@ai-desci
Summary of Anthropic’s Claude Multi-Agent Research System
> Anthropic uses the “orchestrator-worker architecture”, where a lead Claude agent spawns and coordinates subagents to explore complex queries in parallel (a multi-agent system).
> This lifted Anthropic’s internal research task success rate by ~90 % over single-agent setups.
> The system scales reasoning capacity efficiently but costs ~15× more tokens and is reserved for high-value questions.
> It improved agent performance via tailored prompts and Claude-driven self-optimization, cutting task times by 40%.
> It uses LLM-as-judge, scoring with rubrics for factuality + human testing to catch failures👇
https://www.anthropic.com/engineering/built-multi-agent-research-system
0 reply
0 recast
0 reaction