Diagnosing difficult bugs has often been considered the "hard part" and coding agents seem quite good at it?

So I'm not sure this is a good rule of thumb. AI is better at doing some things than others, but the boundary is not that simple.