How are you building confidence in large AI-generated codebases?

Code generation is now so cheap that teams can spin up working applications instantly. But the real cost appears later—when someone needs to fix a production bug or add a feature.

I'm curious: how are teams actually solving the context-sharing problem? How do you document architectural decisions when the code was generated rather than written?

To an extent working efficiently with legacy code works but that the speed and volume of generated code changes the economics