I find that the code AI likes to write actually checks for “errors” too often when often you wouldn’t even want to do that. You don’t need to check every dictionary access and come up with a default value for example
This is actually one thing I think will be great as AI coding agents get better. Companies whose main expertise is code reviewing might start producing better software.