Multi-Agent Systems

A multi-agent system is made up of several autonomous agents that interact in a common environment—cooperating, coordinating or competing—to solve problems beyond the capacity of a single agent. It helps to start with what an agent is: an entity—a program, a robot—situated in an environment, able to perceive it and act on it autonomously to achieve its goals.

What characterizes them

Their defining features are each agent's autonomy, local view—none has global knowledge of the problem—decentralization—there is no central controller; if there were, the system would reduce to a monolithic architecture—and interaction among agents, whether communicating directly or acting on the shared environment. Moreover, each agent may have its own goals, which sometimes conflict.

Coordination, cooperation and competition

That is why multi-agent systems revolve around coordination: mechanisms for one agent's actions to fit with others'. Agents can cooperate toward a common goal or compete in defense of their own interests, and they communicate through agreed protocols. Game theory provides the formal framework for analyzing those strategic interactions.

Where they are used

They have established uses in simulation (climate, epidemics, crowds), swarm robotics and drone coordination, markets and auctions, and logistics. As a more recent application, since around 2023 multi-agent systems based on large language models have grown, in which several language agents divide up roles and collaborate on complex tasks; it is an active but still immature field, with challenges of coordination and reliability.