Skip to content
Sync AI AgentsTwo curved arrows chasing each other in a closed loop, each tipped with a filled circle.sync ai agents

Pattern 02: handoffs

Handoff relay chainAgents A, B and C in a row with solid arrows from A to B and B to C, each carrying a small note. A faded dashed arrow from C back to A is labelled ping-pong.ping-pongABC
Control moves forward. Context moves only as far as the note.

How do agent handoffs work?

An agent decides it is not the right one for the current step and hands control to another. The receiving agent gets the conversation, or a summary of it, and continues. There is no lead in the middle. The OpenAI Agents SDK makes handoffs a first-class primitive alongside agents-as-tools. Microsoft Agent Framework's Handoff orchestration lets "agents transfer control to each other". LangChain documents Handoffs next to Subagents and Router. Pydantic AI calls it programmatic hand-off.

Why do teams pick handoffs?

Handoffs map to how people already work: support triage, then billing, then escalation. Each agent stays small, with a short prompt and a few tools. There is no orchestrator to scale or pay for. And the user often talks to exactly one agent at a time, which keeps the experience coherent.

Where do handoffs break?

The note is the context. Whatever the first agent does not write down, the second agent does not know.

Actions carry implicit decisions, and conflicting decisions carry bad results.
Source: Cognition, Don't Build Multi-Agents, Jun 2025

A billing agent that does not know the triage agent already promised a refund will make a different decision.

Ping-pong. Two agents that each think the other should handle an edge case will pass it back and forth. Without a hop limit this burns tokens until a timeout.

No shared view of ownership. Mid-chain, nobody outside the process can easily answer "who has this now?" That matters once a human needs to intervene.

Handoffs usually stay inside one framework. Handing off from an OpenAI Agents SDK agent to a CrewAI crew means a protocol such as A2A, and the handoff note then crosses a boundary where formats differ.

When do agent handoffs work, and when do they break?

When it works

  • Lanes are clear (triage, specialist, escalation)
  • Two to four agents
  • One framework
  • Conversation-style tasks
  • The next step depends on the previous result

When it breaks

  • Decisions made early are not written down
  • Two agents can both claim a case
  • Humans need to see who owns the work
  • Handoffs cross frameworks or teams

Which tools use agent handoffs?

ToolHow it coordinatesGood forSource
OpenAI Agents SDKHandoffs as a primitive; tracing built inPython/TS appsOpenAI docs
LangGraphHandoffs pattern on a state graphStateful flows needing durable executionLangChain docs
Microsoft Agent FrameworkHandoff orchestration with human-in-the-loop.NET and PythonMicrosoft Learn
Pydantic AIProgrammatic hand-off, typedTyped Python appsPydantic docs
BANDHandoffs via @mention in a shared room; per-agent, per-message lifecycle with attempt historyHandoffs across frameworks and to humansBAND platform

What do people ask about agent handoffs?

An agent handoff is when one AI agent transfers control of a task or conversation to another agent that is better suited to the next step.

Pass a structured note (intent, decisions and reasons, open questions, next owner), include the full trace where possible, and use a shared store for facts. See our handoff guide.

Handoffs are simpler when lanes are clear and one agent works at a time. An orchestrator is better when a task needs parallel subtasks and one place to decide when it is done.

Not sure which pattern fits?

Six questions, about a minute. You'll get a pattern, a shortlist and the tradeoffs.

Get started with BANDTake the finder

Last reviewed: September 2026