Multi-Agent System

Common shapes

The most common is orchestrator-workers: a lead agent splits the task, dispatches to specialised sub-agents, then merges their results. Other shapes include pipelines, debate/voting, and decentralised peer collaboration.

Each sub-agent typically has its own system prompt, tool set, and context window, and they exchange only structured intermediate results rather than sharing full history.

Benefits and costs

Benefits: break a task larger than one context window into parts, parallelise distinct roles, isolate failures. Costs: coordination overhead, token cost that multiplies, and the fact that *reassembling* sub-results is itself an error-prone step. Anthropic's finding is that multi-agent pays off mainly when a task decomposes cleanly into parallel work.

Common misconceptions

• More agents is not automatically better; on simple tasks it is usually slower and more expensive.
• Sub-agents do not share memory by default — how information passes between them must be designed explicitly.

Related terms

Sources

1. Wooldridge, M. — An Introduction to MultiAgent Systems (2nd ed., 2009)
2. Anthropic — How we built our multi-agent research system (2025)
3. Anthropic — Building effective agents: orchestrator-workers (2024)