How we built our multi-agent research system
by Anthropic
Anthropic's production playbook for orchestrator-worker multi-agent systems that search and reason in parallel.
Overview
Published June 13, 2025 by Anthropic's engineering team (Jeremy Hadfield, Barry Zhang, and colleagues), this guide dissects the production multi-agent system powering Claude's Research feature. It uses an orchestrator-worker pattern in which a lead agent decomposes a query and spawns specialized subagents that explore in parallel, then synthesizes their findings and hands off to a citation agent. The article shares hard-won prompt-engineering principles (teach the lead to delegate, scale effort to query complexity, design careful tool interfaces, let agents use extended thinking to plan), an evaluation methodology combining LLM-as-judge rubrics (factual and citation accuracy, completeness, source quality, tool efficiency) with human testing, and candid production lessons on stateful error handling, debugging, and rainbow deployments. It quantifies the trade-offs: their Claude Opus 4 lead with Claude Sonnet 4 subagents outperformed a single-agent Opus 4 by 90.2% on internal research evals, while multi-agent systems burn roughly 15x the tokens of a chat, so the pattern only pays off on high-value, parallelizable tasks.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Guide
- Cost
- Free
- Duration
- ~25 min read
- Provider
- Anthropic
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Design an orchestrator-worker (lead + parallel subagents) architecture for open-ended research tasks
- ✓Prompt a lead agent to delegate effectively and scale effort to query complexity
- ✓Evaluate agents with LLM-as-judge rubrics plus human testing instead of rigid golden answers
- ✓Reason about the token economics of multi-agent systems (~15x a chat) and when they are worth it
- ✓Handle production concerns: stateful errors, debugging non-determinism, and deploying to running agents
Highlights
- •First-hand account from the team that shipped a real multi-agent product, not a toy demo
- •Concrete numbers: 90.2% eval uplift over single-agent, ~4x/15x token multipliers
- •Practical prompt-engineering and tool-design heuristics you can apply immediately
Who It’s For
Best For
- ✓AI engineers building multi-agent or deep-research systems
- ✓Teams deciding whether a task justifies multi-agent orchestration
- ✓Engineers designing agent evaluation and observability
Prerequisites
- •Familiarity with LLM agents, tool use, and prompting
- •Basic understanding of agent evaluation
FAQ
What is How we built our multi-agent research system?
An in-depth Anthropic engineering write-up on how they built the multi-agent system behind Claude's Research feature, for AI engineers designing production agents that plan, delegate, and search in parallel. It details the orchestrator-worker architecture, prompt engineering, evaluation, and the token economics of going multi-agent.
Is How we built our multi-agent research system free?
How we built our multi-agent research system is free to access.
What level is How we built our multi-agent research system for?
How we built our multi-agent research system is aimed at a advanced audience. Recommended background: Familiarity with LLM agents, tool use, and prompting, Basic understanding of agent evaluation.
How long does How we built our multi-agent research system take?
Expect roughly ~25 min read. Most learners work through it at their own pace.
What will I learn from How we built our multi-agent research system?
You'll learn: Design an orchestrator-worker (lead + parallel subagents) architecture for open-ended research tasks; Prompt a lead agent to delegate effectively and scale effort to query complexity; Evaluate agents with LLM-as-judge rubrics plus human testing instead of rigid golden answers; Reason about the token economics of multi-agent systems (~15x a chat) and when they are worth it; Handle production concerns: stateful errors, debugging non-determinism, and deploying to running agents.