Context Threading

Context threading supports parallel processing patterns where multiple agents work on different aspects of a problem simultaneously within the same session.

How It Works

The system creates session "threads" that inherit context from parent sessions while maintaining independent execution paths. Thread results are merged back into the parent session on completion.

Parent Session (user_123)
    |
    |-- Thread A: Technical Analysis
    |   (inherits context, independent execution)
    |
    |-- Thread B: Financial Analysis
    |   (inherits context, independent execution)
    |
    +-- Thread C: Risk Assessment
        (inherits context, independent execution)
        |
        v
    Results merged back to parent session

Use Cases

Parallel analysis: Multiple specialist agents analyze the same input simultaneously
Fan-out / fan-in: One agent distributes work, waits for parallel results, then synthesizes
Independent subtasks: Complex requests decomposed into independent pieces

Implementing with YAML Configuration

Context threading is configured through can_call rules. When an agent calls multiple downstream agents, they execute with shared session context:

agents:
  coordinator:
    name: "coordinator"
    model: "gpt-4o"
    can_call:
      - agent: "technical_analyst"
        condition: "always"
      - agent: "financial_analyst"
        condition: "always"
      - agent: "risk_assessor"
        condition: "always"

  technical_analyst:
    name: "technical_analyst"
    model: "gpt-4o-mini"
    yields:
      technical_score: "number"
      findings: "array"

  financial_analyst:
    name: "financial_analyst"
    model: "gpt-4o-mini"
    yields:
      financial_score: "number"
      projections: "array"

  risk_assessor:
    name: "risk_assessor"
    model: "gpt-4o-mini"
    yields:
      risk_level: "string"
      factors: "array"

The coordinator's intelligence function can synthesize parallel results:

async def coordinator(llm_response, input_data, context):
    # Results from parallel agents arrive via the mesh
    return {
        "summary": llm_response.get("summary", ""),
        "recommendation": llm_response.get("recommendation", "")
    }

Next Steps

State Management — State patterns
Session Management — Session lifecycle
Multi-Agent Coordination — Coordination patterns

LeafMesh — Parallel processing with shared context