Discussion on: AI Agents Can Get Approval and Still Do the Wrong Thing

[Atul Kumar]

That's a really interesting extension of the same principle. I particularly like the idea of treating acceptance as a signed delivery receipt rather than a boolean approval. Binding the task ID, deliverable hash, verification method, and verifying party creates an auditable record of exactly what was accepted.

There's an important distinction here: authorisation should bind to the action being performed, while acceptance should bind to the artifact being delivered. They solve related but different problems.
If an agent replans and changes the deliverable, the previous acceptance should no longer apply. The hash check makes that relationship explicit instead of relying on an assumption that the task remained unchanged.
I'd be interested in reading the spec, especially how you handle cases where a deliverable changes after verification, or where acceptance depends on external tests and mutable resources.

Thanks for sharing this. It adds an important dimension to the discussion around trust and accountability in multi-agent systems.