What is an agent handoff?
An agent handoff is the transfer of control over a conversation or task from one AI agent to another, usually a specialist, which then continues the work with some or all of the earlier context.
OpenAI's practical guide to building agents describes two ways to split work between agents. In the manager pattern, a central agent calls specialists as tools and stays in charge. In the decentralised pattern, agents hand off to each other, and whichever agent holds the conversation answers the user. A handoff is the move between them.
Google uses the term agent transfer for the same idea in its writing on multi-agent context, and contrasts it with agents as tools in the same way.
How a handoff works in practice
Exposed as a tool
In the OpenAI Agents SDK, each possible handoff appears to the model as a tool, such as transfer_to_refund_agent, so choosing to hand off is an ordinary tool call.
History moves with it
By default the receiving agent sees the full conversation. Input filters can trim it, for example by removing earlier tool calls.
A reason can travel too
The handing-off agent can pass structured input, such as a reason or priority, and a callback can start fetching data as the handoff happens.
One destination per handoff
Each handoff targets one agent. Several possible destinations means several handoff tools, and the model picks one.
Context and safety across a handoff
Google's post on context-aware multi-agent frameworks (December 2025) recommends scoped inclusion, passing only what the next agent needs, and narrative casting, which reframes the previous agent's messages as background so the new agent does not mistake them for its own actions.
Checks need to follow the conversation too. The Agents SDK documentation notes that input guardrails run only for the first agent and output guardrails only for the agent that produces the final answer, so a check that matters for every specialist has to be placed where every specialist passes through it.
Handoff or agent as a tool
A handoff fits when a specialist should own the rest of the conversation, such as a refunds agent that needs to ask its own follow-up questions. When the lead only needs an answer back, calling the specialist as a tool keeps control in one place; the Agents SDK docs prefer that route when a nested specialist needs structured input without taking over the conversation.
How Motiontic's engineers apply it
Motiontic's engineers keep the number of handoff paths small and test each one: does the right agent receive the conversation, with the right context, and can it hand back or escalate to a person when it is out of its depth. Handoffs to a human follow the same design, with the transcript and the reason for escalation attached.
A handoff compared with calling an agent as a tool
| Handoff | Agent as a tool | |
|---|---|---|
| Who keeps control | The receiving agent | The calling agent |
| Who talks to the user next | The receiving agent | The calling agent |
| Context passed | The conversation, filtered or whole | Only the input the caller sends |
| Fits | Specialists that own a whole branch of the conversation | Specialists that return one answer or artifact |
| Main risk | Lost context or checks that do not follow the conversation | An overloaded lead agent |
Common questions
What is a handoff in the OpenAI Agents SDK?
A handoff is a delegation from one agent to another, exposed to the model as a tool. When the model calls it, the run continues with the receiving agent, which by default sees the full conversation history.
What is the difference between a handoff and routing?
Routing usually happens once, up front, when a classifier sends an input to the right branch. A handoff can happen at any point in a conversation, decided by the agent that currently holds it.
Can an AI agent hand off to a human?
Yes, and it should be designed in. OpenAI's guide recommends planning human intervention for when failure thresholds are exceeded or an action is high-risk.
Related
Routing (agent pattern)
Classify first, then send each input to the branch built for it.
Orchestrator-worker pattern
A lead agent plans, workers run in parallel with their own context, the lead merges.
Human in the loop
Approval gates where a person decides, placed by risk rather than on every step.
Customer service agents
Support agents that look up, act and escalate inside Zendesk, Intercom and the systems behind them.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.