Building loops is the start. Closing them is the unlock.

There is a line going around: "I don’t prompt anymore. I build loops." It is true, and it is the most under-explained sentence in AI. People say it like a flex and then change the subject. Let me say the part that usually goes unsaid.

A loop did not replace the prompt. It moved it.

You used to type the prompt. Now you write the thing that types it for you: on a schedule, with the context already loaded, with a check at the end that decides whether to run again. The model still gets a prompt every turn. You just stopped writing each one by hand. So when someone says they only build loops, what they have done is automate the typing. That is real. It is just smaller than it sounds.

Most loops stop one camp short

Usually the loop runs in a codebase, and it ends when the tests go green or the pull request merges. Quick translation if you do not live in code: a green test means the software ran the automatic checks it wrote for itself and they all passed, the little status light flips from red to green. A merged pull request means that new code got accepted into the product. Credit where it is due: the sharp ones ship the pull request, and some even deploy it. But the loop still closed inside the software. Nothing in it touched the business. No customer got handled, no number moved, no decision got made and lived with. A green test is not a closed loop. It is a closed demo: the code checking its own homework. And getting there is its own climb; earlier on, you are still the glue holding every step together by hand. That is an open loop. You close the window and the work is gone.

A few people do close it, in narrow slices. Support agents now get paid only when they actually resolve a ticket, not when they reply. That is closing the loop against a real outcome, and it is real. It just lives in one lane. Few have made it the way they run a whole business yet, which is exactly the open road ahead.

What makes a loop worth building

It closes against the real result. You set the goal and the bounds. The system finds the work, builds it, checks it, ships it, then measures what actually happened, grades it, and starts the next run higher than the last. That second half has a name: recursive action. The loop acts on what it learned, and each action sets up the next. Let an agent roam with no goal and no grade and you get expensive slop. Box it in and grade it every step and it gets better every pass. It is easy to ship and guess. The whole game is to ship, measure, and raise the floor.

The climb, plainly. Same agent, one more thing it needs at each camp.

• Base campAsk it stuff. Where we all started.
• AutomatePut it on a schedule. The typing stops.
• The fleetMany agents at once. You set the direction.
• The brainYour own data, in a shape it can use.
• The loopIt does the work, and grades itself.
• SummitLearns every step and acts on what it learns. The first half runs today. The second is the work now.

Most of the build-loops crowd is parked around automate and the fleet. The interesting air is the loop, where you grade every run and close it against the real result.

Where this goes

I did not just write the argument. I built it: a model I tune to each venture, that grades its own work and gets smarter while I sleep. That is the longer story, and it is the one I actually want you to read next. Start with The Loop Is the Product. This piece was just to clear the air first.

If you are building loops, good, you are climbing. Just be honest about which camp you are on, and do not mistake a green test for a closed one. Ship, measure, raise the floor. That is the whole job.

The same next step shows up with brains. Building one is great. The exciting move is making it act and grade itself.

A note from later

Written in June, and it still holds as mechanism. What I would change is the emphasis. Closing a loop matters because of what the closing leaves behind, not because closed is a nicer state than open. The full version of that correction is The loop was never the point.