Type: url
Original: https://www.anthropic.com/engineering/multi-agent-research-system
Status: generated
Summary
Anthropic’s engineering account of its Research feature, built as an orchestrator-worker multi-agent system in which a lead agent plans, spawns subagents that search in parallel with their own context windows, and synthesises what they return. It reports a 90.2% improvement over a single-agent baseline on an internal research eval, attributes most of the measured performance variance on BrowseComp to token usage, and states the cost plainly at roughly 15x the tokens of a chat — while naming domains it considers a poor fit, most coding tasks among them. The rest covers the prompt-engineering principles that made delegation work, how to evaluate a system that never takes the same path twice, and the production problems that followed: compounding errors in stateful agents, non-deterministic debugging, rainbow deployments, and a synchronous-execution bottleneck.