How to check everything an AI agent did on my behalf?
Check one action at a time: what was requested, what permission covered it, which material it used, what actually happened and what was charged. A conversation transcript helps reconstruct intent, but an execution record is needed to distinguish a proposed action from one that ran.
Find the smallest action record available, rather than stopping at the task summary. Compare its input and output with the instruction you gave. If a task produced three files and published one, those are separate events worth checking. A title saying “done” cannot tell you which file was sent, who could see it or whether a later attempt changed the result.
A useful record names the instruction, material and authorization used at the time. A broad label such as “your preferences” gives you little to disagree with. Where available, inspect the actual source or version. Missing information should stay missing: a plausible explanation reconstructed later is not the same evidence as a source captured when the action ran.
A proposal is not a completed action, and an accepted request is not always a completed result. Check whether failed, cancelled and unknown outcomes have distinct states. When a provider response was lost, the honest state may remain unknown until it is reconciled. Do not count every generated plan or notification as work delivered.
Compare task details with the usage or settlement record for that request. A retry may reuse a known result, repeat failed work or start a new action, depending on the path. Ask which outputs are preserved and which work will run again. Neither a retry button nor an idempotency field on its own proves that the whole workflow cannot duplicate effects.
A record maintained by the executing platform helps you review its activity. It does not by itself make the output correct or establish independent verification of every event. For important work, retain the original source and the final artifact, inspect the available export, and involve the responsible reviewer. Reversibility is action-specific; a message already read cannot simply be made unread.
OneOneTalk links supported task steps with their results and receipts. Review the actual step, its authorization, output and recorded cost; a completion label alone is not acceptance evidence. A receipt records platform activity and is not independent proof that every claim in the output is correct.
Approval applies to the action shown. The Web approval control can continue execution immediately, so review before confirming. When an interrupted request has an unknown outcome, reconcile its result and charges before retrying; cancellation does not automatically reverse an external effect or refund a charge.
More on the product in the English overview.
The public primary material this page is built on. We do not restate their conclusions as our own evidence — they are listed so you can check for yourself.