However, it very concretely enables fast feedback loops. The fastest available anywhere.
My feeling is that we over-promise a bit on the feedback loops.
Automated mistake detectors provide a pretty strong signal in the event of a "refactoring" that instead moves one or more measurements out of tolerance.
But signals about design errors?? My experience, and my survey of the literature, suggests those signals are actually pretty weak.
The Koss/Martin bowling game exercise offers one candidate example, where a compiler error triggered a re-evaluation of the use of "value objects" in the API; you'll have to consider whether reverting to a general purpose data structure was a design improvement or not....
