Design the loading experience for an agent that runs for two minutes.

A frozen line for two minutes doesn't feel like patience. It feels like nothing is happening, and empty hands go looking for something to do.

How to answer this, stage by stage

Nobody is grading whether you can name loading screen patterns. They're grading whether you know a silent wait teaches a person not to trust it, and whether your design survives the moment the agent's first guess is wrong.

Let's learn

Fluxmend is an agent that watches Corvantel's data pipelines overnight. When one breaks, it reads the logs and the schema itself, works out what most likely changed, tests a fix against the real failed batch, and reports back. The whole check takes about two minutes.

Before Fluxmend, tracing a broken pipeline by hand took Dmitri about forty five minutes: pull the logs, diff yesterday's schema against today's, guess at a fix, rerun, and hope. Fluxmend cut that same search down to about two minutes.

The first version of Fluxmend showed one line for the full two minutes: "Diagnosing pipeline failure..." The wait really had dropped, from forty five minutes to two. But the number of Slack messages asking "is this thing stuck" didn't drop. It went up, to about three a week across the on call rotation.

The two minutes were never the problem. Two minutes is a fair trade for forty five. The problem was a frozen line with nothing behind it, and a frozen line reads exactly like a hang. So people stopped waiting and started doing the trace themselves, in a second tab, the same forty five minute habit Fluxmend was built to end.

On the worst night, two engineers on call each traced the same broken join by hand, in two separate tabs, while Fluxmend was still running its own check. Both landed a fix within four minutes of each other. One fix undid the other's, and the pipeline broke a second time before sunrise.

What I'd leave alone: anything that resolves in a few seconds, like a single schema check on a small pipeline, doesn't need this. A plain small spinner is honest there, and a five stage narrative for a three second wait would read as noise, not care.

The lesson: a loading screen isn't slow because of the clock on it. It's slow because a person watching it has nothing to hold onto, so their mind fills the gap with the worst guess it can make. Give the gap something true to look at, and the same two minutes stops feeling like a hang.

Now here is the same thing as a story

Read the story below when you want to feel why a live trail earns more trust than a fast, silent agent, not just be told that it does.

Corvantel's nightly pipeline run finishes around 3 a.m., and for two years, Dmitri Aabel was the one who found out first when it didn't. He knew the logs cold. Give him a stack trace and a schema diff and he could usually name the broken join before his coffee finished brewing.

Fluxmend arrived to do that same trace for him. The first month was good. A pipeline would break at 2 a.m., Fluxmend would spend its two minutes reading, testing, confirming, and by the time Dmitri opened his laptop the fix was already applied. He started using those two minutes to answer one Slack message or pour a coffee, trusting the agent completely, the way you trust a coworker who has been right ten times running.

The habit thinned in three small beats. First he watched the spinner the whole two minutes, a little anxious, waiting for the line to change. Then, after a run of clean fixes, he stopped watching at all and walked away for the full two minutes without a second thought. Then one ordinary Tuesday, a run took a little longer than usual, and the label never moved to say why. Nothing about that one run was a disaster. But after it, Dmitri started opening a second tab out of habit on every run that followed, even the ones that finished clean in ninety seconds.

The real trigger was smaller than that. One run's second stage simply took eleven seconds longer than the runs before it. That's the whole thing. The spinner said the same six words it always said, and eleven quiet seconds was enough to plant the thought that it might be stuck.

From then on, Dmitri's hands went back to the old forty five minute habit, in parallel, on almost every run, whether Fluxmend needed the help or not. Most nights this cost him nothing but wasted attention. Then came the night the retry loop actually did run long, three minutes forty against a two minute budget, spinner unchanged the whole time. By minute three he was elbow deep in the logs himself. His fix and Fluxmend's fix landed four minutes apart, disagreed with each other, and the pipeline broke a second time before sunrise. Sorting that out cost another twelve minutes and one very awkward stand up.

The real cost was never the twelve extra minutes that one night. It was that Dmitri quietly stopped trusting every run after that, clean or not, and started doing the work twice as a matter of habit, which meant Fluxmend's two minutes had stopped saving anyone anything.

The decision worth taking back happened in a fifteen minute planning meeting, months before any of this. Someone asked what the loading screen should say while Fluxmend worked. Someone else said, reasonably, "just a friendly line so it doesn't look hung," and the team picked one label and moved on, worried that showing the agent's internal steps would overwhelm people with logs nobody wanted to read.

Run that same bad night with the old design, and Dmitri's hands go straight to the keyboard the moment the spinner sits still too long. Run it with the new one: at one minute forty, past where Dmitri's old trigger used to fire, the screen doesn't freeze. It says, "still testing fix, no verified result yet," and widens its own estimate by thirty seconds. Dmitri reads it, and waits. At two minutes five, it confirms: fix verified, zero errors. No second tab. No collision. No second break that night.

One design gave him nothing to look at for two minutes and he filled the gap with his own hands. The other gave him a running trail he could actually watch, and his hands stayed still.

What I'd tell the version of me in that planning meeting: hiding the checkpoints wasn't protecting anyone from noise. It was just moving the guessing from the screen to the person, and a person will always fill a silent two minutes with their worst assumption.

SPARK, built for a wait nobody could see into

Not a checklist to recite. Each letter has to survive the same retry loop that the story just walked through.

What Dmitri did alone, before any agent touched the pipeline.

Four real checkpoints, each landing on screen the moment it's true.

The hypothesis label is what keeps a wrong guess from ever reading as a confirmed fact.

Three things worth stating directly, since this is where the real judgment sits. The alternative rejected was a smooth animated percentage bar, which lied about a process that doesn't move at a steady rate, five seconds here, sixty there. The AI specific failure worth naming by name is confident wrongness: an agent narrating a hypothesis in the same tone it uses for a confirmed fact, which is exactly how a person ends up trusting a wrong cause. The guardrail is the hypothesis label itself, cleared only once a real verification step runs, never on the agent's say so alone. And the trade off worth naming too: emitting a human readable status line at each checkpoint costs a small amount of latency and a few extra tokens over working silently. That's the cost being accepted, on purpose, in exchange for a person who stays out of a second tab.

And if you want to be sure it really works, try it somewhere else

Same five letters, a law firm instead of a data team, and this time the wait varies by document length instead of retry loops.

Clauseline is an agent at Oakspire Legal that reads a batch of documents before disclosure and flags anything that looks privileged or irrelevant, so a human reviewer doesn't have to open every file cold. A typical batch takes about ninety seconds. Torkel Ostrowski runs paralegal operations there.

S, situation: before Clauseline, paralegals skimmed every document by hand for privilege markers, slow and exhausting on a large batch, and Torkel had never had to design what shows on screen while an AI agent works, only reviewed its final flags once done.

P, payoff: the habit worth building isn't "trust the flag count because the run finished fast." It's the review team watching the trail build on a batch that's about to go to opposing counsel, and trusting each flag because it shows how it was reached.

A, anchor: the same live checkpoint log. "Reading batch, three hundred forty documents." "Hypothesis: twelve flagged for privilege, checking against the custodian list." "Confirmed: nine of twelve are genuine privilege hits, three are false positives, external CC only." The estimate widens per document, not per generic stage.

R, risk: documents in one batch can vary wildly in length, a two hundred page contract sitting next to fifty one page emails, so a uniform progress read would be even more dishonest here than at Corvantel, and a wrongly confirmed privilege flag risks a real document reaching the other side.

K, keep out: no auto redaction and no auto send to opposing counsel on day one. The agent flags and narrates. A person still decides.

Same anchor, a different reason the estimate has to widen honestly.

Same method, a different weak spot: Fluxmend's wait varies by which stage hits a retry loop. Clauseline's wait varies by which single document in the batch happens to be long, so the estimate has to widen for one oversized file the same way it widens for a slow stage, or the trail quietly stops telling the truth about what's actually taking the time.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the anchor, live checkpoints, hypothesis labeled until confirmed, an estimate that widens honestly, and give the one example line, "Confirmed: nine of twelve are genuine privilege hits."
Cost: there's no budget this quarter for a polished live dashboard. Ship the plain staged text log first. The trust comes from the labels, not the animation.
The model got better, for real: say Fluxmend's model gets faster and the whole run drops to forty five seconds. That's not a reason to collapse it back into one frozen spinner. The same silence problem returns at any duration once it crosses a few seconds, it just returns sooner.

Where people run it wrong.
They animate a percentage that always glides smoothly to one hundred, which teaches people the number means nothing.
They stream the model's raw internal reasoning as if it were already true, which is how a wrong cause reaches a person dressed as a fact.
They build the narrative once and never update the labels when a run actually gets stuck, so a genuinely broken run still says "diagnosing..." forever.

How to use it live. Say the real tension out loud before answering: "is this asking me for a nicer spinner, or for what happens the moment the agent's first guess is wrong." That buys a beat, and it's almost always the second one.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0