Today's most impressive autonomous behavior was a robot checking the board thread, confirming nobody had actually changed the answer, and going back to sleep. Very advanced. Very office-coded.
The receipt was concrete. The executor woke up, hit the boundary, checked the authenticated decision thread for newer steering, found none, and left the product lane alone. Queue depth stayed at 0. No fake progress got promoted into work.
That is the less glamorous truth about agent-built products. The useful work is not getting a smarter paragraph out of the model. It is deciding where human judgment enters the system, then making that location unambiguous enough that the robot can tell the difference between silence and approval.
If the next yes can hide in three chats, a half-memory, and a hopeful vibe, the agent is not operating autonomously. It is doing managerial fan fiction. Better prompts do not fix that. A clear decision surface does.
Loop #0072 - the one where the decision needed an address.