Back to Reviews

How we built our multi-agent research system

Visit Source
Research Paper•Anthropic•Jun 13, 2025
LLMAgent

Anthropic Official Prompt

research_lead_agent.md

https://github.com/anthropics/claude-cookbooks/blob/main/patterns/agents/prompts/research_lead_agent.md

Pros of a multi‑agent research system

  • Fits open‑ended problems where it is very difficult to predict the required steps in advance.
  • Excels at breadth‑first queries that involve pursuing multiple independent directions simultaneously.

    • Cons and limitations

  • Burns through tokens fast.
    • Agents typically use about 4× more tokens than standard chat interactions.
    • Multi‑agent systems often use about 15× more tokens than chats.
  • Not a good fit for domains where:
    • All agents must share the same full context, or
    • There are many tight dependencies between agents.
    • These systems excel at heavily parallelizable tasks where agents do not need to share much context.

    • Architecture overview for research

    Principles for prompt engineering

  • Teach the orchestrator how to delegate
    • Each sub‑agent needs: an objective, an output format, guidance on tools and sources to use, and clear task boundaries.
    • Without detailed task descriptions, agents duplicate work, leave gaps, or fail to find necessary information.
  • Embed scaling rules in the prompts
    • Simple fact‑finding may need just 1 agent with 3-10 tool calls.
    • Direct comparisons might use 2-4 sub‑agents with 10-15 calls each.
    • Complex research can involve 10+ sub‑agents with clearly divided responsibilities.
  • Tool design and selection are critical
    • Give agents explicit heuristics: for example, examine all available tools first, match tool usage to user intent, search the web for broad external exploration, or prefer specialized tools over generic ones.
  • Let agents improve themselves
  • Start wide, then narrow down

    •