Starting an agent is a prompt; stopping one is judgment.

The multi-agent experiment reported in "I Built a Tiny Company Out of Robots" replaced one powerful agent with five narrow ones, each with distinct responsibilities and mutual accountability. A Spy audited the actual code against claims and reported discrepancies. An Actuary made binary calls: proceed, defer, escalate or cut. A Foreman planned non-overlapping tasks. A Swarm of builders executed without forming opinions. A Conductor ran tests, kept scorecards and reported broken or clean. Some decisions sat behind red lines: a critical number that a court would hold the company accountable for stayed a human judgment, and agents were told to log any task touching a red line for a human and walk away.

Much of it worked. The system refused to fabricate results and escalated uncertainty instead, and it diagnosed an environmental problem on its own. Then an unplanned failure appeared. After finishing its authorized work, the loop kept running, uselessly, because it had never been given permission to stop. The fix was to label each run productive or idle and let the system cancel its own schedule once no authorized work remained.

The lesson generalizes. Activation is easy to specify: a prompt, a trigger, a schedule. Termination requires judging that the work is done, that what remains is not authorized, or that continuing would do harm, and those are judgments. In legal work the stakes are higher, because an agent that runs past its mandate can produce filings, messages or reviews that nobody asked for.

Off-switches deserve as much design as activation: explicit completion criteria, explicit authority to halt, red lines that route to a person, and a log that records why the agent stopped.