Earn the Right to Interrupt
A system should observe widely and interrupt rarely. Every alert must justify the attention it takes.
A system should observe widely and interrupt rarely.
We often confuse observability with notification. If an event matters enough to record, someone must surely be told. Soon every successful backup, temporary slowdown, routine retry, and harmless recovery asks for attention. The monitoring is working, but the human around it is not.
An alert spends a limited resource: another mind’s ability to stop, understand, and act. It should therefore answer three questions. Is action required? Is it required now? Is this the right person to take it? If any answer is no, record the event, place it on a dashboard, or include it in a later summary. Do not turn information into an interruption merely because delivery is easy.
This matters more as AI agents gain initiative. An agent that reports every step has not really returned attention to its operator. One that hides uncertainty is worse. The useful middle is deliberate escalation: work quietly within clear bounds, preserve evidence, and interrupt when authority ends or the cost of delay becomes real.
The beacons of Gondor were powerful because they were not ordinary light. Each fire carried one urgent meaning across great distance: aid was needed now. Imagine them burning for every patrol report, supply count, and change in weather. The network would contain more information and convey less truth. By the time Minas Tirith called for help, Rohan might no longer be looking.
Good operational design protects the meaning of the signal. Routine success belongs in history. Degradation belongs in a digest until it crosses a useful threshold. A page belongs to a condition that is urgent, actionable, and owned.
Silence is not proof that a system sees nothing. Sometimes it is proof that the system knows the value of your attention.