想当 agent engineer？斯坦福这门课值得收藏。

强烈推荐斯坦福的 CS329Z《Engineering AI Agents》，一句话：有志成为 agent engineer 的同学可以看看。

agent 开发正在从调 API 的手艺，变成一门正经工程学科，课程、评测、方法论都在成型。与其在网上东拼西凑，不如跟一套成体系的课走一遍。

Stanford CS329Z | Engineering AI Agents

Source: https://cs329z.stanford.edu/ 1 Wed Sep 23Foundations & LandscapeIntroduction — What Are Agentic Systems?Course logistics, a brief history of agents, and the anatomy of an agent: LLM core, reasoning, memory, tools, and environment. Compound AI systems and workflows vs. agents, plus the key challenges of reliability, training, safety, and evaluation.SlidesReadings:- Zaharia et al. “The Shift from Models to Compound AI Systems.” BAIR Blog (2024).

• Weng. “LLM Powered Autonomous Agents.” Lil’Log (2023).
• Kapoor et al. “AI Agents That Matter.” TMLR (2025).

2 Mon Sep 28LLMs for BuildersLanguage 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.SlidesReadings:- Vaswani et al. “Attention Is All You Need.” NeurIPS (2017).

• Katharopoulos et al. “Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention.” ICML (2020).
• Rajasekaran et al. “Effective Context Engineering for AI Agents.” Anthropic (2025).

2 Wed Sep 30Building BlocksRetrieval-Augmented Generation (RAG)Why retrieval helps, then the RAG pipeline end to end: indexing (chunking, contextual retrieval, RAPTOR, HNSW), retrievers from BM25 and DPR to hybrid search, reranking, and ColBERT, agentic RAG (Self-RAG, Search-R1), and common failure modes. Plus a course project overview.SlidesReadings:- Lewis et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS (2020).

• Karpukhin et al. “Dense Passage Retrieval for Open-Domain Question Answering.” EMNLP (2020).
• Khattab & Zaharia. “ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT.” SIGIR (2020).

3 Mon Oct 5Tool Use & Function CallingWhy LLMs need tools and how they learn to use them: Toolformer, scaling to thousands of APIs with Gorilla and ToolLLM, standard interfaces from REST to function calling to MCP, evaluating tool use with BFCL, ToolBench, and τ-bench, and common failure modes.SlidesReadings:- Schick et al. “Toolformer: Language Models Can Teach Themselves to Use Tools.” NeurIPS (2023).

• Patil et al. “Gorilla: Large Language Model Connected with Massive APIs.” NeurIPS (2024).
• Model Context Protocol Specification, revision 2025-06-18. modelcontextprotocol.io (2025).

3 Wed Oct 7Frameworks & Agent DesignFrameworks & OrchestrationThe data, system, and evals pillars of agent engineering, with live demos of DSPy (signatures, modules, optimizers) and LangGraph (state, nodes, edges). A tour of open-source agent harnesses (mini-SWE-agent, OpenHands, Pi), durable orchestration with Temporal, observability tooling, and best practices.SlidesCodeReadings:- Khattab et al. “DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines.” ICLR (2024).

• LangChain. “LangGraph Overview.” LangChain Docs.
• Arena.ai. “HarnessTax: How Much Does the Harness Matter for Coding Agents?” Arena Blog (2026).

4 Mon Oct 12Agent Design Patterns & ScaffoldsWorkflows vs. agents and the design patterns for each: five composable workflow patterns, reasoning and acting patterns (ReAct, Reflexion, CodeAct, ReWOO), and context-management patterns (event logs, compaction, observation masking, context isolation). Case study of the Pi harness.SlidesReadings:- Yao et al. “ReAct: Synergizing Reasoning and Acting in Language Models.” ICLR (2023).

• Schluntz & Zhang. “Building Effective AI Agents.” Anthropic (2024).
• LangChain. “Context Engineering.” LangChain Blog (2025).

Additional readings- Shinn et al. “Reflexion: Language Agents with Verbal Reinforcement Learning.” NeurIPS (2023).

• Wang et al. “Executable Code Actions Elicit Better LLM Agents.” ICML (2024).

4 Wed Oct 14Memory & Multi-Agent SystemsAgent Memory ArchitecturesWhy agents need memory and the four types (working, episodic, semantic, procedural). Four architecture patterns: skill libraries, tiered hierarchies like MemGPT, reflective consolidation, and memory maintenance. Memory benchmarks and open problems in forgetting, privacy, and over-personalization.SlidesReadings:- Sumers et al. “Cognitive Architectures for Language Agents.” TMLR (2024).

• Packer et al. “MemGPT: Towards LLMs as Operating Systems.” arXiv (2023).
• Park et al. “Generative Agents: Interactive Simulacra of Human Behavior.” UIST (2023).

5 Mon Oct 19Multi-Agent SystemsWhen to use single vs. multi-agent systems and how to build them: orchestration patterns (orchestrator, swarm, hybrid), communication through handoffs, message boards, and the A2A protocol, and coordination infrastructure (task ledgers, shared memory, worktree isolation). Why multi-agent systems fail.SlidesReadings:- Wu et al. “AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation.” COLM (2024).

• Cemri et al. “Why Do Multi-Agent LLM Systems Fail?” NeurIPS Datasets & Benchmarks (2025).
• Kim et al. “Towards a Science of Scaling Agent Systems.” arXiv (2025).

5 Wed Oct 21OptimizationOptimizationWhen optimization pays off and how to do it: context optimization (bootstrapped few-shot, MIPROv2, OPRO, TextGrad, GEPA), weight optimization (SFT on agent rollouts, RL with PPO and GRPO, on-policy distillation), agents as optimizers (Darwin Gödel Machine, autoresearch), and when to optimize prompts vs. weights.SlidesReadings:- Agrawal et al. “GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning.” ICLR (2026).

• Opsahl-Ong et al. “Optimizing Instructions and Demonstrations for Multi-Stage Language Model Programs.” EMNLP (2024).
• Shao et al. “DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.” arXiv (2024).

6 Mon Oct 26📺 Guest Lecture: Terminal-BenchRyan Marten,Alex ShawRyan Marten and Alex Shaw (Laude Institute), co-creators of Terminal-Bench, on benchmarking how well AI agents accomplish real-world tasks in a terminal environment.SlidesReadings:- Terminal-Bench project site (leaderboard, tasks, and harness).

• Merrill et al. “Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces.” ICLR (2026).

6 Wed Oct 28Data for Agentic SystemsWhat Data Do Agents Need?Traces, demonstrations, and feedback, data for optimization vs. evaluation, data flywheels, synthetic data generation, and collecting data from human-agent interaction.SlidesReadings:- 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 2Data Selection & QualityFinding 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 4Evaluation for Agentic SystemsEvaluation Fundamentals & Benchmark DesignWhy 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📺 Fireside Chat with Andrew NgAndrew NgA fireside chat with Andrew Ng, hosted by Diyi Yang. Professor Ng is the founder of DeepLearning.AI, co-founder of Coursera, founding lead of Google Brain, and a Professor of Computer Science at Stanford.8 Wed Nov 11LLM-as-Judge & Evaluation InfrastructureThe 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).

9 Mon Nov 16📺 Guest Lecture: SWE-benchCarlos Jimenez,Alexander WettigCarlos Jimenez and Alexander Wettig (Anthropic), creators of SWE-bench, on the evolution and future of evaluations for coding capabilities and beyond.Readings:- Jimenez et al. “SWE-bench: Can Language Models Resolve Real-World GitHub Issues?” ICLR (2024).

9 Wed Nov 18SafetyAgent Safety & GuardrailsPrivacy 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).

10 Mon Nov 23No Class — Thanksgiving Recess10 Wed Nov 25No Class — Thanksgiving Recess11 Mon Nov 30Coding Agents & Proactive AgentsCoding & Software AgentsHow coding agents work end to end: the SWE-agent, Claude Code, and OpenHands architectures, scaffolds as design decisions, SWE-bench and the 4-tuple framework in practice, and 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).

11 Wed Dec 2Proactive AgentsFrom reactive to proactive agents: 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).

— Finals Week (Dec 7–11)Open Problems & Final DemosFinal Project Demo DayHeld during the end-quarter examination period. Exact time and location TBA.