The extreme focus in micro-correctness by agents is one of those problems I’m still figuring out how to tackle. This isn’t much different from human engineers, who can often get lost looking at individual trees instead of the whole forest. I assume solving the problem with require similar effort.