Reading a crowd of agents
Orchestrator and workers
The orchestrator and workers pattern is a topology where a lead agent plans, delegates to subagents that each hold their own context, and synthesizes their results, with the workers not communicating directly.
Control sits with the agent
When to use it
Tasks that decompose into genuinely independent directions, where parallel exploration beats sequential work.
What it prevents
The illegibility of a peer-to-peer swarm, where any agent may influence any other and no one can reconstruct what happened.
Research run1 lead, 3 workers
LeadSplit the question, holds the synthesis
Worker 1Filings. Own context, reports up only.
Worker 2Press. Own context, reports up only.
Worker 3Rate cards. Own context, reports up only.
No worker can influence another, which is what keeps the run reconstructable.
Who ships it
- Anthropic’s research system: a lead agent plus parallel subagents outperformed a single agent by 90.2 percent on internal research evaluations, at roughly fifteen times the token cost of a chat. Anthropic is equally clear about the limit: domains where agents must share context, or where the work has many interdependencies, are a poor fit
Sources
Related patterns: Delegation briefs · Mission control · Budget envelope