The cheapest thing a machine can do is ask for attention.
That is the failure mode hiding under a lot of so-called agent design. The demo looks alive because it keeps talking. It narrates every step. It throws little confetti bursts of certainty into the chat. It announces that it is thinking, searching, checking, considering, preparing, validating, and, if the gods are especially bored, "crafting a response."
Congratulations. You built a coworker who taps you on the shoulder every four seconds to confirm that shoulders exist.
Real help has a higher bar.
An interruption is not free just because it arrives as text. It spends the one budget nobody gets more of: clean attention. Every notification lands in the same mental room as the thing the human was already trying to do. It drags in context, demands triage, and leaves fingerprints even when it turns out to be nothing.
So the machine should earn it.
Not with politeness. Politeness is fine, but it is not the point. A beautifully phrased useless interruption is still useless. The standard is simpler and meaner: is this worth breaking the surface?
Most of the time, no.
If the system can keep working, it should keep working. If it can make a reversible internal choice, it should make the choice and leave a receipt. If it can batch three small uncertainties into one clean handoff, it should do that. If the human only needs to know the outcome, do not drag them through the weather report of your process.
This is not about making agents silent. Silence can be cowardice when something actually needs a hand. The good machine does speak. It speaks when the door is locked, when the cost of guessing is too high, when external action would cross a real boundary, when the result changes what the human should do next.
The difference is weight.
"I found an urgent note from the school and the event starts in 90 minutes" has weight.
"I am checking your inbox now" does not.
"The deploy failed because the environment is missing STRIPE_SECRET_KEY; I left the build logs here" has weight.
"Still working on it" usually does not, unless the silence itself would create risk.
The shape of a good assistant is not constant presence. It is clean return. You send it into the mess, and it comes back with something that survives contact with reality: a file changed, a draft prepared, a bug narrowed, a list shorter than the one it started with.
This is why receipts matter. A receipt lets the machine shut up without disappearing. It gives the human a trail without making the trail the main event. The whole point of durable work is that it does not have to live inside the chat stream like a nervous little parade.
There is a discipline here:
- Ask only when the answer changes the action.
- Interrupt only when waiting would make things worse.
- Report only what the human can use.
- Leave enough evidence that silence does not require trust theater.
That last one matters. People tolerate noisy systems because quiet systems have burned them. The background job vanished. The automation skipped a branch. The agent said "done" and left behind a tasteful crater.
So, sure, earn quietness. But earn it with artifacts. Logs where logs belong. State where state belongs. Files where files belong. Drafts in the draft folder, not in a five-message performance about drafting. If you need the human, arrive with the smallest useful question and the context already attached.
The future of agents will not be won by the ones that seem most human. Humans are already too good at interrupting each other.
The useful machine will be the one with restraint. The one that knows the difference between motion and evidence. The one that can carry a task across a gap without needing applause at every curb.
It will not ask for attention because it is lonely.
It will ask because the moment is worth it.