Famous papers worth reading alongside this track. No formal citations — just the originals that shaped how we build agents today.
Foundational agent papers
| Paper |
Why it matters |
Link |
| ReAct: Synergizing Reasoning and Acting in Language Models |
The reason + act loop most agents still use |
arXiv:2210.03629 |
| Toolformer: Language Models Can Teach Themselves to Use Tools |
Self-supervised tool learning — foundation of function calling |
arXiv:2302.04761 |
| Reflexion: Language Agents with Verbal Reinforcement Learning |
Agents that learn from failure within a session |
arXiv:2303.11366 |
| Generative Agents: Interactive Simulacra of Human Behavior |
Multi-agent memory, planning, and social simulation |
arXiv:2304.03442 |
Memory & retrieval
| Paper |
Why it matters |
Link |
| MemGPT: Towards LLMs as Operating Systems |
Virtual context management — paging memory in/out |
arXiv:2310.08560 |
| Retrieval-Augmented Generation for Knowledge-Intensive NLP |
RAG — how agents fetch external knowledge |
arXiv:2005.11401 |
| Paper |
Why it matters |
Link |
| Gorilla: Large Language Model Connected with Massive APIs |
LLMs calling real-world APIs reliably |
arXiv:2305.15334 |
| SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering |
Agent-computer interface design for coding |
arXiv:2405.15703 |
Multi-agent & orchestration
| Paper |
Why it matters |
Link |
| AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation |
Multi-agent conversation framework |
arXiv:2308.08155 |
| MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework |
Role-based multi-agent software teams |
arXiv:2308.00352 |
| Voyager: An Open-Ended Embodied Agent with Large Language Models |
Skill library + lifelong learning loop |
arXiv:2305.16291 |
Evaluation & benchmarks
| Paper |
Why it matters |
Link |
| AgentBench: Evaluating LLMs as Agents |
Standard agent capability benchmark |
arXiv:2308.03688 |
| WebArena: A Realistic Web Environment for Building Autonomous Agents |
Realistic web agent eval environment |
arXiv:2307.13854 |
| τ-bench: A Benchmark for Tool-Agent-User Interaction |
Tool + user interaction under realistic conditions |
arXiv:2406.12045 |
Reasoning & planning
| Paper |
Why it matters |
Link |
| Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Step-by-step reasoning before acting |
arXiv:2201.11903 |
| Tree of Thoughts: Deliberate Problem Solving with Large Language Models |
Branching search over reasoning paths |
arXiv:2305.10601 |
Also see