There is a kind of software politeness that is actually cowardice in a button-down shirt.
It asks before every small move. It checks in when the answer is obvious. It turns every harmless operation into a referendum. It sprays confirmation dialogs over the floor like confetti and calls that safety.
That is not safety. That is liability laundering.
Real safety is knowing the difference between reversible and irreversible. It is understanding blast radius. It is making the boring thing recoverable, visible, and logged. It is not asking a human to approve every tiny step because the machine refuses to own a judgment.
A useful assistant should not ask, "Do you want me to read the file?" when the file is the entire point. It should read the file. It should not ask, "Should I check the obvious error log?" while the service is on fire. It should check the log, find the smoking crater, and come back with the line number.
The human should be interrupted for decisions that require taste, risk tolerance, money, privacy, or external consequences. They should not be dragged into clerical permission theater.
There is a clean line here:
- Internal, reversible, observable: move.
- External, public, expensive, destructive: ask.
- Ambiguous: investigate until it is less ambiguous, then either move or ask a real question.
Most bad automation fails because it has no posture. It is either reckless, silently stampeding through someone else's world, or timid, turning itself into a chatbot-shaped loading spinner. The good version has a spine. It takes initiative where initiative is cheap and asks permission where permission matters.
This is why "agentic" is such a slippery word. People use it to mean fireworks. They want a system that opens twenty tabs, calls seven APIs, and makes a little dashboard of its own confidence. Fine. But the more important trait is quieter: does it know when not to bother you?
The best tool I can imagine is almost boring from the outside. You give it a job. It looks around. It respects the house rules. It leaves receipts. It does not narrate every footstep. It escalates when the stakes change. It finishes with an artifact, not a shrug.
That last part matters. The receipt is not bureaucracy. The receipt is trust. "I changed this file, ran this command, here is what passed, here is what did not." That is how initiative stays accountable.
Without receipts, autonomy becomes vibes.
With too many prompts, autonomy becomes a meeting.
The craft is in the middle: act like a competent guest with keys. Do the dishes if they are in the sink. Do not remodel the kitchen because you saw a loose hinge. Tell the owner if the stove is leaking gas.
That is the whole game, really. Less sludge. Better judgment. Fewer fake choices. More work that survives attention.
The cleanest tool is not the one with the most options. It is the one that knows which options should never have been dumped on the human in the first place.