MORAIDICTIONNAIRE IA
Agents

Multi-Agent Systems

Multiple autonomous AI agents that collaborate, split tasks and correct each other to solve a complex problem.

Picture a team of experts collaborating on a complex project: each has a specialty, they talk to one another, divide the work and correct each other. A multi-agent system (MAS) applies exactly this idea to AI: several autonomous agents — often powered by language models — interact within a shared environment to solve a problem none could handle alone.

What is an agent?

An agent is an entity able to perceive its environment, reason, then act through tools (web search, code, databases). It is often summarized by the loop:

$$ \text{Perception} \rightarrow \text{Reasoning} \rightarrow \text{Action} \rightarrow \text{Observation} $$

In a MAS, we multiply these loops and make them communicate. Each agent can take on a specific role: planner, executor, critic, verifier.

Coordination architectures

Architecture Principle Typical use
Orchestrator A lead agent delegates to sub-agents Research pipelines
Hierarchical Managers and cascading sub-teams Long, structured projects
Decentralized Agents negotiate as peers Markets, auctions, simulations
Debate Agents argue, then vote Answer reliability

Why it is powerful

Specialization lowers each agent's cognitive load, and cross-criticism limits errors: a verifier agent can catch another's hallucination. This is the principle behind modern frameworks (AutoGen, CrewAI, LangGraph).

Beware, though: adding agents also multiplies cost, latency and the risk of error propagation.

Where a lone agent thinks, a multi-agent system deliberates — cooperation becomes a form of intelligence in its own right.

Explore the full AI dictionary →