Stanford AI Agent Course Is Free on YouTube, and the Comments Are Already Arguing

Stanford CS329Z Engineering AI Agents course videos are available to watch directly on YouTube, and the course website and GitHub are also open. The first lecture starts with the foundations and landscape of agentic systems; the instructors are Diyi Yang, Michael Ryan, and John Yang. The course is broken down by week and walks people through building agents step by step.

Michael Guo recommended the course in a tweet, saying it is underrated on YouTube. His advice is direct: instead of scrolling short videos on the weekend, pick a topic you're interested in and follow along. He also mentioned that Stanford courses usually follow a 10-week syllabus, so you can go from start to finish in 2.5 months. Someone in the comments replied that a free Stanford agent course is right there, yet some people still go buy a $400 cohort.

Course website: https://cs329z.stanford.edu/

YouTube playlist: https://www.youtube.com/watch?v=b4y3rjGpg_g&list=PLVKEyLMx8AmU

A commenter posted the course GitHub: https://github.com/cs329z

Along these lines, Michael Guo also mentioned another Stanford course, CS146S The Modern Software Developer. In the Fall 2026 syllabus published by Mihail Eric, 85% of the previous version's content was scrapped and redone. The 10-week course is basically a capability map for an AI-native software engineer: the underlying principles of coding agents and the agent loop; context engineering, spec-driven development, MCP; agent skills, CLI, AGENTS.md / CLAUDE.md, hooks; how to make a repo agent-ready; agentic code review and security; scheduling background agents in parallel; MCP portal, model routing, permissions, and cost; and in the final week, building a Software Factory. The course partners with projects such as Browserbase, Semgrep, OpenHands, Milvus, CrewAI, Vercel, Warp, Unsloth, and Anyscale, and students submit PRs to real production-scale codebases. The course starts on September 22, and classes and resources are public.

One section of CS329Z is singled out for discussion: Agent memory. Michael Guo turned Lecture 7 into a 9-minute explainer. The core metaphor is clear: an agent only sees what is inside the context window, like a small pool of light in a dark room; everything else is in darkness. A memory system has to decide three things: what to keep, where to store it, and when to bring it back into the light. The video breaks down 4 types of memory: working, episodic, semantic, and procedural; 4 design patterns: Voyager & skill files, MemGPT, Generative Agents, and Mem0; evaluation from LoCoMo to MemoryArena; and unsolved problems including what should be forgotten, poisoned memories, and over-personalization.

The comments are closer to the engineering reality than the syllabus. One commenter said lectures easily skip what happens after compression: stored facts can survive, but persistent rules may drop to zero because no one actively retrieves a rule; rules need a layer that is loaded every session. Michael Guo replied that this is exactly the video's cold open, and the fix is MemGPT's pinned working context: put rules in a persistent layer and facts in search. The tradeoff is that the persistent layer cannot be too large; otherwise you trade forgetting for over-personalization. Another commenter said personal assistants first break on "what to forget": a fact from March, "I'm in San Francisco," will confidently come back in October; giving each memory a last-confirmed timestamp and decaying by type is more useful than swapping in a better retriever. Another commenter mentioned that agentic code review needs auditable execution traces—model, tool calls, approvals, failures, and repair paths can all be replayed—not just the final diff.

This course is not without controversy either. Some commenters pointed out that the videos were generated by an individual using AI, or are just the course slides played with AI voiceover. That reminder is useful. The course website and GitHub are both there; slides, reading papers, and assignments can all be accessed directly. The videos are good for building a map, but to actually get hands-on, you still need to read the materials, run the code, and submit assignments.

Another option is the CMU 11-768 AI Agents course: https://www.cmu-agents.com/#/ . A commenter posted the link.

If you want to test the waters over the weekend, one lecture is enough. Start with CS329Z Lecture 1, get the agent loop and tool calls running, then pick the memory or code review lecture and follow along. Bookmarking won't make an agent grow by itself.