A handoff is the act of one agent passing control of a task — along with the relevant context — to another agent better equipped to continue it. Rather than finishing every part of a request itself, an agent can recognize that the work has shifted into another agent’s area of responsibility and transfer the conversation, so the second agent picks up from the current state instead of starting fresh.
Why handoffs matter
When a task spans several areas of expertise, forcing one agent to cover all of them produces a sprawling prompt and a context crowded with details relevant to only part of the work. The agent becomes harder to focus, harder to evaluate, and more likely to make mistakes outside its core competence. Splitting the work across specialists is the usual remedy, and handoffs are how control moves between those specialists once the work is underway.
The benefit is that each agent stays narrow and well-scoped, while the system as a whole can still address a request that crosses boundaries. The risk is in the transfer itself: if a handoff drops context, hands to the wrong agent, or bounces control back and forth without progress, the coordination overhead can outweigh the gain. A clean handoff carries enough state for the receiver to continue and makes the transfer visible so it can be traced.
How it works
A handoff generally proceeds in a few steps:
- Recognition — the active agent determines that the task has moved beyond its role or available tools.
- Selection — it identifies the target agent, commonly by invoking a dedicated handoff action that names the agent to receive control.
- Transfer — control and the relevant context pass to the target, which may be the full conversation history or a focused summary.
- Continuation — the receiving agent resumes the task from the handed-off state and either completes it or hands off again.
Frameworks expose this in different forms. A handoff strategy can let any agent in a group pass control to another, and a handoff operator between two agents can declare that the first may transfer to the second. In both cases the underlying action is the same transfer of control and context.
Handoff vs. routing
Routing and handoff both direct work to the right agent, but at different moments. Routing is a front-door decision: a classifier inspects the request once and sends it to the appropriate agent before any work happens. A handoff is a mid-task decision made by an agent that is already running, when it determines someone else should continue. A system often uses both — routing to choose the first agent, handoffs to move control as the task evolves.
In practice
Handoffs are reliable only when the transfer of control and context is recorded and survives interruption. A durable, observable runtime persists each agent’s steps and the handoffs between them server-side, so a crash during a transfer resumes rather than losing the run, and every change of control is visible in the trace. Handoffs are a coordination mechanism within a multi-agent system, they are one way to implement orchestration, and they contrast with the decentralized model of choreography. For the available patterns, see multi-agent strategies.