CS 329Z: Engineering AI Agents (Stanford, Fall 2026)
by Stanford University (Diyi Yang, Michael Ryan, John Yang)
A from-scratch blueprint for building, evaluating and securing agentic systems before you reach for a framework.
Overview
CS 329Z: Engineering AI Agents is a 3-unit Stanford course taught by Diyi Yang, Michael Ryan and John Yang in autumn quarter 2026, meeting Mondays and Wednesdays from September 23 to December 2, with final demos December 7 to 11. It teaches how to build agentic systems by progressing from simple LLM pipelines to compound AI systems to autonomous agents, with emphasis on problem decomposition, component selection, data collection and evaluation design. The public lecture schedule runs from agentic systems foundations and LLMs for builders (APIs, structured I/O, context engineering) through retrieval-augmented generation, tool use and function calling, frameworks such as DSPy, LangChain and LlamaIndex, and design patterns including ReAct, plan-and-execute and reflection. Later weeks cover memory and multi-agent orchestration, optimization through prompts, fine-tuning and test-time compute, data selection and quality for agentic systems, evaluation fundamentals and benchmark design, LLM-as-judge infrastructure, agent security and guardrails, coding and software agents, and proactive mixed-initiative agents. Core components such as RAG, tool use and agent loops are built from scratch before frameworks are introduced. Homework 1 has students build a company's internal AI assistant without agent frameworks, combining LLM pipelines, retrieval over a corporate email archive, tools, terminal access, memory and human-in-the-loop interaction. Homework 2 builds an evaluation suite with code-based graders, LLM-as-judge evaluation, benchmark tasks and error analysis, and each homework is followed by a short quiz on design decisions. A quarter-long project is worth 50% of the grade. Recordings are restricted to enrolled students on Canvas, so outside learners get the syllabus, schedule and assignment specifications rather than videos.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 10-week Stanford quarter (Sep 23 to Dec 11, 2026), two 80-minute lectures/week; recordings enrolled-only
- Provider
- Stanford University (Diyi Yang, Michael Ryan, John Yang)
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Build RAG, tool use and agent loops from scratch without agent frameworks
- ✓Apply agent design patterns such as ReAct, plan-and-execute and reflection
- ✓Design memory and orchestration for multi-agent systems that coordinate on tasks
- ✓Optimize agentic systems through prompts, fine-tuning and test-time compute
- ✓Select and curate data for building and improving agentic systems
- ✓Build evaluation suites with code-based graders, LLM-as-judge and benchmark tasks
- ✓Add security guardrails and understand coding agents and proactive mixed-initiative agents
Highlights
- •Builds core agent components from scratch before introducing DSPy, LangChain and LlamaIndex
- •Homework 1 is a realistic build: an internal company assistant over an email archive with terminal access and memory
- •Treats data and evaluation as first-class topics, with dedicated lectures on benchmark design and LLM-as-judge
- •Instructors are associated with DSPy (Michael Ryan) and SWE-agent (John Yang), according to independent course comparisons
- •Caveat: lecture recordings are Canvas-only, so external learners work from the public schedule and assignment specs
Who It’s For
Best For
- ✓AI engineers moving from single LLM calls to production agent systems
- ✓Engineers who want to understand agents below the framework layer
- ✓Engineering leads designing evaluation and guardrails for agent products
Prerequisites
- •An NLP course such as Stanford CS224N, CS224U, CS224V or CS336, or equivalent background
- •Solid Python and hands-on experience calling LLM APIs
FAQ
What is CS 329Z: Engineering AI Agents (Stanford, Fall 2026)?
Stanford's autumn 2026 course on engineering agentic systems, taught by Diyi Yang, Michael Ryan and John Yang. It is for engineers with an NLP background who want a rigorous blueprint for building RAG, tool-using and multi-agent systems from scratch, then evaluating, optimizing and securing them, before reaching for a framework like DSPy or LangChain.
Is CS 329Z: Engineering AI Agents (Stanford, Fall 2026) free?
CS 329Z: Engineering AI Agents (Stanford, Fall 2026) is free to access.
What level is CS 329Z: Engineering AI Agents (Stanford, Fall 2026) for?
CS 329Z: Engineering AI Agents (Stanford, Fall 2026) is aimed at a advanced audience. Recommended background: An NLP course such as Stanford CS224N, CS224U, CS224V or CS336, or equivalent background, Solid Python and hands-on experience calling LLM APIs.
How long does CS 329Z: Engineering AI Agents (Stanford, Fall 2026) take?
Expect roughly 10-week Stanford quarter (Sep 23 to Dec 11, 2026), two 80-minute lectures/week; recordings enrolled-only. Most learners work through it at their own pace.
What will I learn from CS 329Z: Engineering AI Agents (Stanford, Fall 2026)?
You'll learn: Build RAG, tool use and agent loops from scratch without agent frameworks; Apply agent design patterns such as ReAct, plan-and-execute and reflection; Design memory and orchestration for multi-agent systems that coordinate on tasks; Optimize agentic systems through prompts, fine-tuning and test-time compute; Select and curate data for building and improving agentic systems; Build evaluation suites with code-based graders, LLM-as-judge and benchmark tasks; Add security guardrails and understand coding agents and proactive mixed-initiative agents.
Topics
Sources
This page was written from 2 sources, 1 on domains other than cs329z.stanford.edu.