What is Handoff?
Also called agent handoff, escalation, transfer.
A handoff is the transfer of an in-progress task or conversation from one agent to another, or from an agent to a person, along with the context needed to continue. Unlike delegation, responsibility moves and typically does not return. Handoffs are common in support routing, escalation, and specialist agent designs.
The defining question is what travels with the task. Options range from the full conversation history, which preserves everything but imports irrelevant material and confusion, to a written summary, which is compact but loses whatever the summary omitted. Many systems send a structured briefing: the objective, what has been established, what was attempted, and the open question.
Handoffs to people have extra requirements. The person needs to know why the transfer happened, what the agent already told the customer, and what remains unresolved, all without reading the whole transcript. A handoff that dumps a raw log and stops is why escalation often feels like starting over. Clear triggers, such as an expressed complaint or a policy boundary, should decide when it fires.
Handoff loops are a known failure. Two agents can each decide the request belongs to the other, passing it back and forth until a limit stops it. Context can also silently degrade across successive transfers, each summary dropping a little more, until the final holder is working from a description that no longer matches the original request.
The distinction from delegation matters for accountability. In a delegation the original agent still owns the outcome and checks the result. In a handoff ownership moves, so the record must make clear who holds it now. Related concepts are the supervisor pattern, where a coordinator routes instead of transferring, and human review, where a person receives the handoff.
Key points
- Responsibility transfers and usually does not come back.
- A structured briefing beats both a full transcript and a bare summary.
- Loops happen when two agents each reject the request.
- Escalation needs the reason and the open question stated clearly.
In practice
A billing question reaches a general support agent that cannot access invoices. It hands off to a billing agent with a short briefing: the customer's account, the invoice number in question, the two things already checked, and the unresolved discrepancy. The customer sees one continuous conversation. The second agent starts from the briefing instead of re-reading forty messages.
Handoff FAQs
How does an AI agent handoff work?
A handoff transfers an in-progress task or conversation to another agent or person with the context needed to continue. The receiving party becomes responsible for the next step and should know what has happened, what was tried, and what remains open.
What information should an AI agent handoff include?
A useful handoff includes the objective, relevant facts, actions already taken, the reason for transfer, the open question, and any constraints. A structured briefing lets the receiver continue without rereading an entire transcript.
What is the difference between an AI agent handoff and delegation?
A delegation sends a bounded subtask out and returns the result to the original agent, which keeps accountability. A handoff moves responsibility for the task or conversation to the receiving agent or person.
How can AI systems avoid handoff loops?
Set clear ownership, transfer criteria, and a maximum number of hops. The record should show who owns the task now, so two agents cannot keep sending it back to each other without a decision or escalation.