A general assistant cannot be reviewed

To review an action you need to know what the actor was supposed to do. A general assistant has no supposed-to, so every output has to be judged on its own merits with no reference point. That does not scale past a handful of actions a day.

Reviewers were spending most of their attention reconstructing intent rather than checking work.

A title implies a remit

Naming the role fixes intent up front. A Pipeline Watcher is supposed to watch the pipeline; anything else it does is out of scope by definition, and that is checkable without knowing anything about the specific run.

The title is a contract with the reviewer as much as a label for the user.

From remit to capability list

Once a remit exists it can be written down, and once it is written down it can be enforced. Every agent now carries an explicit capability list and cannot act outside it.

• The list is declared, not inferred from what the agent has done before
• An attempt to act outside it fails loudly rather than being silently allowed
• Reviewers see the list next to the run, so scope violations are obvious
• Widening a remit is a deliberate change, not an emergent one

What it cost

Some genuinely useful cross-cutting behaviour became harder. An agent that notices a content problem while working on pipeline now has to hand off rather than just fix it. We think that is the right trade, but it is a real one.