How we built our multi-agent research system
by Anthropic
Anthropic's own postmortem on shipping Claude Research: the architecture, the prompts, the evals and the 15x token bill.
Overview
Published 13 June 2025 by Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox and Daniel Ford, this is Anthropic's engineering account of the system behind Claude's Research feature, organized into six sections plus an appendix: benefits of a multi-agent system, architecture overview, prompt engineering and evaluations, effective evaluation of agents, production reliability and engineering challenges, and a conclusion. The architecture is an orchestrator-worker pattern: a lead agent analyzes the query and develops a strategy, spawns subagents that explore different aspects in parallel with their own context windows, then compiles their findings. The numbers are the reason to read it. A Claude Opus 4 lead with Claude Sonnet 4 subagents outperformed single-agent Opus 4 by 90.2% on Anthropic's internal research eval; token usage alone explains 80% of performance variance on BrowseComp; agents use roughly 4x the tokens of a chat and multi-agent systems roughly 15x, which is the honest constraint on when the pattern pays. Eight prompt-engineering principles are given concretely, including teaching the orchestrator to delegate with explicit objectives and output formats, embedding effort-scaling rules so a simple query does not spawn fifty subagents, designing tools with distinct purposes, and searching broad before narrowing. The evaluation section argues for starting at roughly 20 test queries with an LLM-as-judge rubric covering factual accuracy, citation precision, completeness, source quality and tool efficiency, kept alongside human review, which is what caught the agents favoring SEO content farms over authoritative sources. The production section covers durable state across long runs, compounding errors, full tracing, and rainbow deployments that avoid disrupting agents mid-run.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Guide
- Cost
- Free
- Duration
- ~25 min read
- Provider
- Anthropic
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Structure an orchestrator-worker agent system with parallel subagents and separate context windows
- ✓Judge when multi-agent parallelism is worth roughly fifteen times the token cost
- ✓Write orchestrator prompts that delegate with explicit objectives, boundaries and output formats
- ✓Embed effort-scaling rules so simple queries do not spawn dozens of subagents
- ✓Build an agent eval from about twenty queries with an LLM-as-judge rubric
- ✓Keep human review in the loop to catch source-quality biases that automation misses
- ✓Handle production realities: durable state, compounding errors, tracing and rainbow deployments
Highlights
- •First-party numbers rather than vendor marketing: 90.2% improvement over single-agent, 80% of variance explained by token spend, a ~15x token multiplier
- •States plainly where the pattern fails, including tasks needing shared context such as most coding work, instead of selling multi-agent universally
- •The eight prompt principles are drawn from observed failures, including early agents spawning 50 subagents for a trivial query
- •Simon Willison, previously a multi-agent skeptic, credits this post with changing his mind, and his notes are a useful independent read alongside it
- •Carries an appendix beyond the six main sections, so there is implementation detail past what secondhand summaries reproduce
Who It’s For
Best For
- ✓Engineers designing agent orchestration beyond a single ReAct loop
- ✓Teams costing out an agentic feature before committing engineering to it
- ✓Anyone building evals for agents where the output has no single correct answer
- ✓Platform engineers responsible for keeping long-running agents alive in production
Prerequisites
- •Familiarity with LLM tool use and a basic agent loop
- •Comfort reasoning about token economics and context windows
- •None beyond that: it is prose, with no code to run
FAQ
What is How we built our multi-agent research system?
Anthropic's engineering write-up of the multi-agent system behind Claude's Research feature, published June 2025 by the team that built it. It covers the orchestrator-worker architecture, eight prompt-engineering principles for lead agents and subagents, how to evaluate agents with small samples and LLM judges, and the production failures around state, errors and deployment that only appear at scale.
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 intermediate audience. Recommended background: Familiarity with LLM tool use and a basic agent loop, Comfort reasoning about token economics and context windows, None beyond that: it is prose, with no code to run.
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: Structure an orchestrator-worker agent system with parallel subagents and separate context windows; Judge when multi-agent parallelism is worth roughly fifteen times the token cost; Write orchestrator prompts that delegate with explicit objectives, boundaries and output formats; Embed effort-scaling rules so simple queries do not spawn dozens of subagents; Build an agent eval from about twenty queries with an LLM-as-judge rubric; Keep human review in the loop to catch source-quality biases that automation misses; Handle production realities: durable state, compounding errors, tracing and rainbow deployments.
Topics
Sources
This page was written from 2 sources, 1 on domains other than anthropic.com.