Multi-Agent Systems
Multiple AI agents that cooperate, compete, or communicate to solve problems beyond any single agent's capability.
Multi-Agent Systems (MAS) refers to a paradigm in which multiple autonomous AI agents operate concurrently, interacting through cooperation, competition, or negotiation to accomplish tasks that exceed the capacity of any individual agent. Each agent may have its own knowledge base, goals, and decision-making logic, and the system-level behavior emerges from the interactions between agents rather than from any single controller. MAS research spans distributed AI, game theory, and more recently, LLM-based agent architectures where language models serve as the reasoning engine for each agent.
The fundamental mechanics involve communication protocols (message passing, shared state, tool-sharing), coordination mechanisms (centralized orchestration vs. decentralized consensus), and conflict resolution strategies. In LLM-based MAS, popular architectures include role-playing frameworks where agents take assigned personas and negotiate outcomes, hierarchical setups where a supervisor agent delegates sub-tasks to specialized workers, and emergent cooperation setups where agents develop shared protocols without explicit programming. The landmark 2023 Stanford "Generative Agents" paper demonstrated 25 LLM-powered agents living in a virtual village, exhibiting believable social behaviors emerging purely from agent-agent interaction.
MAS excels at problems that are naturally distributed — both in data and in authority — such as multi-robot coordination, decentralized finance, supply chain optimization, and collaborative research. The paradigm also naturally handles adversarial settings where agents must reason about other agents' strategies. The primary tradeoff is complexity: designing reliable communication protocols and avoiding agent-level failure modes (e.g., agents getting stuck in infinite negotiation loops or amplifying each other's errors) is significantly harder than building a single-agent system.
Key open questions include how to ensure alignment and safety when multiple agents interact in ways that produce emergent behaviors not anticipated by any individual designer, and how to scale communication overhead as the number of agents grows. There is also active debate about whether multi-agent cooperation is best designed explicitly through protocols or emerges implicitly from prompting strategies — and whether foundation model agents are fundamentally different from classical software agents in how they coordinate.