Stanford Put Its Self-Improving AI Agents Course on YouTube for Free

Stanford has posted the core lectures of one of its newest and most current graduate courses — CS329A, Self-Improving AI Agents, taught in Autumn 2025 by Azalia Mirhoseini and Aakanksha Chowdhery — free on YouTube: nine lectures, roughly ten hours, no paywall, no enrolment form. The subject is the live frontier of applied AI: how to build agents that get better at their jobs by checking their own work, learning from feedback and searching for better answers, rather than by being retrained from scratch. It is a real Stanford seminar, with the real reading list attached. It is also, in the ways that matter to a career, not a Stanford course at all — and the distance between those two facts is the story.

What CS329A actually teaches, in plain English

The course follows one question through the term: how does an AI system improve itself? It starts with test-time compute — the discovery that a model often gets better answers simply by thinking longer, sampling many candidate answers and picking among them, than by being made bigger. Then comes verification: training systems to check work step by step, because an agent that can catch its own errors can improve without a human grading every attempt. From there the lectures move through learning from feedback — code that is run and judged by its execution, arguments refined against a constitution of principles — into multi-step reasoning and planning, reinforcement learning at scale, agents that design other agents, deep-research agents that search and synthesise, and finally the hard problem of evaluation: how to measure whether an agent can be trusted with long tasks that take hours, not seconds. The course site lists the primary papers for every session, which means a motivated viewer can read the same sources the seminar discusses.

Who it is for, and who will bounce off it

This is a graduate research seminar, and it does not pretend otherwise. It assumes fluency with modern machine learning and language models — the level of Stanford’s own NLP and LLM courses — plus Python, comfort reading arXiv papers, and enough AI tooling experience to follow discussions of agent frameworks and API-based experiments. Enrolled students do three homeworks, work in teams on an original research project with API credits provided, and present midterm results. None of that exists for the YouTube viewer: no feedback, no project supervision, no credits, no certificate, and no credit toward any degree. The course’s own policies underline the point — auditing is not allowed even for people physically at Stanford. What the viewer gets is the lectures and the reading list. For a working ML engineer or a strong final-year student, that is a great deal. For a beginner, it will feel like being handed the last chapter of a textbook: the map is real, but the road to it is elsewhere.

The “$850K skill” framing, and what is actually true

The lectures spread online wrapped in a salary claim: coverage sharing them framed the course as an “$850K skill” — the idea being that agent-building expertise commands extraordinary pay, and that the expertise is now sitting free on YouTube. Treat the number as marketing. No course page, Stanford’s included, promises any salary, and no employer hires a playlist. The defensible version is quieter and still significant: the skills CS329A teaches — evaluating agents, building verification loops, orchestrating tools — are precisely the skills companies now struggle to hire for as agents move into production, where questions of agent identity and accountability are becoming engineering requirements rather than abstractions. A hiring manager cannot verify that you watched ten hours of lectures. They can verify what you built with them. The course is best understood as unusually good raw material for that building.

Part of a bigger opening

Stanford is not alone, and this is not a sudden conversion. Elite universities have been publishing courseware for two decades, and the releases are getting closer to the research frontier. In the same weeks, the University of Michigan’s robotics department drew wide attention for putting its course materials — lectures, textbooks and projects for several courses, published freely online — within anyone’s reach. Other institutions are experimenting at the admissions end instead, guaranteeing places to qualified applicants rather than rationing them. The pattern is consistent: the content of elite education is becoming abundant, while the credential, the supervision and the network remain scarce. CS329A is a clean example of the split — a frontier syllabus, free to the world, attached to a seminar almost nobody can join.

So watch it, if you have the background: ten hours with two researchers teaching the field they are helping to invent, with the papers linked and the guest lectures drawn from the labs building these systems. Just keep the ledger straight. The lecture-hall door is open. The diploma office is not, and Stanford never said otherwise. Information verified through 1 November 2026.

Photo: Yan Krukau, via Pexels.

Sources

1. Stanford CS329A, Self-Improving AI Agents — course site, syllabus and policies (Autumn 2025): https://cs329a.stanford.edu/
2. CS329A lecture playlist, Stanford Online, YouTube (nine lectures, free): https://www.youtube.com/playlist?list=PLoROMvodv4rO1dBrD2POvwRhJxIJWIonX
3. Michigan Robotics — about the department (several courses freely available online): http://robotics.umich.edu/about/