An agent handoff should transfer a bounded task and its evidence, not merely a conversational summary. The receiving agent needs to know what is requested, what is already established, and which actions remain outside its authority.
Structured handoffs are contracts
A structured agent handoff is a typed message connecting one workflow stage to another. It should identify the task, current intent revision, expected output, relevant evidence, uncertainty, deadline, and permitted action scope.
For a proposed contract, keep task identity separate from message identity and execution attempt. Include a source-agent reference for diagnosis, but do not treat the sender’s self-description as authority. The trusted orchestrator controls permissions and ownership.
Keep completed business operations explicit. A receiving agent must not repeat a write merely because the handoff asks it to “finish the request.” Include authoritative result references and unresolved steps in separate fields.
Transfer a fictional scheduling task
An intake agent has confirmed the service address and caller preference. It sends a scheduling specialist the supported service, confirmed fields, current revision, and a request to identify available options. The specialist may return options; it is not authorised to book them yet.
A caller correction arrives while the specialist works. Its response should carry the input revision so the orchestrator can reject stale options. A polished reply based on the old address is still unusable for the current task.
| Handoff field | Purpose |
|---|---|
| Task and operation reference | Preserve the logical business identity |
| Intent revision | Detect superseded work |
| Expected result type | Bound the receiving task |
| Evidence references | Support claims without a transcript dump |
| Known uncertainty | Prevent unsupported certainty |
| Allowed scope | Limit effects at the execution boundary |
Replay a fixture with missing required fields, a stale revision, and a duplicate delivery. Assert that the receiving stage either proceeds within the contract or returns a safe unresolved result.
Prefer explicit acceptance over assumed ownership
The orchestrator should know whether the handoff was accepted, rejected, timed out, or completed. A message being delivered does not prove the specialist took responsibility. Keep one accountable owner for any consequential business action.
Structured contracts add schema and compatibility work. They also make failures testable and reduce reliance on agents reconstructing intent from long conversations. Keep the packet small enough to inspect and rich enough to preserve consequential meaning.
This contract is proposed. The voice integration case study describes related workflow experience without claiming this multi-agent architecture. Pair it with agent contract versioning.
Agent handoff FAQ
Should every prior message be copied?
No. Pass relevant facts and controlled evidence references under the receiving task’s access scope.
Does a handoff grant tool permission?
No. Enforce authority through trusted application policy.
What makes a handoff complete?
The intended owner accepts it, returns the expected result or explicit unresolved state, and preserves the task’s business identity.