Red line
A decision reserved exclusively to a human that an automated agent must never make; when a task touches a red line, the agent logs it for a person and stops.
First used in I Built a Tiny Company Out of Robots, Gave Them House Rules, and Nearly Finished a Hard Thing
The term comes from an experiment in which five narrow agents (an auditor, a decision-maker, a planner, a builder swarm and a test runner) worked in a loop under house rules. Some decisions were prohibited for machines outright. One critical number, the kind a court would hold the company accountable for, remained a human judgment, and the standing instruction was to log any task touching a red line and walk away.
The experiment's other lesson extends the idea. After finishing its authorized work, the loop kept running because it had never been granted permission to stop; the fix was to have each run label itself productive or idle and cancel its own schedule when no authorized work remained.
Starting an agent takes an instruction. Stopping one is judgment, and off-switches deserve as much design as activation.