How we built our multi-agent research system

Author: Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford (Anthropic) · Type: practitioner · Status: draft · Published: 2025-06-13 · URL: https://www.anthropic.com/engineering/built-multi-agent-research-system

Licence: Copyright Anthropic; free to read, quote with attribution.

Summary

Anthropic’s engineering write-up on the orchestrator-worker architecture behind Claude Research: a lead agent plans and delegates to parallel subagents. Documents a 90.2% improvement over single-agent Opus 4 on internal research evals, but roughly 15x the token cost of a chat interaction, plus hard-won lessons on prompt design, error recovery, and why human testers remain essential.

Insights

• The orchestrator-worker pattern (lead agent plans, parallel subagents execute) is the production-proven baseline for agent teams, and its main mechanism of value is simply spending more tokens on the problem. (high, draft) orchestration cost agents in production

  Multi-agent systems work mainly because they help spend enough tokens to solve the problem. — Section on why multi-agent systems work / performance evaluation

• Multi-agent orchestration carries roughly 15x the token cost of a single chat interaction, so it only pays off on high-value, parallelizable tasks. (high, draft) cost orchestration

  Agents use approximately 4x more tokens than chat interactions… multi-agent systems consume roughly 15x more tokens than standard chats. — Performance and cost discussion

• Even a sophisticated eval suite misses failure modes that human review catches, which argues for keeping human spot checks in any agent pipeline. (high, draft) human in the loop evaluation content quality

  People testing agents find edge cases that evals miss. These include hallucinated answers on unusual queries, system failures, or subtle source selection biases. — Evaluation section

• Agentic systems are fragile to small changes and need durable-state error recovery, not restart-from-scratch behavior. (high, draft) agents in production reliability orchestration

  In agentic systems, minor changes cascade into large behavioral changes, which makes it remarkably difficult to write code for complex agents. … We built systems that can resume from where the agent was when the errors occurred. — Production reliability and engineering challenges section