Houston's airport once had a problem with complaints at baggage reclaim. It took passengers about a minute to walk from the gate to the carousel, and then they had to wait seven more before their bags started to appear. Eight minutes isn't a long wait. People complained anyway.
So the airport moved the arrival gates further out and sent the bags to the furthest carousel. Passengers now walked six times as far, and complaints dropped to almost nothing. Nobody got their bags any faster. Alex Stone told the story in the New York Times in 2012. The airport fixed the complaint without touching the thing everyone thought was the problem.
A new kind of wait
I've got a version of this now. I hand an AI agent something big, a piece of research or a folder of files to sort, and it goes away to work on it. Sometimes it's back in thirty seconds. Sometimes it's twenty minutes. I can't tell which when I press go, so I can't plan anything around it.
Other people have noticed. Rachael Hodder at Atomic Object calls it "the agent wait", "a new category of time", and she's written 26 things to do with it, sorted into four sizes: the blink (10 to 30 seconds), the pause (one to three minutes), the breather (three to seven) and the actual break (seven to twelve). Brian Hempel asked on X for a better name than "waiting for AI". Nobody's come up with one that stuck.
Hodder's list is a good one. But there's a catch in it: you can only pick from it if you know which size of wait you're in. And the tool won't tell you.
Known waits feel shorter
None of this is new, which makes it more annoying. In the mid-1980s David Maister wrote "The Psychology of Waiting Lines", and one of his principles was that uncertain waits feel longer than known, finite ones. In 1993 Jakob Nielsen put the same idea into screen terms. After about ten seconds, he wrote, people want to get on with something else while the computer works, "so they should be given feedback indicating when the computer expects to be done."
What most AI tools give you instead is a spinner, or the agent's working scrolling past faster than anyone could read it. Maister had a principle for that too: occupied time feels shorter than empty time. The tools use it the wrong way round. They fill the wait with something to watch, so the time they've just handed back gets spent watching them. (One tech boss told Axios in April that agents "operate like slot machines", which is the less charitable way of saying the same thing.)
Adi Leviim made the opposite case in UX Collective this week. AI jobs vary too much to estimate, he argues, and "a time estimate is a best guess dressed as a promise", so tools should describe the work rather than guess a number. I can see why he'd say that. Four minutes that turns into fifteen is worse than no number at all. But Hodder's four sizes are wide enough for a tool to get right most of the time. Tell me whether this is a blink or a break, and I'll sort out the rest.
Call it the interval
Nobody in a theatre sits and stares at the curtain during the interval. You know how long it is, because it's printed in the programme, fifteen or twenty minutes. You go and get a drink. A bell tells you when to come back. And the show hasn't stopped: behind the curtain the crew are changing the set, which is rather the point.
That's what this time is. An interval, handed back to you while the work carries on without you. I wrote last month about the handover, the moment an AI system passes control back to a person. The interval is the stretch in between, after you've handed the work over and before it comes back.
It only works as an interval if the tool does what a theatre does: say roughly how long, and ring when it's done. That isn't a big ask. Many agents already write themselves a list of steps before they start, so a rough size is sitting right there.
If you're designing one of these tools, put the running time in the programme.