Skip to main content
Five agents receive the identical prompt with no coordination. Their outputs diverge because LLMs are nondeterministic. A curator agent then reads all five candidates and selects the strongest one. Run it twice — you’ll get different winners.

What it demonstrates

  • swrm fan-out: all five agents run concurrently
  • on_failure: continue — the swrm proceeds even if an individual agent fails
  • Synthesis referencing each agent’s output via {{ generate.agents.<id>.output }}
  • Output variance as a feature, not a bug
  • Mixing OpenAI and Anthropic within a single swrm node

Run it

Workflow

docs/cookbook/1000-monkeys/workflow.yaml

Graph

Next steps

Market Analysis

Parallel specialists, each with a different task.

Swrm Reference

Full swrm node documentation.