When your agent can't read the result back, what counts as proof?
A user asks what counts as proof of completion when an agent cannot read the result back itself. The thread gets at the line between a demo and a production agent: a tool's success message is not evidence. The practical answer is that the result itself must carry something independently checkable — a receipt, an ID, or a confirmation number — or the action should be treated as not done.