Handoff Pattern

Overview

The Handoff Pattern creates clean boundaries between agent responsibilities. When one agent completes its part of a task, it explicitly transfers control—along with all relevant context—to the next agent.

Core Mechanics

Handoff Message

{
  "from_agent": "research_agent",
  "to_agent": "synthesis_agent",
  "task_id": "task_123",
  "context": {
    "findings": [...],
    "sources_consulted": [...],
    "open_questions": [...]
  },
  "instructions": "Synthesize these findings into a summary"
}

Handoff vs. Agent-as-Tool

Aspect	Handoff	Agent-as-Tool
Control	Full transfer	Temporary delegation
Return	May not return	Always returns to caller
Context	Explicit transfer	Passed as parameters
Ownership	New agent owns task	Original agent retains ownership

Implementation

OpenAI Agents SDK

from openai_agents import Agent, handoff

research_agent = Agent(name="researcher")
writer_agent = Agent(name="writer")

@research_agent.on_complete
async def research_complete(result):
    await handoff(
        to=writer_agent,
        context=result,
        message="Write an article based on this research"
    )

LangGraph

Uses Command(goto="next_agent") to transfer control between nodes in the state graph.

Context Transfer Strategies

Full Context

Pass entire conversation history. Safe but can hit token limits.

Summarized Context

Compress history into summary. Efficient but may lose details.

Structured Context

Extract key fields into schema. Explicit but requires schema design.

Reference Context

Pass IDs; receiving agent fetches what it needs. Efficient for large context.

Common Pitfalls

Context Drift

Each handoff risks losing important details. Mitigate with:

• Explicit context schemas
• Validation at handoff boundaries
• Summary verification

Infinite Loops

Agent A hands to B, B hands back to A. Prevent with:

• Maximum handoff depth limits
• Cycle detection
• Explicit termination conditions

Orphaned Tasks

Handoff fails and no agent owns the task. Require:

• Acknowledgment from receiving agent
• Timeout and escalation paths
• Dead letter queues

When to Use

Good fit:

• Clear sequential workflows
• Specialized agents with distinct responsibilities
• Audit trail requirements

Poor fit:

• Highly dynamic, unpredictable workflows
• Real-time collaborative tasks
• Simple single-agent scenarios