Multi-Agent System — AI Glossary

Reference details

Related terms

• AI Agent
• Agent Handoff
• Agentic Loop
• Orchestrator Agent

How it works

1. The topology is the design decision, and the three common shapes fail differently rather than merely performing differently.
2. In an orchestrator-workers shape one agent holds the plan and dispatches scoped tasks, so the failure mode is a bottleneck and a single point whose context must hold the whole problem.
3. In a peer-to-peer shape agents message each other directly, which removes the bottleneck and introduces conversations that do not terminate — two agents can defer to each other indefinitely.
4. In a hierarchical shape the orchestrator pattern nests, which scales the decomposition and makes attribution hardest, since a bad result arrives through several summarising layers.
5. Every shape needs a termination condition at the system level, because each agent only knows when its own task is complete and none of them can observe that the system is looping.
6. Routing is either a fixed topology or a model decision, and only the first is inspectable before a run — a model-chosen next agent makes the execution path as unpredictable as a single agent's tool selection.

Example: Where the errors go

A planner, a researcher and a critic each at 90% reliability do not give you 90% end to end. Errors compound along the chain: 0.9 × 0.9 × 0.9 ≈ 0.73. Add a step and it falls again. Multi-agent designs buy specialisation and pay for it in compounding failure and multiplied token cost.

1 agent  @ 0.90            = 0.90
3 agents @ 0.90 in series = 0.90^3 = 0.729
5 agents @ 0.90 in series = 0.90^5 = 0.590

Commonly confused with

Multiple prompts is not multiple agents. Calling a model three times with different instructions is a pipeline. A multi-agent system has agents that decide, delegate and react to each other's output — which is also what makes it hard to debug.

When to use it

Reach for it when:

• Sub-tasks genuinely need different tools, context or permissions
• One agent's context window cannot hold the whole problem
• A separate critic or verifier measurably improves output quality

Reach for something else when:

• A single agent with good tools would do — this is the common over-build
• You cannot trace which agent caused a bad outcome
• Cost matters: every agent adds a full model call, and they multiply

Referenced by

This term is used by the following ModelRefs references:

• supervisor worker
• debate
• chain of agents

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Multi-Agent System — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path