Reference · Glossary

Multi-agent system

Last updated

A **multi-agent system** coordinates several LLM agents (researcher, critic, coder) that pass messages or share tools to complete a job.

#When to use

Complex workflows with distinct roles and checkpoints — research then critique, plan then execute — when one prompt is too overloaded.

#When not to

Simple Q&A. Extra agents add latency, cost, and failure modes; start with one agent plus tools.

#Quality checklist

  • Clear role contracts and stop conditions
  • Shared scratchpad / state with size limits
  • Eval the whole graph, not only one node
  • Human approval for irreversible actions

#Learn next