Multi-Agent System

A multi-agent system is a collection of autonomous agents that interact to achieve goals that no single agent could achieve alone. The agents may be software processes, robots, biological organisms, or human actors. The defining feature is not the number of agents but the emergence of collective behavior from local interaction rules — a systems-level phenomenon that cannot be predicted from the properties of individual agents.

The design of multi-agent systems faces a fundamental tension inherited from bounded rationality: each agent has limited information, limited computation, and limited time. The system as a whole must therefore decompose problems, coordinate solutions, and resolve conflicts without centralized control. This produces architectures that mirror the nearly decomposable systems described by Herbert Simon: subsystems that solve local problems and integrate solutions through interfaces. A market is a multi-agent system. A swarm of drones is a multi-agent system. An immune response is a multi-agent system. The common structure is local decision-making with global consequences.

The critical challenge is alignment: ensuring that the emergent collective behavior serves the intended purpose. In decentralized systems, local optimization does not guarantee global optimality. Agents pursuing individual goals may produce collectively destructive outcomes — the tragedy of the commons, market bubbles, information cascades. Designing multi-agent systems that are both efficient and aligned requires understanding not just individual agent behavior but the coupling topology that determines how local decisions propagate.