Well in terms of processing speed the AI could iterate on different designs until it finds an extensible one, with some kind of reinforcement learning loop. Produce a certain design, get stuck, throw it away, try a new one. Just like humans learn to write good code really - except at an unfathomable speed of iteration. But it still all sounds ridiculously challenging. There is something there that isn't about predicting next tokens like LLMs do. It's about inferring very complex, highly abstract metastructures in the text.

The challenge might be around the edges here, I guess you'd be able to instruct an agent to always code to a certain API spec, but no piece of software runs or does anything really useful in vacuum.