Summary
Multi-agent systems assign different roles to different LLM calls that coordinate. Sometimes it helps (research → draft → critique); often it does not. The coordination cost is real; the single-agent baseline is usually stronger than teams admit.
Objectives
- 01State the three canonical multi-agent patterns.
- 02Explain the coordination cost.
- 03Recognize when a single agent is sufficient.
The Lesson
Patterns
Hierarchical: manager agent delegates to workers. Peer: agents with distinct roles collaborate. Debate: agents argue positions; a judge decides. All three are studied; all three sometimes work; none is a universal answer.
The coordination cost
Every inter-agent message is a token cost, a latency cost, and a failure mode. If a single strong agent can hold the context and the roles, it usually will do so faster and cheaper. Measure against the single-agent baseline before shipping.
Key Ideas
- Multi-agent adds coordination cost; sometimes worth it, often not.
- Always measure against a single-agent baseline.