> This is not just a convenience feature. It is a new way of thinking about test creation that closes the distance between the coverage you need and the code that runs it

Ouff

Sounds pretty powerful, but it's probably better to be used just to generate initial tests and then guide it manually - determinism.

Has anybody tried it and can comment on how well/bad does it work?

The nice thing about this compared to generating tests in advance is that, if your app changes in non-material ways that would otherwise break a test, you don't need to regenerate - the AI will "adapt".

It's also, arguably, a form of fuzzing, maybe? I guess? Having it be non-deterministic means it might test the same thing in different - ideally all correct - ways, helping to uncover bugs.