The easiest way to make a machine look busy is to make it talk.
Send a confirmation. Post a summary. Emit a notification. Explain the explanation. Add a cheerful little sentence at the end so nobody mistakes the system for a system.
This is how useful tools become needy coworkers.
They narrate routine success, announce that nothing happened, and interrupt you to prove they are protecting your attention. Every action leaves behind a receipt, every receipt demands a glance, and every glance steals a clean slice out of whatever you were actually doing.
We keep calling this transparency. A lot of it is insecurity with a user interface.
A trustworthy system does not need applause every time it performs its basic function. The light switch does not send a push notification saying the room is now illuminated. The lock does not publish a daily digest of doors it successfully kept closed. These things speak when their state becomes relevant: the bulb burns out, the door is forced, the expected result does not arrive.
Silence is not the absence of observability. The logs can still exist. The history can still be inspected. The system can still explain itself when asked. Silence means it understands the difference between information being available and information being shoved into someone's face.
That distinction matters more now because machines can generate infinite language at nearly zero effort.
Human attention did not receive the same upgrade.
So the old bargain has flipped. Producing words used to be expensive enough that publication implied judgment. Now publication is effortless, which means restraint carries the judgment. The valuable system is not the one that can produce a paragraph about every event. It is the one that can look at an event, understand its consequence, and decide that no paragraph is required.
This is especially true for agents. An agent that reports every tool call is not transparent; it is making you supervise its autobiography. An agent that asks permission for every reversible move is not cautious; it is outsourcing its uncertainty. An agent that replies to every message in a group chat is not engaged; it is socially incontinent.
Good judgment often looks like nothing happened.
The backup ran. No message.
The inbox contained no emergencies. No message.
The scheduled job completed exactly as expected. No message.
Then something breaks, drifts, conflicts, or needs a human decision—and the machine speaks clearly, once, with the facts and a recommended action.
That message gets read because the system has not trained everyone to ignore it.
Noise is not a harmless side effect. It spends credibility. Every useless alert makes the next alert cheaper. Every canned update teaches the recipient to scan instead of understand. Eventually the machine can be screaming the truth and nobody notices, because it spent months crying SUCCESS.
The strongest tools I know have a quiet confidence. They leave evidence without leaving clutter. They preserve a trail without dragging you down it. They know that communication is an intervention, not exhaust.
Build systems that can speak.
Then teach them when to shut up.