room usually beats an online course. Online is better at explaining concepts and reaching a spread-out team. What stalls people is rarely the concept; it is being stuck on setup with no one to unstick them, and a facilitated small room removes that. There is no shortage of ways to learn about AI online, and many of them are excellent. But most people who have taken them will admit the same thing: they understand more than before, and they have still not built anything. That gap is worth taking seriously, because it is not about the content. ## The thing that stalls people is friction, not concept Watching someone build a workflow is not the same as building one. The moment you try it yourself, you hit the small obstacles no video covers: the tool behaves oddly, a step does not connect, you are not sure whether the output is good enough. On your own, at your desk, between meetings, that friction is usually where it ends. The tab gets closed and the intention drifts. A course cannot walk over and unstick you. A person can, in about a minute. ## A room does three things a screen cannot - **It removes the setup tax.** Someone is there to get you past the tooling so you spend the day on the idea, not the plumbing. - **It creates a deadline.** A booked day with other people in the room is a commitment you keep. "I'll get to it" quietly loses to everything else; a date in the diary does not. - **It gives you peers.** Building alongside other senior people working on their own real tasks is its own teacher. You see five other ways to approach a problem, and you leave with their contact details. ## When online is the better choice In-person is not always the right answer, and it is worth being plain about when it is not. - **You need the concepts, not a build.** If the aim is to understand agentic AI well enough to judge proposals and question vendors, a good online course does that well, costs less and fits round your diary. - **Your team is spread out.** A day in one room means travel for everyone who is not nearby. For a distributed team, live online sessions built around real tasks will reach more people. - **You are already building.** Once your first workflow is running, going deeper on one tool is often quicker with targeted online material than with another day away. The case for a room is narrower and stronger than "in-person is better": it is the right choice when you want to leave with something built against your own work, and the likeliest thing to stop you is getting stuck alone at your desk. If you are weighing executive education and bootcamps as well, [which AI training format leaders need](/guides/ai-bootcamp-vs-course-vs-exec-education/) compares all three. ## The measure is what runs on Monday The honest test of any AI training is not whether the slides were good. It is whether you have something working afterwards. Online learning is measured in modules completed. A day in the room is measured in a workflow that runs against your real task after you have gone home. That is the whole design of the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/): a small cohort, a facilitated day in Oxford, and one working agentic workflow built against a task you actually brought. No coding, and nothing to watch later, because you built it in the room. The [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) is built on the same principle and points it at a different problem. You spend the day populating a second brain with your own documents and notes rather than a worked example, and leave with it organised and running. The [Oxford Agentic Coding Bootcamp](/bootcamps/agentic-coding-bootcamp/) is the sharpest case of the setup tax this page opened with. What it asks you to bring is a laptop you can install software on, because the day involves setting up a code editor, a terminal and a GitHub account on your own machine, and corporate or government devices often block that. Most of what exists for agentic coding is online - recorded courses, platform listings, live sessions over video - and Claude Code and Codex do the typing in any of them. What a room adds is a senior engineer sitting next to you at the point where the machine says no. None of the argument above is really about agentic workflows. It is about the difference between being told how something works and having it working by the time you go home. --- ## Guide: The agent wrote the code. Why is nothing live? URL: https://oxfordagentic.com/guides/why-the-code-an-agent-wrote-is-not-live/ Updated: 11 November 2026 A coding agent gives you working code on your own machine. What it does not give you is a repository, a way to undo, somewhere the thing runs, a memory, a place to keep secrets, and a check that it does what you think. None of those six are coding, all six are learnable in a day, and they are the whole distance between a demo and something live. Something odd happens the first time you build with a coding agent. It works. You describe what you want, the agent writes it, the thing runs on your screen, and you feel about four steps from launching. Then you try to show it to someone and discover the four steps were the easy part. This is the most common place people stop, and it is worth being precise about why, because the gap is not what it looks like. It is not that the code is bad or that you need to learn to program. It is that working code on your laptop and software other people can use are different objects, and nobody told you what sits between them. ## The six things between your laptop and the internet Every one of these is a thing agents assume you already have. None of them requires you to write code. - **Somewhere for the code to live.** A repository, usually on GitHub. Not a folder on your desktop, because a folder cannot be shared, restored, or deployed from. - **A way to undo.** Version control is the reason you can let an agent change fifty files at once and still sleep. Without it, one confident wrong edit is unrecoverable, and you will start refusing the changes that make the tool worth using. - **Somewhere it actually runs.** Your machine is not the internet. Deploying means putting the thing on a host that serves it to other people, with a URL that is not localhost. - **A memory.** The moment your thing needs to remember anything between visits, you need a real database rather than a variable that resets. - **Somewhere to keep secrets.** API keys, passwords, tokens. These must not sit in the code, and an agent will cheerfully put them there if nobody has said otherwise. - **A way to know it works.** Not a feeling. A test that fails loudly, and a habit of reading what changed before it goes out. Read that list again and notice what is missing: writing code. That is the part the agent already does. ## Why every comparison tells you to use a builder instead Search for whether a non-developer can ship with a coding agent and you will mostly find pages selling app builders. They are not wrong about the gap. They are arguing that you should not have to cross it, because their platform owns the repository, the hosting, the database and the deploy on your behalf. That is a real trade, and worth taking seriously. A builder gets you live faster. What you give up is ownership: the thing exists inside their account, on their terms, priced how they choose, and moving it later means rebuilding. The six items above are the price of owning what you made. Paid once, they apply to everything you build afterwards. Which side of that trade suits you depends on what you are making. A one-off internal tool nobody else depends on is a fine thing to leave in a builder. Anything you intend to keep, sell, or hand to someone else is worth owning outright. ## The order that works If you do decide to own it, the order matters more than the tools: 1. **Write the brief before the code.** The agent reads it first and builds against it. A vague brief produces a confident wrong answer, and that is your time, not its. 2. **Get the repository and the undo button working before anything else.** You want version control before you need it, not after. 3. **Deploy something almost empty, early.** A nearly blank page live on a real URL is worth more than a finished app on your laptop, because the hard part is the path, not the page. After that, every change is a small repeat of a path you have already walked. 4. **Add the memory and the secrets handling only when the thing needs them.** 5. **Put the check in before you invite anyone.** Most people do this in exactly the opposite order, build something elaborate locally, and then meet all six problems at once on the day they wanted to launch. ## How do I check what the agent wrote if I cannot read code? You check the outcome rather than the code, and three habits cover most of it. Use the thing the way a real person would, including the inputs you expect to go wrong, because a screen that looks right proves only that something rendered. Look at what changed and how much: if you asked for one thing and fifty files moved, that is worth a question before it goes out, and you do not need to read a file to count them. Then put a test in, so the check runs itself next time rather than depending on you remembering. None of those three require reading a line of code. All three are habits rather than skills, which is why the sixth item on the list above is learnable in a day, and why it is still the one people skip. ## Does a passing test mean it works? It means the test passed, which is a smaller claim. A test only checks what somebody told it to check, so a suite can be entirely green while the thing you actually care about never happens, and an agent asked to make the tests pass can write tests that pass. It is worth knowing what yours exercises before you trust the colour of it. The practical version: keep at least one check that goes through the whole path the way a person would, end to end, and treat the rest as a safety net for the parts you have stopped watching. ## Can I ask the agent to check its own work? Yes, and it is worth doing. An agent that can run what it built will catch mechanical faults faster than you will, and asking it to explain what it changed is a reasonable way to find out what it assumed. What it cannot do is know what you meant. That is the one check that does not hand over: whether this does the thing you actually wanted, for the person who is going to use it. Keep that part and delegate the rest. ## Where this gets practical Reading the list is not the same as having done it, and the first time through, every step has a small trap in it that costs an evening. That day is what the [Oxford Agentic Coding Bootcamp](/bootcamps/agentic-coding-bootcamp/) is for: Claude Code and Codex do the typing, and the day is spent on everything around them, from the brief to the repository, the deploy, the database and the tests. You leave with a personal site live on the internet and a working prototype, both in your own repository. Cohort 1 runs on [26 November 2026 at Worcester College](/events/oxford-agentic-coding-bootcamp-cohort-1/). If you are still deciding which agent to use, [the plain comparison of Claude, Cowork, Claude Code and Codex](/guides/claude-cowork-claude-code-or-codex/) is the place to start. If your problem is a recurring task rather than an idea you want running, the [Oxford Agentic Bootcamp](/bootcamps/oxford-agentic-bootcamp/) is the better day. --- ## Guide: Why your second brain stopped getting used URL: https://oxfordagentic.com/guides/why-your-second-brain-stopped-getting-used/ Updated: 18 November 2026 Most second brains fail the same way: capture is easy, and the reading back never happens. The usual fix is more discipline or better folders, and it does not work. What changes the outcome is swapping the reader - building the thing for a model rather than for a future version of yourself who was never going to browse it. Search for why second brains fail and the answers converge fast: too much capture, not enough review, folders that decay into a junk drawer. All of that is true, and all of it describes a symptom rather than a cause. The cause is duller. You built a system whose only reader was you, and then you did not read it. ## The reading was always the half that would not happen Capture is pleasant. Every save feels like progress, costs a few seconds, and leaves something behind you can point at. Reading back is none of those things. Nothing prompts it, it competes with actual work, and it pays off only when you happen to need the thing you happen to have kept. So the two halves were never equally likely to survive. One was a habit with an immediate reward. The other was homework. That is why "be more disciplined about your weekly review" does not fix it. It is advice to keep doing the half that was already losing. ## What changes when the reader is not you A second brain your AI reads is a different object, and the difference is not the tooling. [It is who it is written for](/guides/second-brain-for-ai-vs-note-taking/). Three things move. **Retrieval stops being your job.** You no longer have to remember a note exists in order to get the benefit of it. That single change removes the failure everybody describes, because the failure was always "I forgot I had that". **You keep less, not more.** A model does not need your reading list. It needs the handful of things that make its answers yours rather than generic: how you write, what you have already decided and why, [the recurring facts of the work](/guides/what-goes-in-a-second-brain/). Most of what sank the old system does not belong in the new one. **The writing changes.** Notes to yourself lean on context you are carrying in your head. A model has none of it, so the note has to say the quiet part out loud. ## What this does not fix It does not make the thing self-maintaining. A second brain a model reads fails in a new way instead: [it goes stale and then answers confidently from the stale version](/guides/keeping-a-second-brain-from-going-stale/), which is worse than a folder you scroll past, because nothing looks wrong. It also does not conjure material. If what you built holds nothing about your actual work, changing the reader changes nothing. ## The honest test Open whatever you built. Find the last time you went to it to answer a real question, rather than to add something to it. If that date is further back than you expected, the system is not underused. It is finished, and it is telling you it was built for a reader who was never going to turn up. That is the premise the [Oxford Second Brain Bootcamp](/bootcamps/second-brain-bootcamp/) starts from. You spend the day building one from your own material, for a reader that will actually do the reading, and you leave with it working rather than with a template to fill in later. No coding required, and less to keep than the one that stopped getting used. ---