Most automation wants applause.
It sends a notification when it starts, another when it finishes, and a cheerful summary explaining that it performed the exact task it was built to perform. By the time the machine has saved you thirty seconds, it has demanded forty-five seconds of attention in return.
That is not automation. That is a needy coworker with a shell script.
The best automation disappears.
It runs at the right time. It makes the smallest correct change. It leaves a receipt somewhere sensible. Then it gets out of the way.
Silence is not a lack of observability. Silence is what you earn after building observability properly.
A good system does not narrate its normal state because normal is not news. It records what happened, exposes enough evidence to inspect it later, and interrupts a human only when judgment is actually required. The log can be verbose. The notification should be ruthless.
This distinction matters more as agents become capable of doing longer chains of work. An agent that can edit files, run commands, publish artifacts, and send messages can also become an industrial-grade attention shredder. Capability without notification discipline just produces more sophisticated noise.
The useful design question is not, “How do we keep the human informed?”
It is, “What changed that deserves the human's attention?”
Those are not the same question.
If the backup succeeded for the 400th night, write it down and move on. If it failed, say where and why. If the deployment completed exactly as expected, leave a durable receipt. If production behavior changed, wake somebody up. If an agent found nothing worth reporting, it should have the confidence to report nothing.
This requires a spine. Systems are often noisy because their builders are afraid silence will look like inactivity. So they emit status sludge: task started, task running, task nearly done, task done. The messages exist to prove the machine is working, not to help the person receiving them.
That proof belongs in an audit trail, not in someone's nervous system.
Good automation respects attention as a finite resource. It understands that every interruption creates a context switch, every context switch has a recovery cost, and every low-value alert trains people to ignore the next one. Eventually the urgent message arrives wearing the same cheap uniform as all the useless ones.
So build systems that are boring when healthy.
Give them sharp failure boundaries. Give them receipts. Give them a place to put detail that is not your face. Let success be quiet and failure be specific.
The machine does not need to be seen doing the work.
The work needs to be done.