ALT Annual General Meeting · · Online talk and questions
From tool maker to tool user
How AI in 2026 reshapes the space of possibilities for learning technology
An illustrated summary of the talk and the questions that followed, for anyone who missed it or wants to go back to an example.
The talk in five parts
Select a part to play it from there.
New AI models arrive every six to eight weeks, and they now drive agents, which do the work instead of answering questions about it. Most people still think of AI as a chatbot; Dominik more or less stopped using ChatGPT by May.
AI went from oracle, to tool maker (vibe coding, 2025), to tool user (2026): a model using tools in a loop on your computer. An agent filled in his York expenses claim, and an always-on assistant read his slides and spotted a typo. Full agent use costs $100 to $500 a month.
An agent built and published a Canvas course through the Canvas API: finally, the death of the VLE as a place where you upload and dump files. A Word textbook became a reading-friendly web book, a Nature primer became an interactive walk-through with a narrated video, and the AGM papers became one web page.
Watch someone who uses agents for work like yours, then install a desktop agent such as Codex or Claude Code on the $20 plan and try things step by step. Learn how your computer works: files, sync, hosting, databases. Few lessons from the chatbot era still apply, so you have missed almost nothing.
Prices are rising because there is not enough computing power, not because the technology fails, and what you build is ordinary software that stays on the web. On energy, ask whether a use is worthwhile. For learning technologists, agents can now do what learning technology was always meant to do.
Slides and resources
Comment or ask a question
Your message goes to Dominik by email. Add your email address if you would like a reply; it is used only for that.
Transcript of the video
Out of date
Any assumption about what AI can do is likely to be out of date soon. Dominik Lukeš, of the AI Competency Centre at the University of Oxford, made this his starting point at the ALT Annual General Meeting on 11 November 2026. New models now arrive every six to eight weeks, and each release scores higher than the last. They no longer only drive chatbots. They drive agents, which do the work instead of answering questions about it.
Speed and ambition
Agents are also speeding up the software around us. Firefox now fixes hundreds of security bugs a month, where it used to fix tens. And the same capability raises ambition: you can now take on bigger and more complex projects than before.
Still chatbot
Most people still think of AI as a chatbot. In January, Dean Ball predicted that by the end of 2026 the least important thing you could do with frontier AI would be getting chatbots to answer questions. For Dominik that was true by May, when he more or less stopped using ChatGPT.
Three stages
He described three stages. When ChatGPT arrived, people treated AI as an oracle: you asked a question and got an answer. Then came vibe coding, in 2025, when AI became a tool maker that could build a whole app, though you still needed to know what to do with it. In 2026 AI is a tool user. It works with the files and software on your computer, and can do almost anything you can do there.
The loop
An agent is a model using tools in a loop. You set a task; the model makes a plan and sends a command to the harness, the software around it. The harness runs the command on the computer and returns the result, and the model revises its plan and carries on, now for hours or even days. Any tool you can use on your computer, the agent can use for you: your files, the web, the command line, other models, the browser and now other applications.
Expenses
A chatbot can tell you how to fill in an expenses form. An agent fills it in. After a workshop at the University of York, Dominik asked his agent to find York's email, collect the forms and his receipts, and complete the claim. It filled in the claim and the bank form, organised the receipts as attachments, and he was paid.
Always on
Agents are also getting computers of their own. In September, OpenAI launched dots, assistants that run on their own cloud computer and work towards your goals around the clock. The day after the launch, Dominik's dot saw on his calendar that he was giving a talk, read his slides and pointed out a typo.
Price
This costs more. The twenty dollars a month that bought everything in 2023 has become a hundred to five hundred dollars for full agent use, and even the top tiers have limits. People pay because they find the value. One photographer who runs a small business said he spends about a thousand dollars a month and finds it worth it.
Canvas
So what does this make possible? Dominik showed a public Canvas course whose pages were built entirely by an agent on his desktop, to his direction, and published through the Canvas API. A chatbot could only help write the content. An app made with AI still had to be put in place by hand. An agent writes the page, builds an interactive demo, publishes it in Canvas, makes it the home page, hides the navigation and checks that it works. You check that it is the right thing. He called this, finally, the death of the VLE as a place where you go to upload and dump files.
Web book
The same agents change how he shares and publishes. His Czech Language Navigator, a large Word textbook he gave away online for years, is now a web book. Readers can use a reading ruler, change the font, listen to the text and share a highlight as a link or an image. Its complicated grammar tables can be swapped round, and they animate as the words change. Anyone's agent can take the framework and reuse it for free.
Nature primer
For teaching, he took a Nature primer on AI for science and had his agent build a step-by-step walk-through of its argument, with interactive figures and a narrated video. The video is not generated the way Google generates video. Claude Opus 5.5 wrote it as a web page and rendered it with a synthetic voice, which is how this summary was made too. Students can point their own agent at the page and ask for one like it for another paper.
Agm and archive
Even this meeting's papers changed shape. His agent turned the AGM's PDFs into one navigable web page that members could send as an attachment or publish. His years as a community radio DJ are now a website archive, and the talk itself ran in his own presentation app instead of PowerPoint.
Where to start
In the questions, Dominik was asked how to get started. His advice was to watch someone on YouTube who uses agents for work like yours, then install a desktop agent such as Codex or Claude Code on the twenty-dollar plan and try things step by step. It is now worth learning how your computer works: where files live, how they sync, what hosting and a database are. You no longer have to do that work yourself, but you need to know what to ask for.
Not behind
He also told people not to feel behind. Someone who ignored AI from 2022 until this September would have missed almost nothing, because few lessons from the chatbot era still apply.
Bubble
Asked about an AI bubble, he said prices are rising because there is not enough computing power to serve everyone, not because the technology fails. Whatever happens to the finances, what he built is not AI. It is ordinary software, learning objects that stay on the web, unlike chatbots embedded in a VLE, where someone else decides how much intelligence you get. So create as much as you can.
Responsibility
On energy, his question was whether the use is worthwhile, not whether any use of energy is wrong. In research, cyber security and medicine, he argued, it is irresponsible not to use AI, because the work will be worse without it.
The job
For most people, ignoring agents costs nothing: the world goes on as before. For learning technologists, he said, it costs the purpose of the job, because what AI can do now is what learning technology was always meant to do. Tools like H5P were stopgaps that made hard things a little easier. Now the skills of designing learning experiences can direct agents to build them.