Back office ops · Production

Anthropic builds Claude's multi-agent Research feature: orchestrator-worker architecture outperforms single-agent by 90.2%

The problem

Research tasks are open-ended and path-dependent, making linear pipelines inadequate; information relevant to complex queries also exceeds single context windows.

First attempt

Early agents made coordination errors, including spawning excessive subagents for simple queries, searching endlessly for nonexistent sources, and consistently preferring SEO-optimized content farms over authoritative sources.

Workflow diagram · grounded in source
1
User query submission
trigger
“When a user submits a query, the lead agent analyzes it, develops a strategy, and spawns subagents to explore different aspects simultaneously”
2
Lead agent strategy planning
ai_action
“the lead agent analyzes it, develops a strategy”
3
Parallel subagent spawning
routing
“spawns subagents to explore different aspects simultaneously”
4
Subagent information gathering
ai_action
“the subagents act as intelligent filters by iteratively using search tools to gather information”
5
Final answer compilation
output
“returning a list of companies to the lead agent so it can compile a final answer”
Reported outcome

The multi-agent system with Claude Opus 4 as lead and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on the internal research evaluation, and users report saving up to days of work.

Reported metrics
Multi-agent vs single-agent performance on internal research eval90.2%
BrowseComp performance variance explained by token usage80%
BrowseComp performance variance explained by three factors combined95%
Agent token usage relative to chatabout 4×
Show all 6 reported metrics
multi-agent vs single-agent performance on internal research eval90.2%
BrowseComp performance variance explained by token usage80%
BrowseComp performance variance explained by three factors combined95%
agent token usage relative to chatabout 4×
multi-agent token usage relative to chatabout 15×
user time savedup to days of work
Reported stack
Claude Opus 4Claude Sonnet 4Google Workspace
Source
https://www.anthropic.com/engineering/built-multi-agent-research-system
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The multi-agent system with Claude Opus 4 as lead and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on the internal research evaluation, and users report saving up to days of work.

What tools did this team use?

Claude Opus 4, Claude Sonnet 4, Google Workspace.

What results were reported?

Multi-agent vs single-agent performance on internal research eval: 90.2%; BrowseComp performance variance explained by token usage: 80%; BrowseComp performance variance explained by three factors combined: 95%; Agent token usage relative to chat: about 4× (source-reported, not independently verified).

What failed first in this deployment?

Early agents made coordination errors, including spawning excessive subagents for simple queries, searching endlessly for nonexistent sources, and consistently preferring SEO-optimized content farms over authoritative…

How is this back office ops AI workflow structured?

User query submission → Lead agent strategy planning → Parallel subagent spawning → Subagent information gathering → Final answer compilation.