Design the interruption and takeover experience for a running agent.

Harrow Logistics moves refrigerated freight across five states. Marrow Route is the agent that reroutes and rebooks shipments during storms and port delays. Casey Lindqvist supervises the night shift and used to keep the live action feed open on a second monitor every single hour.

How to answer this, stage by stage

Seven stages. The interviewer is watching for whether your design assumes continuous human attention, because that's the assumption that always breaks.

Let's learn

Marrow Route watches weather and port data and reroutes or rebooks freight shipments in real time, sometimes touching dozens of shipments in a single disruption event.

Before Marrow Route, a dispatcher rerouted shipments by hand during a storm, one call at a time, catching problems immediately because they were the one making every decision.

With Marrow Route live, a storm night that used to take a dispatcher three hours to work through by hand now resolves in under ten minutes of agent time. That's the number in the ops review. It's also not the number that should have worried anyone.

The turn. The extra speed isn't the problem. What changed is Casey's own behavior. A single-turn routing suggestion tool never has this problem, because a person reads every suggestion before acting on it by definition. An agent that acts on its own removes that built-in read, and whether anyone's watching becomes a design choice instead of an automatic fact.

What I would leave alone: low-stakes, easily reversible reroutes, a minor delay adjustment on a non-perishable shipment, genuinely don't need a forced checkpoint. Adding one there just trains people to acknowledge without reading, which recreates the exact problem the checkpoint exists to prevent.

The lesson: a monitoring design that depends on someone choosing to keep watching will eventually stop being watched, and it will stop exactly when things have been going well, which is the worst possible time for it to stop.

Now here is the same thing as a story

The short version above is what you'd say defending this checkpoint design to Harrow's operations leadership. Read this one for how the gap actually got found.

Casey Lindqvist has supervised Harrow's night shift for five years. She can tell which storm cell will actually disrupt a route before the weather service updates its own forecast.

Nothing dramatic happened at any single point on this timeline. That's exactly what made it easy to miss.

Marrow Route launched in the spring, and for the first six weeks Casey kept the live feed open every hour of her shift, watching every reroute as it happened. It never once got a decision wrong that mattered. By week fourteen, she'd settled into checking only the next-morning digest, since the live feed had stopped showing her anything she hadn't already seen play out correctly a hundred times before.

The exact same number of reroutes happened both weeks. Only one thing changed, and it wasn't on any dashboard.

Then came a night in October when a coastal storm forced Marrow Route to reroute forty shipments in under twenty minutes. One of them, a truck of refrigerated seafood, got rerouted onto a route with a documented three-hour weigh-station delay the agent's own data had access to but weighed as acceptable against a slightly shorter total distance.

The monitoring design was built for a dial that slowly needed less attention. What actually happened was a switch that turned off entirely.

Nobody caught it until the next morning's digest, by which point the seafood had sat past its safe window and the shipment was a total loss. Casey hadn't been careless. She'd done exactly what six weeks of a perfect track record taught her to do: trust it.

Harrow's team rebuilt the monitoring experience around a forced checkpoint instead of a passive feed: every ten actions, or every fifteen minutes during an active event, Marrow Route pauses and shows a plain-language summary of what it just did, and Casey has to actively acknowledge it before it continues.

The third box didn't exist before October. It's the only thing that changed, and it's the whole fix.

Replayed under the new design: the same coastal storm forces the same forty reroutes. At action twenty-three, the checkpoint pauses and shows Casey a plain summary including the seafood truck's new route and its flagged delay. She catches it in the eight minutes it takes to reach that checkpoint, reroutes that one truck again herself, and the shipment arrives on time.

I built the monitoring feed as passive because in week one, that lighter design felt like the right amount of friction for someone watching constantly anyway. It took losing a full truck of seafood to see that the same lightness becomes the exact gap a long, genuine streak of success would eventually walk right through.

FLIPS, the five lettersThe I step, over-trust, is the flip family that fires when things are going well, which is exactly why it's easy to miss.

The recap in one image: five letters, and the I step is the one this whole design problem lives inside.

The recap, one line per letter: find the person is Casey watching every reroute at launch, locate the habit is her watching thinning from hourly to daily, identify the flip is watching everything versus watching nothing with no in-between, pinpoint the old decision is the passive feed with no forced pause, and show the replay is catching the bad reroute at minute eight instead of minute 360.

And if you want to be sure it really works, try it somewhere elseSame five letters, a pharmacy refill agent instead of a freight router. A completely different flip family this time: not over-trust, but a workaround, someone building their own private check because the takeover view gave them nothing to trust.

Coldkeep Pharmacy runs a refill-authorization agent that approves routine prescription refills on its own. Benedek Farkas is the pharmacist who reviews anything the agent flags as unusual.

Find the person: Benedek, five years at Coldkeep, knows which patients' refill patterns are worth a second look before the agent even flags them. Locate the habit: he used to trust the agent's takeover view completely when he needed to step in. Identify the flip, a different family this time, the workaround flip: he doesn't stop watching, he starts keeping his own paper log of every takeover, because the agent's takeover screen shows only its current recommendation with no history of what it already decided or reversed on that same patient. Pinpoint the old decision: the team never built a run-history or comparison view into the takeover experience, assuming a single current recommendation was enough context for anyone stepping in. Show the replay: with a redesigned takeover view that shows the agent's last three decisions for that patient side by side, Benedek stops needing his private paper log entirely, since the actual product now gives him the comparison he was building by hand.

A dosage change flagged for review needs a hard stop even though it's rare. A routine refill needs almost nothing, even though it happens constantly.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "design for the moment nobody's watching, not the moment someone is, with a forced checkpoint and a one-tap takeover," and stop.
Cost: there's no budget this quarter for a full checkpoint system. Say so honestly, and start with a single mandatory acknowledgment at the halfway point of any multi-action event, since even one forced pause beats a purely passive feed.
The model gets better, for real: if Marrow Route's routing accuracy improves further, that's still not a reason to remove the checkpoint, since a longer streak of good decisions is exactly what makes the next miss harder for a person to catch on their own.

Where people run it wrong.
They build a visible "stop" button and call the interruption experience done, without asking whether anyone will actually be watching closely enough to press it.
They design the checkpoint cadence around a fixed clock instead of the pace of actions, missing a fast-moving disruption entirely.
They force the same heavy checkpoint on every action regardless of stakes, which trains people to acknowledge without reading.

How to use it live. If you get stuck designing an interruption experience, ask yourself one question: does this design still work on the day the agent has been right for six straight weeks? If your honest answer is "only if someone happens to still be watching," you haven't designed the interruption yet, you've designed a feed.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0