Loop. If this summer at the beach you hear a word said with the air of someone in the know, this is it. The myth that comes with it: AI that corrects itself, improves with every round and breaks through — while you sunbathe.
The intern from last time has learned to check their own work and try again, without you.
What it actually is: a cycle. The agent tries, judges the result, corrects, tries again. This continues until the result is good enough.
The idea is sound and it works. But inside that “until it’s good enough” lives the devil.
You know the hot-and-cold game? One person searches, the others guide: cold, warmer, hot. The loop is exactly that: attempt, response, correction. Now take away the voices. The child wanders the beach at random. Every now and then they stumble on the treasure, but that’s luck, not searching.
The voices are the measure. A loop is only as good as its measure: something that tells the AI, at every round, how far it is and which way to go. A clear measure, and the next round is better. A vague measure, and the next round is just different.
Notice this too: the demos that leave people speechless are almost all about software code, games, mathematics. It’s no coincidence. These are the worlds with a free referee: the test passes or it doesn’t, the game keeps score, the calculation checks out. A clear voice, at every round, at zero cost. Take away the referee and the magic disappears. And your real problem, usually, doesn’t come with a free referee: if every check costs a meeting, the loop becomes the most expensive way to go in circles.
Then there’s the person who tells you that “you only break through if you write your own loop.” That’s like saying that to roast a joint you need to build your own oven thermostat. The thermostat is a loop: it reads the temperature, heats up, reads again, until it holds at 180. And its hot signal is built in: the probe. The manufacturer designed it once, well. You turn the dial and put the roast in. The loop is infrastructure, not a Saturday afternoon project.
With that, the toolkit is complete: the message (prompt), the hands (agent), the cycle (loop). In the next article we open the bonnet: what the model actually knows. And above all, until when.
In the meantime, when someone says “loop” with the air of someone in the know, ask them just one question: who’s telling you “warmer”?