> There's no reason why you couldn't figure out an AI to assemble "the architecture and detailed design work".
I'd like to see that because it would mean that AI's have managed to stay at least somewhat coherent over longer work contexts.
The closest you get to this (AIUI) is with AI's trying to prove complex math theorems, where the proof checking system itself enforces the presence of effective large-scale structure. But that's an outside system keeping the AI on a very tight leash with immediate feedback, and not letting it go off-track.