Workflow · Production

Anthropic's multi-agent research system: lessons from prototype to production

The problem

Research tasks are inherently open-ended and path-dependent, making them impossible to solve with fixed, linear pipelines; the unpredictability requires AI agents that can dynamically adapt their approach based on intermediate findings.

First attempt

Early agents in the system made coordination errors such as spawning excessive subagents, searching endlessly for nonexistent sources, and showing consistent bias toward SEO-optimized content farms over authoritative sources.

Workflow diagram · grounded in source
1
User submits research query
trigger
“When a user submits a query, the lead agent analyzes it, develops a strategy”
2
Lead agent plans strategy
ai_action
“the lead agent analyzes it, develops a strategy”
3
Spawn parallel subagents
routing
“spawns subagents to explore different aspects simultaneously”
4
Subagents search and filter
ai_action
“the subagents act as intelligent filters by iteratively using search tools to gather information”
5
Subagents return condensed results
output
“returning a list of companies to the lead agent so it can compile a final answer”
6
Lead agent compiles final answer
output
“the lead agent so it can compile a final answer”
Reported outcome

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

Reported metrics
Multi-agent vs single-agent performance improvement90.2%
Agent token usage vs chat interactionsabout 4×
Multi-agent token usage vs chatabout 15×
token usage explains BrowseComp performance variance80%
Show all 6 reported metrics
multi-agent vs single-agent performance improvement90.2%
agent token usage vs chat interactionsabout 4×
multi-agent token usage vs chatabout 15×
token usage explains BrowseComp performance variance80%
three factors explain BrowseComp performance variance95%
user time savedup to days of work
Reported stack
Claude Opus 4Claude Sonnet 4Google Workspace
Source
https://www.anthropic.com/engineering/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 agent and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on internal research evaluations, and users report saving up to days of work on co…

What tools did this team use?

Claude Opus 4, Claude Sonnet 4, Google Workspace.

What results were reported?

Multi-agent vs single-agent performance improvement: 90.2%; Agent token usage vs chat interactions: about 4×; Multi-agent token usage vs chat: about 15×; token usage explains BrowseComp performance variance: 80% (source-reported, not independently verified).

What failed first in this deployment?

Early agents in the system made coordination errors such as spawning excessive subagents, searching endlessly for nonexistent sources, and showing consistent bias toward SEO-optimized content farms over authoritative…

How is this workflow AI workflow structured?

User submits research query → Lead agent plans strategy → Spawn parallel subagents → Subagents search and filter → Subagents return condensed results → Lead agent compiles final answer.