Coding agents are fast, and most of what they do is safe. The problem is the small share of commands that are not: a force push, a dropped table, an rm -rf in the wrong folder. One of those can undo a week of work in a second.

In Quiel, some commands always stop, even when an agent runs in full autonomy. The list of stop patterns is yours to extend, but it can never be emptied. When an agent hits one, the task pauses and waits.

The request goes to the person you named for that role. They get a notification in the app, in Slack or by email with the exact command, the task and the machine, and they approve or refuse it from there.

Trust profiles decide what each role may do without asking. A developer agent can run tests all day; the same agent touching production credentials has to ask first.

Every decision lands in the append-only log, including refusals. When someone asks who approved what, and when, the answer is one search away.