Multi-Agent Patterns

Sub-Agents (agent_ref)

Delegate work to another agent by reference. The parent flow pauses while the sub-agent executes, then resumes with the result.

{
  "type": "agent_ref",
  "data": {
    "type": "agent_ref",
    "agentRef": "research-assistant",
    "refType": "slug",
    "inputMapping": { "message": "{{input.query}}" },
    "outputVariable": "research_result",
    "inheritContext": true,
    "timeout": 60000
  }
}

Input Mapping

Map parent variables to sub-agent inputs using {{variable}} interpolation.

Context Inheritance

When inheritContext is true, the sub-agent receives the parent’s user context (userId, session, etc.).

Agent Pools

Run multiple agents and combine their outputs using a strategy.

Strategies

Strategy	Behavior
best_of	Run all agents, use an evaluator LLM to pick the best response
aggregate	Concatenate all responses
vote	Return the most common response
fastest	Return the first response that completes

{
  "type": "agent_pool",
  "data": {
    "type": "agent_pool",
    "agents": [
      { "ref": "writer-agent", "refType": "slug" },
      { "ref": "editor-agent", "refType": "slug" }
    ],
    "strategy": "best_of",
    "evaluationPrompt": "Pick the response with the best writing quality.",
    "evaluatorModelId": "openai/gpt-4o",
    "maxConcurrency": 3,
    "agentTimeout": 30000
  }
}

Map-Reduce

Process an array of items in parallel, then combine results.

1. Map: For each item in the input array, run a sub-agent
2. Reduce: Combine all results using a strategy

{
  "type": "map_reduce",
  "data": {
    "type": "map_reduce",
    "inputVariable": "articles",
    "mapAgentRef": "summarizer",
    "mapAgentRefType": "slug",
    "itemTemplate": "Summarize this article: {{item.title}}\n\n{{item.content}}",
    "reduceStrategy": "concatenate",
    "maxConcurrency": 5
  }
}

Reduce Strategies

Strategy	Behavior
concatenate	Join all results with newlines
merge	Deep merge JSON results
custom	Use reduceExpression to combine