第7讲：Agent
参考文献
第 7 讲 Agent 的核心论文、综述、框架文档与规范
推荐综述（先读这两篇）
-
Lilian Weng. LLM Powered Autonomous Agents. OpenAI Blog, 2023-06-23.
https://lilianweng.github.io/posts/2023-06-23-agent/
建立 Planning / Memory / Tool Use 三要素框架的奠基博文，所有后续工作的参考起点。 -
Wang, L. 等. A Survey on Large Language Model based Autonomous Agents. Frontiers of Computer Science, 2024（arXiv: 2308.11432）.
https://arxiv.org/abs/2308.11432
该领域第一篇中文主导的系统综述（人大，86 页），覆盖构建、应用、评估。 -
Xi, Z. 等. The Rise and Potential of Large Language Model Based Agents: A Survey. 中国科学:信息科学, 2024（arXiv: 2309.07864）.
https://arxiv.org/abs/2309.07864
复旦 86 页巨作，从哲学起源到现代实现的全景综述。
核心论文（按本讲章节组织）
7.1 经典 Agent 架构
- Fikes, R. E. & Nilsson, N. J. STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving. Artificial Intelligence, 1971.
- Brooks, R. A. A Robust Layered Control System for a Mobile Robot. IEEE Journal of Robotics and Automation, 1986.
- Brooks, R. A. Intelligence Without Representation. Artificial Intelligence, 1991.
- Rao, A. S. & Georgeff, M. P. BDI Agents: From Theory to Practice. ICMAS, 1995.（引用量 5,000+）
- Wooldridge, M. & Jennings, N. R. Intelligent Agents: Theory and Practice. Knowledge Engineering Review, 1995.
- Russell, S. & Norvig, P. Artificial Intelligence: A Modern Approach. 4th Edition, Pearson, 2020.
7.2 LLM Agent 的四要素
- Wei, J. 等. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022.
https://arxiv.org/abs/2201.11903 - Yao, S. 等. ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.
https://arxiv.org/abs/2210.03629 - Schick, T. 等. Toolformer: Language Models Can Teach Themselves to Use Tools. NeurIPS 2023.
https://arxiv.org/abs/2302.04761 - Karpas, E. 等. MRKL Systems: A Modular, Neuro-Symbolic Architecture. AI21 Labs, 2022.
https://arxiv.org/abs/2205.00445
"Modular Reasoning, Knowledge and Language"，工具使用范式的早期形式化。
7.3 规划与推理
- Wang, L. 等. Plan-and-Solve Prompting. ACL 2023.
https://arxiv.org/abs/2305.04091 - Zhou, D. 等. Least-to-Most Prompting Enables Complex Reasoning in Large Language Models. ICLR 2023.
- Xu, B. 等. ReWOO: Decoupling Reasoning from Observations. 2023.
https://arxiv.org/abs/2305.18323 - Yao, S. 等. Tree of Thoughts: Deliberate Problem Solving with Large Language Models. NeurIPS 2023.
https://arxiv.org/abs/2305.10601 - Besta, M. 等. Graph of Thoughts: Solving Elaborate Problems with Large Language Models. AAAI 2024.
- Zhou, A. 等. Language Agent Tree Search. ICML 2024.
- Shinn, N. 等. Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023.
https://arxiv.org/abs/2303.11366 - Madaan, A. 等. Self-Refine: Iterative Refinement with Self-Feedback. NeurIPS 2023.
- Hao, S. 等. Reasoning with Language Model is Planning with World Model (RAP). EMNLP 2023.
- Huang, X. 等. Understanding the Planning of LLM Agents: A Survey. 2024.
https://arxiv.org/abs/2402.02716
7.4 记忆与工具
- Packer, C. 等. MemGPT: Towards LLMs as Operating Systems. 2023.
https://arxiv.org/abs/2310.08560
（后续开源为 Letta 框架） - Patil, S. G. 等. Gorilla: Large Language Model Connected with Massive APIs. NeurIPS 2024.
- Qin, Y. 等. ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs. ICLR 2024.
- Li, M. 等. API-Bank: A Comprehensive Benchmark for Tool-Augmented LLMs. EMNLP 2023.
- Wang, X. 等. Executable Code Actions Elicit Better LLM Agents (CodeAct). ICML 2024.
- Asai, A. 等. Self-RAG: Learning to Retrieve, Generate, and Critique. ICLR 2024 Oral.
- Sarthi, P. 等. RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval. ICLR 2024.
工具调用与协议规范：
- OpenAI Function Calling 文档: https://platform.openai.com/docs/guides/function-calling
- Model Context Protocol (MCP) 规范（Anthropic, 2024-11）: https://modelcontextprotocol.io
"AI 的 USB-C"，2025 年已被 OpenAI、Google DeepMind 采用，2025-12 捐赠至 Linux Foundation。 - MCP GitHub: https://github.com/modelcontextprotocol
7.5 多 Agent 协作
- Park, J. S. 等. Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023.
https://arxiv.org/abs/2304.03442
25 个 agent 在沙盒中自发组织情人节派对的经典工作。 - Wu, Q. 等. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. 2023.
https://arxiv.org/abs/2308.08155 - Hong, S. 等. MetaGPT: Meta Programming for Multi-Agent Collaborative Framework. ICLR 2024 Oral.
https://arxiv.org/abs/2308.00352 - Qian, C. 等. ChatDev: Communicative Agents for Software Development. ACL 2024.
- Li, G. 等. CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society. NeurIPS 2023.
- Du, Y. 等. Improving Factuality and Reasoning in Language Models through Multi-Agent Debate. ICML 2024.
https://arxiv.org/abs/2305.14325 - Chen, W. 等. AgentVerse: Facilitating Multi-Agent Collaboration. ICLR 2024.
- Shen, Y. 等. HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in HuggingFace. NeurIPS 2023.
- Guo, T. 等. Large Language Model based Multi-Agents: A Survey of Progress and Challenges. IJCAI 2024.
其他重要 Agent 系统
- Wang, G. 等. Voyager: An Open-Ended Embodied Agent with Large Language Models. TMLR 2023.
- Nakano, R. 等. WebGPT: Browser-assisted Question-Answering with Human Feedback. OpenAI, 2021.
- AutoGPT. Toran Bruce Richards, 2023-03. https://github.com/Significant-Gravitas/AutoGPT（18 万+ star）
- BabyAGI. Yohei Nakajima, 2023-03.
- Yang, J. 等. SWE-Agent: Agent-Computer Interfaces Enable Automated Software Engineering. NeurIPS 2024.
- Wang, X. 等. OpenHands: An Open Platform for AI Software Developers as Generalist Agents. ICLR 2025.
评测基准
- Liu, X. 等. AgentBench: Evaluating LLMs as Agents. ICLR 2024.
- Zhou, S. 等. WebArena: A Realistic Web Environment for Building Autonomous Agents. ICLR 2024.
- Jimenez, C. E. 等. SWE-bench: Can Language Models Resolve Real-World GitHub Issues?. ICLR 2024 Oral.
- Mialon, G. 等. GAIA: A Benchmark for General AI Assistants. Meta & HuggingFace, 2023.
- Xie, T. 等. OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments. NeurIPS 2024.
- Yao, S. 等. τ-bench: A Benchmark for Tool-Agent-User Interaction. ICLR 2025.
提出 pass^k 指标，揭示 agent 的可靠性危机。 - Chan, J. S. 等. MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering. OpenAI, 2024.
框架文档（本讲实验会用到）
- LangGraph 文档: https://langchain-ai.github.io/langgraph/
- AutoGen 文档: https://microsoft.github.io/autogen/
- CrewAI 文档: https://docs.crewai.com
- smolagents 文档: https://huggingface.co/docs/smolagents
- OpenAI Agents SDK: https://platform.openai.com/docs/guides/agents
- Google Agent Development Kit (ADK): https://google.github.io/adk-docs/（2025-04 开源，Apache 2.0）
- Microsoft Agent Framework 1.0: 2026-04 整合 AutoGen + Semantic Kernel
安全与可靠性
- Hubinger, E. 等. Sleeper Agents: Training Deceptive LLMs that Persist through Safety Training. Anthropic, 2024.
- Sun, L. 等. TrustLLM: Trustworthiness in Large Language Models. 2024.
- Ji, J. 等. AI Alignment: A Comprehensive Survey. 2023.
- Zhang, Z. 等. LLM-based Agents are Affected by Hallucination: A Survey. arXiv: 2509.18970, 2025.
- Levy, S. 等. ST-WebAgentBench: A Benchmark for Evaluating Safety and Trustworthiness in Web Agents. 2024.
课程内部材料
- AI智能体文献综述_从经典架构到LLM驱动的自主系统.md — 本讲的上游材料，覆盖从 STRIPS 到 MCP 的完整脉络，建议作为课后精读材料。
阅读建议
最短阅读路径（3 小时）：
- Lilian Weng 博文（建立总体框架）
- ReAct 原论文（范式基础）
- MCP 规范页（工具生态）
- Du 等人 Multi-Agent Debate 论文（多 agent 代表作）
- τ-bench 论文（看清 agent 的可靠性真相）
2026 年的认知更新：本讲引用的许多论文发表于 2022–2024 年。Agent 领域变化极快，建议每季度跟进一次 arXiv 的 cs.AI agent 方向综述。重点关注：可靠性（pass^k）、安全性（ST-Web）、成本效率（MLE-bench 指出一次评估约 $48k）。