Source: the CS329Z course site (cs329z.stanford.edu), Stanford, Fall 2026. Instructors Diyi Yang, Michael Ryan and John Yang. Surfaced by Hao Hoang's LinkedIn post.
What it is
A graduate course on building agentic systems across the whole range from simple LLM pipelines to compound AI systems to autonomous agents. Students build the core pieces (RAG, tool use, agent loops) from scratch first, then see how frameworks such as DSPy abstract them. Coursework is two applied homeworks and a quarter-long project. Slides and readings are on the site; lecture videos are being posted as the quarter runs (eight video links were live on 15 November).
Syllabus at a glance
| Block | Lectures |
|---|---|
| Foundations | What agentic systems are (history, anatomy, compound systems vs agents); LLMs for builders (decoding, attention and context length, post-training, test-time scaling) |
| Building blocks | RAG end to end (chunking, hybrid search, reranking, ColBERT, agentic RAG); tool use (Toolformer, Gorilla, function calling, MCP, BFCL and tau-bench) |
| Frameworks and design | Frameworks and orchestration; design patterns (workflows vs agents, ReAct, CodeAct, scaffolds) |
| Memory and multi-agent | Memory architectures; when to use one agent or many |
| Optimization | Context and prompt optimization and when it pays off |
| Data for agents | What data agents need (traces, demonstrations, feedback); data selection and quality |
| Evaluation | Benchmark design; LLM-as-judge and evaluation infrastructure |
| Guests | Terminal-Bench and SWE-bench teams; a fireside chat with Andrew Ng |
| Safety and applications | Agent safety and guardrails (tool-access privacy, prompt injection); coding agents; personal agents |
Why it is worth a staff engineer's time
Build-first order. Writing the loop, retrieval and tool calling by hand before touching a framework is the fastest way to know what a framework hides and where it will fail.
Data and evaluation get four lectures. Most agent material stops at patterns; this course spends real time on traces, data quality, benchmark design and grader design, which is where production agents succeed or fail.
Benchmark authors teach their own benchmarks. The Terminal-Bench and SWE-bench lectures are the most direct way to learn what those numbers measure and what they miss.
Workflow vs agent is treated as a design choice, not a progression, matching current practice.
Gaps to fill elsewhere
Little on running agents in production: durable execution, long-running sessions, cost control and observability.
Security is one lecture; teams with tool-using agents need deeper threat modelling.
Videos lag the schedule, and the second half of the quarter is not yet posted.
How to use it
Watch lectures 1 to 4 and build a minimal agent with one chat-completion call and your own tool loop.
Use the design-patterns and memory lectures as a vocabulary check for your team.
Take the data and evaluation block as the template for your own agent evals.
Pair with an open harness-engineering course for production concerns.
Related entries
R023 (building a coding agent from scratch), R022 (AI engineering reading list), R081 (LLM engineering toolkit), R086 (agents on long-horizon benchmarks), R064 (measuring whether a skill helps).