Stanford CS329Z | Engineering AI Agents

Instructors

Diyi Yang, Instructor

Michael Ryan, Instructor

John Yang, Instructor

Welcome!

The shift from monolithic language models to compound AI systems — systems with multiple interacting components including LLMs, retrievers, tools, and optimizers — represents a fundamental change in how AI applications are built for people. This course teaches students how to engineer agentic systems: the full spectrum from simple LLM pipelines to compound AI systems to autonomous agents. Students will learn to pick what types of problems to focus on, decompose problems, select appropriate components, collect and curate data, build evaluations, and reason about the design tradeoffs that arise when building these systems in practice.

Students first build core components (RAG, tool use, agent loops) from scratch, then learn how frameworks like DSPy abstract these patterns. Through two fully applied homework assignments and a quarter-long project, students gain hands-on experience building, optimizing, and evaluating agentic systems.

For more information about coursework, office hours, and logistics, please refer to the Logistics page. You can also view course dates on the Google Calendar.

Class Schedule

Note: the schedule is tentative and subject to change. Classes meet Mondays and Wednesdays, 1:30–2:50 p.m., in Skilling Auditorium. Lecture materials will be linked here as they are released.

Week	Date	Lecture	Course Material
1	Wed Sep 23	Foundations & Landscape Introduction — What Are Agentic Systems? The spectrum from monolithic models to compound AI systems to agents; when compound systems win; the three engineering challenges (decomposition, data, evaluation); course logistics. Slides	Readings: Zaharia et al. "The Shift from Models to Compound AI Systems." BAIR Blog (2024). Additional readings Ng. "Agentic Design Patterns Part 1: Four AI Agent Strategies…" The Batch (2024). Si et al. "Towards Execution-Grounded Automated AI Research." arXiv (2026).
2	Mon Sep 28	LLMs for Builders Language model internals for agent builders: decoding strategies, transformer attention (including linear and hybrid) and context length, the pretraining-to-post-training pipeline, inference and test-time scaling, structured I/O and constrained decoding, and context engineering. Slides	Readings: Schluntz & Zhang. "Building Effective Agents." Anthropic (2024). Additional readings Rajasekaran et al. "Effective Context Engineering for AI Agents." Anthropic (2025).
2	Wed Sep 30	Building Blocks Retrieval-Augmented Generation (RAG) Grounding and hallucination, embeddings and vector stores, chunking strategies, hybrid search, cross-encoders and late interaction (ColBERT). Hands-on: build a RAG pipeline from scratch. Slides	Readings: Lewis et al. "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." NeurIPS (2020). Additional readings Khattab & Zaharia. "ColBERT: Efficient and Effective Passage Search via Late Interaction over BERT." SIGIR (2020).
3	Mon Oct 5	Tool Use & Function Calling The REPL, function-calling APIs, the Model Context Protocol (MCP), designing good tools, code-execution sandboxes, error handling and retries. Hands-on: build a tool-using system from scratch. Slides	Readings: Model Context Protocol Specification. The Linux Foundation (2025).
3	Wed Oct 7	Frameworks & Agent Design Frameworks & Orchestration DSPy (signatures, modules, optimizers), LangChain/LangGraph, LlamaIndex; what frameworks abstract vs. what you built from scratch; choosing the right level of abstraction. Slides	Readings: Khattab et al. "DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines." ICLR (2024).
4	Mon Oct 12	Agent Design Patterns & Scaffolds The workflows-vs-agents taxonomy, five composable workflow patterns, agent patterns (ReAct, plan-and-execute, reflection), and scaffolds as design decisions. Slides	Readings: Yao et al. "ReAct: Synergizing Reasoning and Acting in Language Models." ICLR (2023).
4	Wed Oct 14	Memory & Multi-Agent Systems Agent Memory Architectures Short- vs. long-term memory, memory as tool-based actions, the file system as externalized memory, structured memory paradigms, and cross-agent memory. Slides	Readings: Packer et al. "MemGPT: Towards LLMs as Operating Systems." arXiv (2023). Additional readings Letta. "Agent Memory: How to Build Agents that Learn and Remember." Letta Blog (2025). Chhikara et al. "Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory." arXiv (2025). Park et al. "Generative Agents: Interactive Simulacra of Human Behavior." UIST (2023).
5	Mon Oct 19	Multi-Agent Systems Single vs. multi-agent architectures, orchestration patterns, handoffs and state transfer, delegation and collaboration patterns, and the challenges of coordination and error propagation.	Readings: Wu et al. "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation." COLM (2024). Additional readings Cemri et al. "Why Do Multi-Agent LLM Systems Fail?" arXiv (2025). Neubig. "Don't Sleep on Single-agent Systems." OpenHands Blog (2024). Liu et al. "A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration." COLM (2024).
5	Wed Oct 21	Optimization Optimization The landscape from prompts to fine-tuning; prompt optimization (GEPA, MIPROv2, OPRO, TextGrad); test-time compute scaling; LoRA/QLoRA; distillation; RLHF/DPO at a high level; and when to optimize prompts vs. weights vs. inference compute.	Readings: Snell et al. "Scaling LLM Test-Time Compute Optimally…" ICLR (2025). Agrawal et al. "GEPA: Reflective Prompt Evolution Can Outperform RL." arXiv (2026). Additional readings Soylu et al. "Fine-Tuning and Prompt Optimization: Two Great Steps that Work Better Together." EMNLP (2024). Opsahl-Ong et al. "Optimizing Instructions and Demonstrations for Multi-Stage Language Model Programs." arXiv (2024).
6	Mon Oct 26	📺 Guest Lecture (TBA)
6	Wed Oct 28	Data for Agentic Systems What Data Do Agents Need? Traces, demonstrations, and feedback; data for optimization vs. evaluation; data flywheels; synthetic data generation; collecting data from human-agent interaction.	Readings: Shankar. "Data Flywheels for LLM Applications." (2024). Additional readings Tan et al. "Large Language Models for Data Annotation and Synthesis: A Survey." EMNLP (2024).
7	Mon Nov 2	Data Selection & Quality Finding maximally informative data, filtering and selection strategies, tiny-but-targeted benchmarks, annotation practices, quality assessment, and building datasets from agent traces.	Readings: Yang et al. "SWE-smith: Scaling Data for Software Engineering Agents." NeurIPS D&B (2025). Shankar et al. "Who Validates the Validators? Aligning LLM-Assisted Evaluation with Human Preferences." UIST (2024). Additional readings Zhou et al. "LIMA: Less Is More for Alignment." NeurIPS (2023).
7	Wed Nov 4	Evaluation for Agentic Systems Evaluation Fundamentals & Benchmark Design Why evals are hard, the 4-tuple framework (request, environment, stopping criteria, scorer), designing each component, properties of good benchmarks, realistic scaffolding, and reliability dimensions.	Readings: Zhu et al. "Establishing Best Practices for Building Rigorous Agentic Benchmarks." (2025). Additional readings Press. "How to Build Good Language Modeling Benchmarks." (2024). Polo et al. "tinyBenchmarks: Evaluating LLMs with Fewer Examples." ICML (2024).
8	Mon Nov 9	LLM-as-Judge & Evaluation Infrastructure The three grader types, designing judge prompts, known biases, pairwise vs. pointwise evaluation, non-determinism metrics (pass@k vs. pass^k), harness design, and Anthropic's 8-step roadmap.	Readings: Grace et al. "Demystifying Evals for AI Agents." Anthropic (2026). Zheng et al. "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena." NeurIPS (2023). Ryan et al. "AutoMetrics: Approximate Human Judgements with Automatically Generated Evaluators." ICLR (2026). Additional readings Zhu et al. "AutoLibra: Agent Metric Induction from Open-Ended Human Feedback." ICLR (2026).
8	Wed Nov 11	Safety Agent Safety & Guardrails Privacy risks of tool access, prompt injection (including indirect injection), red-teaming, sandboxing and permission models, output guardrails, liability considerations, responsible deployment, and human-in-the-loop patterns.	Readings: Shao et al. "PrivacyLens: Evaluating Privacy Norm Awareness of LMs in Action." NeurIPS D&B (2024). Zhang & Yang. "Searching for Privacy Risks in LLM Agents via Simulation." arXiv (2025). Li. "Agentic LLMs as Powerful Deanonymizers." arXiv (2026). Additional readings Wen et al. "Contextualized Privacy Defense for LLM Agents." arXiv (2026). OpenAI. "Understanding Prompt Injections: A Frontier Security Challenge." (2025). Anthropic. "Anthropic's Responsible Scaling Policy." (2023).
9	Mon Nov 16	📺 Guest Lecture (TBA)
9	Wed Nov 18	Coding Agents & Proactive Agents Coding & Software Agents How coding agents work end-to-end; SWE-agent, Claude Code, and OpenHands architectures; scaffolds as design decisions; SWE-bench and the 4-tuple framework in practice; the future of software development with agents.	Readings: Yang et al. "SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering." NeurIPS (2024). Wang et al. "OpenHands: An Open Platform for AI Software Developers." ICLR (2025). Additional readings Anthropic. "Claude Code: Best Practices for Agentic Coding." (2025). Young. "Effective Harnesses for Long-Running Agents." Anthropic (2025). Jimenez et al. "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?" ICLR (2024).
10	Mon Nov 23	No Class — Thanksgiving Recess
10	Wed Nov 25	No Class — Thanksgiving Recess
11	Mon Nov 30	Proactive Agents From reactive to proactive; General User Models (GUM); Next Action Prediction; open-source proactive personal agents; privacy and trust implications; and when agents should initiate vs. wait (mixed initiative).	Readings: Shaikh et al. "Creating General User Models from Computer Use." UIST (2025). Additional readings Shaikh et al. "Learning Next Action Predictors from Human-Computer Interaction." arXiv (2026). Steinberger. "OpenClaw: Open-source proactive AI agent." (2025).
11	Wed Dec 2	Open Problems & Final Demos Frontiers & Open Problems Multimodal agents, web agents and computer use, science agents, long-running agent architectures, production and observability (tracing, monitoring, cost management), and open problems in reliability, scalability, and interpretability.	Additional readings Xie et al. "OSWorld: Benchmarking Multimodal Agents in Real Computer Environments." NeurIPS D&B (2024). Yao et al. "WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents." NeurIPS (2022).
—	Finals Week (Dec 7–11)	Final Project Demo Day Held during the end-quarter examination period; exact time and location TBA.

Deadlines

All times are Pacific. Deadlines are subject to change; any changes will be announced in class and posted here.

Week	Deadline	Date	Time
3	HW1 Released	Mon Oct 5	—
3	Project Proposal Due	Fri Oct 9	11:59 p.m.
6	HW2 Released	Mon Oct 26	—
6	HW1 Due	Fri Oct 30	11:59 p.m.
7	Midpoint Demo Video Due	Wed Nov 4	11:59 p.m.
7	Midway Report Due	Fri Nov 6	11:59 p.m.
8	Paper Video Due (10 min)	Fri Nov 13	11:59 p.m.
9	HW2 Due	Fri Nov 20	11:59 p.m.
11	Paper Video Peer Reviews Due (2 videos)	Mon Nov 30	11:59 p.m.
Finals	Final Report Due & Final System Demo	Dec 7–11	TBA

CS 329Z: Engineering AI Agents

Stanford / Fall 2026

Instructors

Welcome!

Class Schedule

Deadlines

Sponsors