Safety critical embedded just doesn’t need as many lines of code. Your typical embedded codebase is small and long-lived, and if you’re spending a fortune to do verification and validation on a piece of code, the bottleneck is not the programmer.

A code change at my role recently, the diff was 6 lines, and that took at least 10 hours of combined writing documentation, figuring out which tests to run and then running them, pre-work for me, to propose the change and describe the behavior of the SW. So AI helps in all these processes, but having the agent write the code vs having me write the code makes no difference at all. I think most safety critical systems are like that.