Multi-Agent System

System agents published

Also known as: MAS, Multi-Agent Architecture, Agent Swarm

Definition

A system composed of multiple agents that interact, collaborate, or compete to achieve individual or collective goals. Agents in a multi-agent system may be homogeneous (same capabilities) or heterogeneous (specialized roles), and may communicate directly with each other or through a shared environment. The system's behavior emerges from agent interactions rather than central control.

What this is NOT

  • Not just parallel execution (agents must interact, not merely run concurrently)
  • Not a workflow with multiple steps (agents have autonomy, steps don't)
  • Not defined by the number of LLM calls (one agent can make many calls)

Alternative Interpretations

Different communities use this term differently:

academic-ai

A computational system where multiple autonomous agents interact within a shared environment, studied for emergent behaviors, game theory applications, and distributed problem-solving.

Sources: Wooldridge: An Introduction to MultiAgent Systems (2009), AAMAS conference proceedings

llm-practitioners

A system where multiple LLM-based agents with different roles or personas collaborate on tasks, often through conversation or message passing. Examples include debate frameworks, role-playing systems, and crew-based architectures.

Sources: AutoGen (Microsoft) documentation, CrewAI documentation, CAMEL paper (2023)

Examples

  • AutoGen conversation between researcher, coder, and critic agents
  • CrewAI crew with manager, researcher, and writer roles
  • Debate framework where agents argue opposing positions
  • Simulation with buyer and seller agents in a market

Counterexamples

Things that might seem like Multi-Agent System but are not:

  • A single agent that reasons in multiple steps (that's an agent loop)
  • Parallel API calls from one agent (no agent interaction)
  • A pipeline where outputs flow one direction without feedback

Relations

  • generalizes agent (A MAS contains multiple agents)
  • overlapsWith agent-orchestration (Orchestration coordinates agents in a MAS)
  • overlapsWith supervisor-agent (Some MAS use supervisor agents for coordination)
  • specializes agentic-system (MAS is one architecture for agentic systems)

See Also

Implementations

Tools and frameworks that implement this concept: