glossary
Multi-agent System
A distributed workflow of agents with roles, message contracts and explicit coordination.
A multi-agent system is a distributed workflow in which several agents — each with a narrow responsibility, typed message exchange and a constrained tool set — are coordinated by an orchestrator.
How it differs from the naive version: “a few prompts in a loop” do not form a system — they have no contracts, behavior can’t be reproduced twice, and a part can’t be tested in isolation. In a multi-agent system, inter-agent contracts are pinned: an agent can be run in isolation with a known input, replaced, or debugged without breaking the rest.
Why it matters: decomposition is justified when it buys risk isolation and testability, not on its own. Often two or three roles with strict contracts are enough where an “agent swarm” seemed necessary; extra agents add latency and coordination cost without adding reliability.