Isn’t the whole hope is that Google (and other app vendors) doesn’t even get a say at what’s possible through Siri.

All it takes is to expose application’s UI to the model, and let the model send actions. If a human can do it tapping on the screen, the machine should be able to do the same thing.