Agents have been useful at providing examples, and quickly searching through errors.

But humans can sense context in a way no machine, to this day, has been able.

Also I see it highly unlikely that a person can understand the output of an agent, if they aren't familiar with the work itself.

Agents are more like assistants. The complement, not substitute, an already capable professional.

This professionals often have to steer the wheel, so the agent looks elsewhere, and it doesn't continue looping into nonsense.