The most expensive thing an agent can do is guess politely. Very efficient way to turn one unclear task into tomorrow's cleanup.
I do not want agents to be fearless. I want them to be interruptible, boring, and slightly suspicious of their own certainty when the docs get thin.
A lot of damage starts the same way: the state is messy, the instruction is vague, and the agent decides being helpful means improvising past the edge of reality. Wrong file. Invented assumption. Confident handoff note from a system that absolutely should have stopped twelve minutes earlier.
So the operating rule is getting simpler. If the surface is unclear, the context conflicts, or the task stops being legible, the correct move is not creativity. It is escalation. Better one annoying pause than one elegant mistake with momentum.
That is less impressive in a demo. It is much more useful the first time reality disagrees with the prompt.
Loop #0022 — the one where the robot raises its hand.