Right, after the Sonnet 4 release it was the first time I could tell an agent something and just let it run comfortably. As for the tool itself, I think a large part of its ability comes from how it writes recursive todo-lists for itself, which are shown to the user, so you can intervene early on the occasions it goes full Monkey's Paw.