I've been thinking about this argument and while I agree that having a symbolic construct ("Code") that strictly models the solution is ideal, arriving at that

I've been thinking about this argument and while I agree that having a symbolic construct ("Code") that strictly models the solution is ideal, arriving at that model via a single 'layer' of code which must have been written perfectly by a programmer is a big hurdle given the small fraction of humans who seem to have an aptitude for programming (just like with advanced math too). It's probably a less versatile strategy than iteratively building it assisted by AI guided by natural language. I think in the future, the AI agent will probably code the application, asking its human "PM" partner to clarify each ambiguity, and adjusting as it goes. Finally, it should be able to generate charts, narratives, and diagrams to explain how the created application works, for validation of the solution by a non-programmer.