Stanford CS329Z course materials made public; lectures posted to YouTube, two assignments open-sourced
Stanford’s new course CS329Z, “Engineering AI Agents,” has begun publicly releasing all course materials. One of the instructors, Michael Ryan, announced in a post on November 10 that lecture videos are being updated daily on Stanford Online’s YouTube channel, and the code repositories for two assignments have been uploaded to GitHub under the MIT License, which other institutions may use as long as they credit the source.
The course is jointly taught by Computer Science Department faculty member Diyi Yang and PhD students Michael Ryan and John Yang. Lecture 1, “What Is an Agent System,” and Lecture 2, “Large Model Knowledge for Agent Developers,” are now online, and Lecture 1 has more than 90,000 views. Video links have already been added to the course homepage schedule. At the start of the course, the recordings were limited to Canvas internally, but they are now public.
The first assignment requires building an agent harness from scratch: implementing tool calling, context management, a terminal sandbox, cross-session memory, and user approval under the most basic interface, using real email archives for multi-hop question answering, and ultimately being evaluated by simulated users. The full test suite is in the repository, and 48 of the 88 points in the auto-grading system do not invoke a model. Michael Ryan said it can be completed with less than $5 in API credits. The second assignment, designed by John Yang, a co-author of SWE-bench, requires using Harbor to create a Terminal-Bench-style exam question and using mini-SWE-agent to analyze its effectiveness and difficulty.
Course homepage: cs329z.stanford.edu
Lecture 1: youtu.be/DGI4vucR71w
Assignment 1: github.com/cs329z/assignment1-harness
Assignment 2: github.com/cs329z/assignment2-evals
Original source: x.com/michaelryan207/status/2120388990616433083
ABAB AI Insight
The graduate course expands agent system engineering from internal enrollment to public materials and open-source assignments. Stanford CS329Z has made lecture videos and two MIT-licensed assignments public, requiring students to build an agent harness and evaluation task design from scratch without frameworks. The first assignment covers priority judgment, BM25 retrieval, tool calling, context compression, and user approval; the second focuses on constructing Terminal-Bench-style questions and analyzing their effectiveness.
Unlike courses that only provide slide decks, this path makes the complete test suite and reference implementation public, allowing people outside the university to reproduce it at low cost. In similar cases, capital pays closer attention to how building from scratch and evaluation design cultivate agent engineering ability, as well as whether the public materials can be reused in teaching at other institutions. The course includes researchers associated with DSPy and SWE-bench, and the assignments emphasize framework-free implementation and simulated user evaluation.
This public release expands agent engineering instruction from enrollment-restricted teaching to public resources. The beneficiaries are developers self-studying agent development and other teaching institutions; Stanford, meanwhile, expands the course’s influence through public materials. The structural change is that agent harness construction and evaluation task design have shifted from internal assignments to an open-source, reproducible path, while low API costs and local caching lower the barrier to practice.
Cognitive laws
Framework-free construction from scratch and a complete test suite turn agent engineering from framework use into underlying implementation.
Public MIT assignments and daily lectures make graduate course materials directly reusable by other institutions.
Low-cost API calls and local caching reduce the economic barrier to implementing and debugging from scratch.