Skip to slide
15 / 41

CH 02 · The Case For Multi-Agent: Anthropic's Research System · 08/08

Read as one page

Key takeaways

  • Anthropic's research system uses an orchestrator-worker design: a lead agent plans and delegates, parallel subagents explore in isolated context windows, and a citation agent verifies sources.
  • It beat a single-agent baseline by about 90 percent, mostly because it spends far more tokens across parallel context windows.
  • Prompt engineering is the main control lever: teach delegation, scale effort to complexity, design tools carefully, start broad then narrow, and guide the thinking process.
  • Evaluate by outcomes, not process: start with small test sets, use a single strong LLM judge, and keep humans in the loop for edge cases.
  • Production demands recovery from failures, tracing for non-deterministic debugging, and rainbow deployments; synchronous execution remains a bottleneck.
  • Reserve multi-agent for high-value, parallelizable, breadth-heavy tasks, not for interdependent work like coding.

Continue to Chapter 3 for the opposing view.

Finished this idea?

Your learning progress stays in this browser.