Explain how feedback design differs for an agent that acts versus an assistant that suggests.

Tanoak Agritech's FlowPilot reads soil-moisture sensors and weather data to manage irrigation. In assistant mode it suggests a schedule change for a person to approve. In agent mode, turned on to keep up during heat waves, it opens and closes valves by itself, inside set limits. Adaeze Mwangi has run irrigation for Elsworth Valley Farms for eight years.

How to answer this, stage by stage

Nobody is grading whether you know agents act and assistants suggest. They're grading whether you can say what that difference actually does to feedback.

Let's learn

For eight years, Adaeze Mwangi could tell an underwatered block from the cab of her truck, just by how the leaves held themselves in the afternoon heat.

Before FlowPilot, adjusting irrigation across Elsworth Valley Farms' twelve blocks took a full morning of manual valve checks. In assistant mode, FlowPilot cut that to twenty minutes: it suggested the day's schedule, and Adaeze approved or edited it over coffee.

Five steps, and the third one is the only genuinely hard one.

In agent mode, FlowPilot opens and closes valves on its own, within limits the farm set, no approval screen involved. That single change is the whole difference this question is really about.

The turn: the extra unopened days weren't the real problem. The real problem was that FlowPilot's own reliability was what let Adaeze stop checking, and the one time it was confidently wrong, nothing in the design gave her a reason to look sooner.

Neither panel looks dramatic. That's exactly why nobody noticed the middle step disappearing.

At its worst: a soil-moisture sensor on a young orchard block failed, reading "wet" while the ground actually dried out, and FlowPilot confidently reduced irrigation there for eleven straight days during a heat wave, before Adaeze happened to notice leaf wilt on a drive-by. That season's yield on that block never fully recovered.

What I would leave alone: routine, small, reversible valve adjustments, a few minutes either way on a well-established block, don't need a heads-up first. Flagging every small action would just rebuild the bottleneck agent mode was built to remove.

The lesson: giving a tool the power to act doesn't just speed up the good days. It quietly removes the one habit that used to catch the bad ones, and something has to be built on purpose to replace it.

Now here is the same thing as a story

The short version above is what you'd say defending the redesign to Tanoak's product team. Read this one for how the gap actually got found.

Adaeze checks the fields most afternoons, phone in hand, moving block to block faster than any dashboard could keep up with her.

For FlowPilot's first year in assistant mode, her mornings had a rhythm: coffee, open the app, glance at the suggested schedule, approve it in under two minutes, and go check the fields herself anyway, out of habit more than need.

Three branches, and only the third one existed anywhere before the redesign.

When agent mode launched, that morning ritual had nowhere to go. There was no draft to approve anymore, just a log of what had already happened. Adaeze checked it daily at first, out of the same habit. Slowly, with nothing new to catch her attention and nothing ever going wrong, the daily check became every other day, then whenever she thought of it.

Nobody can point to the exact day Adaeze stopped opening the log. It just thinned out, over weeks, the way habits do when nothing rewards keeping them.

Eleven days, and the middle stretch is the one nobody was watching.

A faulty soil-moisture sensor on Block 9, a young orchard planted the previous spring, began reporting "wet" the same week its wiring corroded. FlowPilot, trusting its own sensor, reduced irrigation there, confidently, correctly following its own bad data, for eleven days straight during the season's hottest stretch.

We assumed her attention would fade like a dial, a little at a time. It didn't. It stayed off until the day she happened to drive past.

Adaeze caught it on day eleven, not from the log, but from the leaves, the same way she'd caught problems for eight years before FlowPilot existed at all.

Run the same sensor fault forward under the redesign: a reading that flat for that long, on a block that hot, crosses the "hold and ask" bound by day two. Adaeze gets an alert the same afternoon, checks the block in person within the hour, and finds the corroded wire before a single extra day of underwatering happens.

I removed the approve screen the day agent mode shipped because it was, genuinely, the whole point: speed during a heat wave when a slow human approval could cost a block its whole season. It took eleven quiet days and one young orchard block to see that removing the pause without building anything in its place doesn't just speed up the good days. It removes the only thing that used to catch the bad ones.

The five steps, if you want to remember itNot a checklist for irrigation. FLIPS is what tells you where a real habit thinned out, and what decision to take back for it.

Four fields on every logged action, and the fourth one is what makes checking late still worth something.

The recap, one line per letter: find the person is Adaeze losing her morning ritual, locate the habit is the daily glance that agent mode removed, identify the flip is over-trust with no single trigger, pinpoint the old decision is removing the approve screen with nothing to replace it, and show the replay is catching the same fault in an hour instead of eleven days.

And if you want to be sure it really works, try it somewhere elseSame five letters, a port gate instead of a farm valve. Nothing else about the two jobs is alike.

Duskferry Port Authority uses a gate-control tool that can suggest which trucks to hold for inspection, or, in agent mode, open and close inspection lanes on its own based on cargo risk scores. Selin Aydemir supervises the gate floor.

Mapped onto FLIPS, with a different flip family this time, an input flip instead of over-trust: find the person is Selin, who used to type detailed manual notes on unusual cargo for the next shift. Locate the habit is those handoff notes. Identify the flip: once agent mode auto-clears most lanes without her involvement, she stops writing detailed notes at all, since most of what she'd note now clears itself, and starts only noting the rare holds, in a much thinner, less useful shorthand. Pinpoint the old decision: the system was built to only prompt her for a note on a hold, never on an auto-clear, so ordinary but informative cargo details, useful context for spotting a pattern later, stopped getting written down anywhere. Show the replay: a redesign that asks one thin, optional note on auto-clears too restores enough context that the next shift catches a smuggling pattern across normal-looking auto-cleared trucks within a week, instead of after a full season of silence.

A different flip family, a different port entirely, and the same shape: a habit that used to write things down, quietly stopped.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "assistants get feedback in the moment, agents need it after the fact, on a clock, plus a before-the-fact bound for anything unusual," and stop.
Cost: building a real hold-and-ask bound takes engineering time the team doesn't have this sprint. Say so, and ship a same-hour alert first; a late catch still beats no catch at all.
The model gets better, for real: if FlowPilot's sensor accuracy improves overall, agent-mode feedback still matters, because a rarer sensor fault is even easier to write off as a one-time fluke instead of the pattern it might be.

Where people run it wrong.
They copy an assistant's feedback design onto an agent and call it done, assuming a log is the same thing as a review step.
They assume attention fades gradually, like a dial, instead of planning for it to just stop, like a switch.
They build the after-the-fact log but never give anyone a reason or a schedule to actually open it.

How to use it live. When someone asks about agents versus assistants, ask yourself first: by the time a person could give feedback, has the thing already happened? If yes, your whole feedback design has to move to before the action and after it, never during.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0