It's the architecture. That's the correct answer here. The current models produce good implementations. They're also quite good at identifying and planning for edge cases. But, today, a successful software project requires picking the right atoms for the job.

Some of the calculus for that picking will change, since volume of code that must be produced becomes less of an issue. And I don't doubt that models next year and year after will be able to make better formative architectural choices. But as it is now, I'm certain that actual systems handling real workloads require a human designer.

Someone who knows what good software looks like is empowered with agents. Someone without that knowledge isn't going to create a high quality system yet.