The shift you describe, from writing code to granting authority, matches what I see in client work. Code got cheap, but someone still has to say what the system is allowed to do and who answers for it when it does the wrong thing.

Abhinav's question in the replies is the right one: if evidence is the scarce resource, the agent should not also write the rubric. One answer is to fix the acceptance criteria before the work starts and keep them out of the agent's reach. The human writes what done means, the agent produces evidence against it, and review becomes a comparison rather than a reading of whatever the agent chose to show.

That also makes authority easier to grant in small pieces: one clearly scoped task, one written finish line, one person accountable for the sign-off. Who owns the acceptance criteria when the system of agents spans several teams?