How we built our multi-agent research system

Anthropic shares engineering insights from building Claude's Research system, a multi-agent architecture where a lead agent coordinates specialized subagents working in parallel. The system outperformed single-agent approaches by 90.2% on research tasks by distributing work across agents with separate context windows. Key lessons include careful prompt engineering for agent coordination, scaling effort to query complexity, effective tool design, and robust evaluation methods. Production challenges involve handling stateful agents, debugging non-deterministic behavior, and managing deployments without disrupting running agents. Multi-agent systems excel at parallelizable research tasks but consume significantly more tokens than single-agent systems.