When the agent asks first

In Build, the agent does small things on its own and stops in front of the ones with consequences. It shows you a card with the action and what it would do, and waits.

You approve, or you do not. Nothing happens in the meantime.

What counts as consequential

Reading never asks. Looking up a customer, searching your documents, counting hours — that just happens.

Ordinary writing does not ask either: creating a record, filling in a field, ticking off a step in a process. These are the things you would have typed yourself, and the result is visible and correctable.

What asks is the rest: deleting things, sending something out of the house, running a command in a folder you shared. The rule of thumb is whether it would be awkward to undo.

What the card shows

The action, and the details it would use — which record, which recipient, which command. It is the same information the agent is about to act on, not a summary of it.

Saying no

Say no and the action does not run. The agent is told you refused and carries on with the rest of the job; it does not sulk and it does not try the same thing again through another door.

If you would rather stop the whole thing instead of one step, use the stop button while the agent is working.

Not being asked

In Auto the in-house steps run without a card — that is the point of the level, see Ask & Plan, Build, Auto. Actions that leave your system still ask, in every level.

Background jobs ask too

A job running in the background can hit the same gate. It stops and waits, and the card appears in the conversation that started it — so an unanswered question does not quietly hold up work you have forgotten about. See work that runs in the background.