Definition
Multi-agent systems have multiple AI agents acting in a shared environment, sometimes collaborating, sometimes competing, often both. They open up new capabilities (specialization, parallelism) and new failure modes (collusion, runaway loops, emergent strategies) that single-agent setups don’t have.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.
- THREAD: Thinking Deeper with Recurrent Multi-Hop Reasoning
- MAST: A Study of Multi-Agent LLM System Failures
- AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents
- STORM: Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking
- LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion
- Autonomous Chemical Research with Large Language Models
- AgentLab: Scalable and Extensible Framework for Reproducible LLM Agent Experiments
- Emergent Communication through Negotiation
- Improving Factuality and Reasoning in Language Models through Multiagent Debate
- Generative Agents: Interactive Simulacra of Human Behavior
- Mixture-of-Agents Enhances Large Language Model Capabilities
- GPTSwarm: Language Agents as Optimizable Graphs
- TextGrad: Automatic 'Differentiation' via Text
- Optimizing Instructions and Demonstrations for Multi-Stage Language Model Programs