Interface research has a long and productive history with waiting. The 100-millisecond, 1-second and 10-second thresholds are forty years old and still correct. Below 100 ms, the system feels instantaneous. Up to a second, your train of thought survives. Past ten seconds, you go and do something else, and the system needs a progress bar so you know when to come back.
AI agents have opened up a region that taxonomy never had to describe: two minutes.
Too Long to Watch, Too Short to Leave: Designing for Agent Micro-Waits (arXiv:2610.27170) names it a micro-wait and makes it a design problem. Posted 19 October 2026 by Kevin Pu, Anh Truong, Michael Terry, Carrie J. Cai, Michael Xieyang Liu and Savvas Petridis.
Why two minutes is worse than ten
The paper’s formative study interviewed twelve daily agent users, and the finding is a genuine double bind.
Staying put felt like wasted time — you are watching a machine type, which has all the engagement of watching a progress bar with extra steps. But switching away fragmented their attention and cost them the context they needed to pick the work back up.
That second half is the real cost, and it is the thing the 10-second rule does not anticipate. A ten-second wait and then a result is fine, because you held the problem in your head for ten seconds. A two-minute wait is long enough that you will leave, and leaving means that when the agent finishes you no longer remember precisely what you asked, what you had already ruled out, or why you phrased it that way. The output arrives and you have to reconstruct your own intent before you can evaluate it.
So the micro-wait does not cost you two minutes. It costs you two minutes plus the reconstruction, and the reconstruction is the expensive part.
Sidecar
Their system is a parallel assistant in the same interface as the agent. While the agent works, Sidecar:
- reads the agent’s run
- points out implicit decisions the agent made — the choices it did not ask about
- suggests task directions worth exploring
A within-subjects study with 16 participants found three effects. Sidecar helped people stay focused on their main task. It let them explore unfinished ideas without fear of triggering unwanted actions — a real anxiety, since in an agent interface every exploratory keystroke looks like it might commit something. And it helped them notice under-specification in their own prompts.
The second of those is the one I would not have predicted. The reason people stare at an agent instead of thinking alongside it is partly that an agentic interface makes idle thinking feel dangerous; a read-only companion gives you somewhere to think that cannot fire a tool call.
The first finding, though, is where the paper is most interesting and least conclusive. Sidecar does not fill the wait with distraction, which is what a notification or a tips carousel would do. It fills the wait with more of the same task, surfaced from the agent’s own run. If that holds up, the design rule is: do not entertain the user during a micro-wait, and do not ask them to leave — give them a different view of the work they are already waiting on.
The surfacing of implicit decisions is the durable idea
Strip away the waiting framing and that feature stands on its own. An agent’s run is full of choices it made without consulting you: which file, which interpretation of an ambiguous instruction, which thing to skip. Those are invisible in a final answer and visible in a trace that nobody reads because traces are long and boring.
A second model whose only job is to read the trace and tell you “it decided X on your behalf” is useful whether or not you are waiting. It is review, not entertainment — and the fact that it also happens to make the micro-wait bearable may be the secondary benefit rather than the primary one.
Caveats
The abstract reports sample sizes only — 12 and 16 — with no effect sizes and no statistical results, so everything above is a direction rather than a measurement. There is no Comments field and no journal reference on the arXiv entry, so this is an unreviewed preprint. The author list is a Google research group, and the agent being studied is the kind of agent that group builds, which bounds how far the findings travel.
There is also an uncomfortable reading the paper does not take up. Sidecar costs a second model running continuously alongside the first, to make the first one’s latency tolerable. That is a reasonable engineering trade and it is also an admission: the cheapest fix for a two-minute turn is a faster turn, and we are instead designing furniture for the waiting room.
Which does not make the furniture wrong. Micro-waits are not going away while agents do multi-step work, and “too long to watch and too short to leave” is the most useful phrase anyone has produced for the problem.
Sources
- Too Long to Watch, Too Short to Leave: Designing for Agent Micro-Waits — arXiv:2610.27170
- arXiv listing for cs.HC
- Nielsen Norman Group — Response times: the three important limits