Accountable handoffs
From clipboard relay to accountable continuation
Copying an agent answer into another chat is fast. It also strips away the facts that matter later: whose request this was, which revision was approved, what stayed forbidden, and why the second worker believed it could act.
The clipboard version
A reviewer says, “The patch looks fine if you only update the canary.” Someone copies that sentence into a second agent and adds, “Go ahead.” The new context has no durable link to the original request, reviewed diff, approver identity, expiry, denied production scope, or evidence. When a final message says “done,” the team must reconstruct causation from chat history.
The accountable version
- A task packet identifies the exact request, source context, bounded objective, approved scope, and denied authority.
- A human decision binds one proposed action and one request revision to that packet, including expiry and evidence.
- A consumption receipt records that one named successor won the local claim under a declared admission rule.
- A continuation context carries the decision, claim, scope, and start evidence into that successor without widening them.
- An agent report ties returned evidence to the same packet and continuation, states the boundary, and asks for the next human decision.
The second worker still needs a transport and a runtime. HACP does not launch it. The records make the reason for continuation inspectable and give the receiver enough information to fail closed when a binding is missing or stale.
What changes for the reviewer
| Clipboard relay | Accountable continuation |
|---|---|
| Identity inferred from chat | Actor, request, action, packet, and successor identities are explicit. |
| Scope restated in prose | Approved and denied scope remain bound to exact records. |
| “Go ahead” carries ambiguous authority | A human decision records the exact authority-bearing act. |
| Retries look like new work | A consumption record identifies the admitted successor claim. |
| “Done” ends the story | A report returns evidence for another human decision. |
A public example
The HACP success fixtures let you follow the same identifiers and digests from a task packet through a human decision, receipt, continuation context, and report. A stale-context mutation is separately required to fail.
What this does not prove
- HACP does not autonomously route, launch, or authorize agents.
- A valid continuation context does not prove the receiver understood it.
- A receipt does not prove provider execution or exactly-once external effects.
- The public fixtures use synthetic records and do not rely on private implementation or customer details.
- Later evidence cannot retroactively widen the earlier human act.