What do you require before an agent may change something in the physical world?

In August, an agent on my bench decided a board was dead. SSH to it timed out and a network sweep found nothing, so it reported the board as unpowered and started planning around that. The bench camera was pointed at the board the whole time, showing its lit screen, and the power supply was sourcing 0.635 A into it. One quiet channel had turned into "the board is off." So my agents now work on a ladder. Reading instruments and looking through the camera are always allowed, and any claim about the physical world has to come from a reading, not from a silent channel. Moving the camera is allowed too. Cutting power is allowed without asking, because it is the safe direction, but the agent has to tell me right away, and only through a kill path we have tested on the real hardware. Anything that energizes the bench waits for a person. So my question for this sub: where do you draw that line, and how do you prove the agent's kill path works before you trust it? The talk and the code links are in my comment below. Disclosure: I build electronics, and the agents I describe run in a harness I built; Hapimatic is my project at Solid Product Design.