One claim in the brief is honest but still untraced. The only source that would fix it is live Q3 supplier pricing, and it sits in three vendor systems.
Fetching this by hand is an hour of clicking through three vendor portals. An agent does it in seconds, which is exactly why it needs a scope you granted, a stop that always works, and a record of everything it touched.
Autonomy is a grant, not a default. Turn a capability off and the matching step leaves the plan, so the scope and the plan can never disagree about what is going to happen.
A dry run says what it would do and touches nothing. A real run does the same steps for real, and stops the moment you say so.
Not a setting, not buried in a menu, and not only available before it starts. One click, at any point, and it is already stopping.
Written as it happens, not reconstructed afterwards. This is the record you would hand an auditor.
The brief is finished and waiting for your signature. The afternoon produced four things worth remembering, and the assistant would like to keep them.
Go to Recall →Press play to watch this run start to finish.
So the question is never “do you trust it”. It is what it may touch, what it does when it hits something it should not decide alone, and whether stop actually stops.
About 70 seconds, narrated. Synthetic data. Pause or take over at any point.