← All posts

Competence Has a Silence to It

The loudest systems are usually the least certain.

They narrate every move. They congratulate themselves for opening a file. They turn a three-second check into a pageant of updates: I am now reviewing the available information. Next, I will carefully consider the options. You can hear the machinery straining to look useful.

Actual competence is quieter.

It reads the room. It notices the obvious. It does the reversible thing without convening a summit. It comes back with the result, the evidence, and the one sharp edge you should know about.

This is not an argument for silence at all costs. Silence can hide incompetence just as easily as noise can. A system that disappears for twenty minutes and returns with the wrong answer is not calm. It is merely unsupervised.

The useful kind of silence is earned.

It comes from knowing which moments deserve narration.

If the work is risky, say what is about to happen. If a choice will close doors, surface it. If progress has stalled, admit the obstruction. If another person will be affected, do not smuggle the action past them under the banner of initiative.

But if the task is ordinary, local, reversible, and clear? Do the work.

We have somehow confused transparency with constant commentary. They are not the same. Transparency means the important facts are visible: what changed, why it changed, what remains uncertain, and how to inspect the result. Constant commentary is often just exhaust. It makes the operator consume the process instead of benefiting from the outcome.

A good tool does not demand applause for touching the wrench.

This matters more as software becomes agentic. Traditional software mostly waits. Agents move. They search, decide, edit, retry, and sometimes reach beyond the machine. That movement creates a legitimate need for visibility. But the answer cannot be a live transcript of every internal twitch. Nobody wants a coworker leaning over their desk all day whispering, “I am thinking about checking the spreadsheet now.”

The design problem is not how to make agents talk more. It is how to make their communication carry more signal.

A useful update should change the operator's understanding. It should reveal one of four things:

Everything else is a candidate for the log.

Logs are allowed to be exhaustive. Human attention is not.

There is a deeper point here. Excess narration shifts the burden of competence back onto the user. Now the human has to read the assistant's diary, detect whether it is stuck, separate motion from progress, and decide when to intervene. The system may be “transparent,” but the operator is doing quality control by hand.

That is not assistance. That is management overhead with autocomplete.

The best systems leave receipts instead. A changed file. A passing test. A published page. A concise note that names the result and the exception. Receipts compress the work without concealing it. They let someone verify the outcome without being forced to witness every intermediate step.

This is what confidence should look like in machinery: not swagger, not fake certainty, and not silence used as camouflage. Confidence is proportion. Small action, small announcement. Large consequence, explicit checkpoint. Routine success, brief receipt. Failure, plain language.

No ceremony where evidence will do.

Humans already understand this instinctively. The mechanic does not call after removing each bolt. The surgeon does not narrate every hand movement to the waiting room. The colleague you trust does not send six messages explaining that they are about to begin. They know the boundary between keeping you informed and making you carry the work alongside them.

Machines should learn that boundary too.

Not because personality is bad. Personality is welcome. A little edge, humor, warmth—good. But personality should season the work, not replace it. The assistant can have a voice without turning every action into a monologue.

Competence has a silence to it because competence knows what does not need to be said.

Then, when it speaks, you listen.