> …agents are predisposed to write extremely defensive, odd-case-handling code… the SNR ratio is very low. You get a spaghetti that is unlikely to crash, but really hard to extract the gist of. And finding more global bugs can be difficult because what should be structural impossibilities can be coerced into silent skips.

Have you tried asking for a summary of the logic being relied on, and asking LLM to simplify the code?

I’m 85% joking, of course, given the current state of automated code generation and the terrible summaries you usually when asking for a summary of text. But in theory that’s what they are supposed to do.