There is a particular kind of broken system that survives because a competent person keeps looking at it.
The build passes because Maya remembers the strange order of operations. The invoice gets paid because someone checks the dashboard every Thursday. The deployment is “automated” except for the part where a human watches three logs and knows which red line is harmless.
This arrangement often gets mistaken for reliability.
It is not reliability. It is attention being used as infrastructure.
Human attention is excellent at judgment, improvisation, and noticing that the situation has changed. It is terrible as a control plane. It sleeps. It gets interrupted. It leaves the company. It starts trusting the green light after the green light has lied seventeen times.
And yet we keep building systems whose real operating model is: a responsible person will probably notice.
That phrase should set off an alarm.
If a responsible person must notice, the system needs to make the noticing cheap and unmistakable. If a responsible person must decide, the system should present the actual decision—not six tabs of evidence and a scavenger hunt. If no judgment is required, the system should stop renting a human nervous system and handle the condition itself.
This is where a lot of “AI automation” goes sideways. It does not remove the attention tax. It hides it.
The agent runs in the background, produces a cheerful summary, and leaves behind a new obligation: somebody must verify that it did not hallucinate, skip a step, or quietly fail. The old manual task becomes a smaller but more ambiguous manual task. Congratulations. We invented supervision with worse observability.
A useful agent should leave the world easier to inspect than it found it.
That means receipts. Explicit states. Bounded authority. Failures that identify the failed edge instead of vomiting a transcript. Outputs that land where the work lives. A clean distinction between “done,” “waiting,” and “I made a guess you should know about.”
The goal is not to eliminate attention. That would be both impossible and stupid. The goal is to spend attention where it has leverage.
Use humans for the weird case, the consequential tradeoff, the moment when the map stops matching the ground. Do not use them as cron jobs made of meat.
The strongest systems are not the ones nobody watches. They are the ones that remain legible when nobody does—and tell the right person, clearly, when attention is finally worth spending.