Multi Agent System: Definition and Examples

A Multi Agent System (MAS) refers to a set of intelligent agents that interact within a shared environment. Each agent has its own capabilities, goals, and knowledge, and can act autonomously while coordinating with others. This approach is inspired by distributed systems and collective intelligence observed in nature.

In the context of generative AI and prompt engineering, a multi-agent system typically involves multiple instances of language models (LLMs) that assume specialized roles. For example, one agent may act as a researcher, another as a writer, and a third as a critic. Each agent receives a distinct system prompt that defines its expertise, responsibilities, and rules for communicating with other agents.

The main benefit of a MAS lies in decomposing complex problems. Rather than asking a single LLM to handle everything — research, analysis, writing, verification — these responsibilities are distributed among specialized agents. This leads to more reliable results, as each agent can focus on what it does best, while benefiting from the work of others.

Multi-agent architectures vary widely: some are hierarchical (an orchestrator agent delegates to others), others are collaborative (agents negotiate among peers), and still others are competitive (agents propose competing solutions that an arbiter evaluates). Frameworks like AutoGen, CrewAI, or the Claude Agent SDK now facilitate the implementation of these systems.