In the Loop: how agentic AI actually works with humans

I have started a hundred demos with a sentence I no longer believe: "the AI will do the work." It does not do the work. It does a draft of the work, very fast, and then something has to decide whether the draft is any good. For twenty-five years that something was a person reading code at human speed. Now the drafting happens faster than a person can read, and the interesting question of this entire era is what you put at the center to keep it honest.

I made a short film about that arrangement instead of a slide. It is called In the Loop. There is no narration, just light: cold scattered noise that gathers into a single warm core, that core splitting into fast agents, a loop that fails red and then resolves green, and finally a still point at the center of a ring spinning so fast it blurs into one calm circle. This is the essay version. Same five beats.

Foundation: the model comes first

Agentic systems are not magic added on top of a weak model. They are leverage on top of a strong one. If the base model cannot write a competent first attempt, looping it does not produce quality, it produces confident nonsense repeated five times. Every team that "tried agents and it didn't work" that I have talked to was, underneath, running a model that was not yet good enough to be worth orchestrating. The order matters: capable model first, then the loop, then the speed. Reverse it and you have built an expensive way to be wrong quickly.

Speed: fast enough to bypass you

Here is the uncomfortable part. A good agent moves faster than you can supervise it line by line. It will refactor four files, write the tests, and explain itself in the time it takes you to finish reading the first diff. That speed is genuinely useful and genuinely dangerous, because the natural human response to "it is faster than me" is to stop checking. The work blurs into pure motion and you start approving things you have not actually read.

The architectural mistake is to treat that as a supervision problem to be solved with more attention. You cannot out-attention a machine. The answer is not to watch every token. The answer is to change where you stand.

The loop: quality is something you re-run, not something you get

The single most important thing I have learned building these systems is that the first output is a starting position, not an answer. The quality lives in the loop. You run the work, you check it against something real, a test suite, a type checker, a second model asked to find the flaw, and you run it again. Three passes turns a plausible draft into something you would actually ship. In the film this is the ring that fails red on the first revolution and resolves to a calm green by the third. That is not a metaphor I invented for the visual. It is what the build logs look like.

This is why "did the AI get it right the first time" is the wrong metric. Of course it did not. Neither did you. The right metric is whether your system has a fast, honest loop that catches the wrongness and feeds it back in. A great model with no loop is a gifted intern with no code review. A loop with a weak model is review theater. You need both.

The wheel: the human keeps it aligned

Speed and the loop will happily produce a beautifully refined version of the wrong thing. Agents optimize for the goal you gave them, not the goal you meant. Left alone, a field of capable agents drifts, each spinning perfectly, none of them pointed at the same outcome. What pulls them back into formation is not more compute. It is a person stating intent: this is what we are building, this is what "good" means here, this branch and not that one.

That is the human's real job in an agentic system, and it is a promotion, not a demotion. You stop being the typist and become the source of intent and the holder of taste. You make the small number of decisions that the machine genuinely cannot make for you, the ones about what matters, and you let it move at full speed on everything downstream of that decision. The film shows this as the moment the scattered rings sense one beam of intent and all swing to face it.

The point: intense calm

The feeling I was chasing in the film is the one I actually have on a good day of building this way. It is not frantic. It is the opposite. The edges are moving violently, agents drafting and looping and verifying at a speed I could never match, and I am calm at the center, because my job is no longer to keep up. My job is to hold the intent steady and to judge. The faster the ring spins, the more still the center has to be.

That is the whole thesis, and it is why the film ends on a single point of light inside a blur. Agentic work done well is not a person racing a machine and losing. It is a person standing still on purpose while the machine moves, in the loop, not under it. Get the model right, trust the loop to earn the quality, and keep a human at the center deciding what "right" means. Everything else is just speed, and speed without a center is only noise.

The short film In the Loop is live now in the cinema. No words, about a minute, best with sound on.