As someone who has been using AI extensively lately, this is my preferred way of doing serious projects with them:

Let them create the plan, help them refine it, let them rip; then scrutinize their diffs, fight back on the parts I don't like or don't trust; rinse and repeat until commit.

Yet I assume this would still be unacceptable to most anti-AI projects, because 90%+ of the committed code was "written by the AI."

> why would I want to go back and forth with an LLM through PR comments when I could just talk to the agent myself in real time?

Presumably for the same reason you go back and forth with humans through PR comments even when you could just code it yourself in real time. That reason being, the individual on the other end of the PR should be saving you time. It's still hard work contributing quality MRs, even with AI.