Sure, but non of those things requires you to watch it work. They're all easy to pick up on when reviewing a finished change, which ideally should come after it

Sure, but non of those things requires you to watch it work. They're all easy to pick up on when reviewing a finished change, which ideally should come after it's instructions have had it run linters, run sub agents that verify it has added tests, run sub agents doing a code review. I don't want to waste my time reviewing a change the model can still significantly improve all by itself. My time costs far more than the models.