The seven types of AI-wait behaviour (and which one you are)
Watch a room of people waiting for a model to answer and you will see seven distinct species. None of them are wrong, exactly — each one is optimising for something real. The trouble is what each one costs.
1. The Watcher
Behaviour: eyes on the progress indicator. Does not look away, not even to blink properly.
Secretly optimising for: control. If you are watching, the thing cannot surprise you.
The cost: the spinner has never once finished faster because someone watched it. Meanwhile your eyes have been locked at monitor distance and your breathing has gone shallow.
The fix: pick one task a day and physically leave the desk before it completes. Once. That is the whole experiment.
2. The Napper
Behaviour: chair back, eyes closed, four minutes of genuine rest.
Secretly optimising for: the fact that the machine is busy and the human might as well not be.
The cost: genuinely low. The only risk is drift — four minutes becomes forty, and the day quietly reorganises itself around your duvet.
The fix: an alarm. Four minutes is a nap; anything longer is a new phase of the day.
3. The Multitasker
Behaviour: the wait is treated as free time, so a second task is started immediately. By mid-afternoon there are five active threads and none of them remember who you are.
Secretly optimising for: throughput. Refusing to let 180 seconds go unmonetised.
The cost: switching costs compound, and the “quick” second task usually outlives the first.
The fix: try one break with no output at all. Nothing produced. Nobody told. It will feel illegal for about ninety seconds.
4. The Snacker
Behaviour: every wait is a small meal opportunity. The kettle is on before the prompt lands.
Secretly optimising for: the actual, legitimate pleasure of a pause.
The cost: minimal, unless the keyboard crumb situation has become a structural concern.
The fix: move one snack away from the desk. Keep the other one. Moderation, not penance.
5. The Wanderer
Behaviour: the spinner appears, the person does not. Return is announced by the door.
Secretly optimising for: input. You have correctly intuited that the best ideas often arrive away from the screen.
The cost: essentially none, except that you sometimes come back with better ideas than the model’s and have to pretend otherwise.
The fix: carry something to write with. Half of the wandering ideas are good and currently going unrecorded.
6. The Stacker
Behaviour: uses the wait to queue more agent work. Three parallel requests is a slow day.
Secretly optimising for: latency arbitrage — if every wait hides another job, no time is ever idle.
The cost: the waits multiply. You now supervise six progress bars, which is worse than supervising one.
The fix: a hard cap. Two parallel requests. Write it somewhere you will see it.
7. The Reconsiderer
Behaviour: rereads the prompt. Rewrites the prompt. Cancels the run and starts again with a slightly different word.
Secretly optimising for: correctness — or, more honestly, a feeling of agency.
The cost: restarting is the single most expensive thing on this list. Every cancellation throws away all the time already spent.
The fix: let it finish, badly, and fix the output afterwards. A flawed answer you can edit beats a perfect one you cancelled.
Which one is the healthiest?
The Wanderer and the Napper, roughly tied. Both accept the pause instead of fighting it, and neither one tries to convert the pause into productivity.
The two to watch are the Watcher and the Reconsiderer, because both feel like work while producing nothing — which is the most expensive state available to a person at a desk.
Not sure which one you are? Take the quiz. It takes sixty seconds, which is roughly one wait.
Your agent still thinking? That's one wait, and it's already paid for.