What is an agent handoff?
Updated November 2026
An agent handoff transfers active responsibility for a task or conversation from one agent to another. It changes who handles the next step, unlike a subagent call where the caller normally retains control and consumes the result.
What changes during a handoff?
LangChain’s multi-agent documentation describes handoffs through state updates that change routing or agent configuration. Microsoft’s architecture guide describes dynamic delegation between specialists. A handoff therefore needs an explicit control transition, not merely a message mentioning another agent.
For a support conversation, a triage agent might transfer a billing issue to a billing specialist. Record the receiving role, reason for transfer, current objective, unresolved questions, and any restrictions on actions. The destination should know whether it owns the whole conversation or only a bounded phase.
Which context should move with responsibility?
Transfer the minimum sufficient task state rather than copying every message. Include relevant evidence and user commitments, but avoid passing unrelated personal data or stale working notes. Distinguish verified facts from the outgoing agent’s tentative interpretation.
| Handoff field | Purpose |
|---|---|
| Destination and scope | Identify the new owner and assignment |
| Evidence references | Preserve support for important claims |
| Completion or return rule | Explain when ownership changes again |
How do you prevent loops and lost work?
Two specialists can repeatedly transfer a difficult case to each other. Add a transfer budget, detect repeated destinations, and provide an escalation destination when routing cannot resolve the task. Logging the reason for each transfer helps distinguish legitimate progression from circular delegation.
The main tradeoff is specialization versus continuity. Narrow specialists can improve local handling while losing context across boundaries. Summaries reduce token use but may omit qualifications that matter to the next decision.
For distributed execution, define when ownership becomes effective and what happens if the recipient does not acknowledge the assignment. A failed transfer should leave an identifiable owner, not two agents assuming the other is responsible. Preserve unfinished action status separately from conversational summaries.
What else should you know?
Is every delegation a handoff?
No. A tool-like subagent call can leave the caller in control.
Must the whole transcript transfer?
No. Relevant structured context is often safer and smaller.
Can a handoff return to the sender?
Yes, if the workflow defines a return transition.
What is the main safety check?
Keep one clear owner for each consequential action and pending decision.