There is one question I dread after half a year with an agent: "You are better at this than when you started, right?"
The reflex is to nod. But if the next question is better at what, exactly, I freeze. I use it more, my hands are quicker, I am less confused. But is that better — or just more comfortable with the same middling level I reached back in month two?
I could not tell the difference. And the reason I could not tell was not that I was too lazy to look back — it was that I never stopped to look at all.
01Repetition is not learning
There is a comforting belief: do a thing enough times and you naturally get good at it. It holds for things with instant feedback — you miss the hoop, you see it, your hand corrects next throw. But working with an agent has almost none of that feedback. The agent finishes, you get the output, you use it. If it is fine, you move on. If it is slightly off, you fix it and move on too. No hoop tells you whether you just threw well or threw lucky.
Without feedback, repetition does one thing only: it deepens the exact way you already work. If that way is good, you get good faster. If it is mediocre, you get fluent at mediocre — quicker, smoother, more confident, and still mediocre. Smooth is not progress. Sometimes it is just a rut walked enough times to feel like a road.
The agent does not rescue you here, because it does not remember. Every session it returns to the starting line. All improvement in the relationship between you and it flows in one direction only: you change how you work. And you cannot change what you have never looked at.
02The smallest loop that still runs
The good news is that a feedback loop does not have to be a system. No spreadsheet, no dashboard, no athlete's discipline required. It needs exactly three moves, and five minutes at the end of the week is more than enough.
Flag two — note it: one line for what did not work, one line for what did. On paper, in a note, anywhere — as long as it leaves your head.
Flag three — adjust: change exactly one thing for next week. A sentence in a template, a line in your rules file. One. Not ten.
The power is not in the precision — it is in the loop closing. Looking and then dropping it changes nothing. There has to be that "adjust one thing" at the end for the observation to turn into a real change in how you work next week.
Pay attention to the third flag. People tend to do the first two — everyone has sat there thinking "this week was a bit rough with the agent" — and then stop. They think it and leave it there. That is not a loop, that is grumbling. A loop only closes when an observation pushes out an action, however small. Change one sentence a week, and after three months that is twelve passes of your workflow sharpened — by real data, not by feeling.
03Three questions so you do not start from scratch
The hardest part of this loop is not the time — it is sitting down with no idea what to ask yourself, and watching five minutes drift away. So here are three questions to keep on hand and pull out every weekend:
Not "was there an error" but "where did it sit": did I delegate vaguely, skimp on context, or hand it something it could not do? Naming the cause matters more than the fix itself.
What worked deserves as much notice as what did not. If one prompt produced something noticeably cleaner, find what you did differently than usual — that is what is worth promoting into a template.
Force yourself to pick exactly one. Too many and none of them stick. One small change you actually apply next week beats ten good ideas that lived in your head and floated off.
These three questions do not need good answers. They only need to be answered regularly. One rough answer a week, compounded, traces something no single reflection ever could: a direction.
04The small loop is what separates the moving from the stuck
Five minutes a week sounds small enough to seem not worth it. But look at it on a different scale. Two people use an agent for a year. The one without a loop ends the year with twelve months of doing the same thing the same way. The one with a loop ends the year with fifty-two passes of their workflow sharpened, each by a little, along exactly the places that actually rub against them.
The gap between those two people does not open up in a week. It opens up over twelve months — and by the time you notice it, it is already wide.
The five-minute loop does not promise you will get good fast. It promises something humbler and more valuable: it guarantees your next six months are actually six months of progress, not the same month repeated six times. The next time someone asks "better at what?", you will have an answer — because you counted.