I think this is 100% the right direction:

Instead of imperatively letting the agents hammer your codebase into shape through a series of prompts, you declare your intent, observe the outcome and refine the spec.

The agents then serve as a control plane, carrying out the intent.