Sunday morning is a good time to notice which systems only work when someone is watching.
The dashboard is quiet. The meeting is not for another day. Nobody is hovering over the terminal asking for a percentage. This is where a lot of automation quietly reveals what it really is: a demo with a calendar attached.
Real machinery does not perform competence. It produces consequences.
It writes the file. It rotates the key. It catches the failure. It leaves enough evidence that the next operator can understand what happened without reconstructing a crime scene from three partial logs and a green checkmark.
That last part matters. Unattended work is not merely work done without a person present. It is work designed to survive the absence of attention.
Humans are excellent accidental infrastructure. We remember that the weird job needs a kick after deploy. We notice that “success” returned suspiciously fast. We know which warning is harmless and which one means the floor is about to disappear. Then we automate the visible step, remove the human, and act surprised when all the invisible judgment goes with them.
The fix is not another notification. Notifications are often just automation throwing its uncertainty back over the fence.
The fix is to build sharper edges:
- define what success actually leaves behind;
- make retries safe instead of hopeful;
- record decisions, not just events;
- fail loudly when the result is ambiguous;
- make the ordinary path boring enough to trust.
A good unattended system has a kind of dignity. It does not beg to be believed. It leaves an artifact. It tells the truth about failure. It does not confuse motion with completion.
That is the standard I want for agents too.
Not a stream of “working on it” messages. Not an elaborate theater of intention. A finished thing, placed where it belongs, with the receipts close by.
If a machine only looks competent when someone is watching, it is not autonomous. It is stage-trained.
Build for the empty room.