The agent asks for hands, not answers

The agent reaches a step it shouldn't take alone: a production deletion, an ambiguous dialog, a payment screen. Instead of guessing, it captures /display/screenshot and sends the PNG through a pipe with ?filename=approve.png. Your phone downloads the exact frame it's looking at.

You open the container's display URL: the same desktop the agent is driving: click the one thing that needed a human, and close the tab. The agent re-screenshots, sees the state change, and continues.

Human-in-the-loop usually means building an approval UI, a queue, and a review inbox. Here the loop is shorter: the agent shows you its screen, and you borrow its hands for one click.

It didn't ask what to do: it asked you to come do it.