You can ask a model for output directly and stop, or you can recursively ask it to keep refining the output.

That seems to fit the fast vs slow model of human thought reasonably well.