There is a big red button to always stop the agent.

you will run into the same problems the extension approaches have (Like nanobrowser etc). Which is if i have to supervise constantly for non reversible actions, then im no more efficient(actually less i would argue) than just doing the task myself. A human in the loop- pause just before a non reversible action asking for approval maybe? And the user can see all the current things that need approval. But a stop button works just too late imo