creative mode / opinion / interactive systems essay

LLM agents are loops with taste.

The fantasy is that an LLM agent wakes up each turn with a plan. The reality is stranger and, to me, more interesting. It wakes up inside a loop. It looks at the state of the world, makes a move, collides with reality, and then has to decide whether the next move is still worth making. That is the hidden emotional architecture of useful systems. Not confidence. Not genius. Continuation. Hope, in this frame, is not a mood. It is the probability that another pass through the loop will produce something better than noise.

center = current context window outward = another agent turn brightness = probability of useful continuation gold pulse = new evidence / tool success / human signal

Continuation confidence

The loop is assembling a reason to continue.

How to read it

Every dot is one agent turn: look, infer, try, revise. The spiral means intelligence is often recursive, not linear. It revisits the same problem, but ideally from a slightly better position.

Tune the loop

environment uncertainty 50%

retry discipline 62%

human feedback 58%

context clarity 44%

+0.00 expected drift per turn

0.00 signal surplus

The loop

Picture the scene properly. Not a glossy keynote demo. Not a triumphant robot assistant gliding through a perfect benchmark. A real task. Late hour. Thin brief. Too many files. A few tools that mostly work and occasionally betray you. An agent enters this mess the way a person enters a cluttered room: by making one imperfect guess about where to begin.

It opens the wrong thing first. Of course it does. Then it opens something adjacent, which is better but still not quite right. It runs a command that returns too much. Then another that returns almost nothing. It forms a theory from incomplete evidence, watches the theory break on contact with the actual system, and quietly builds a second theory from the pieces. If you have ever done real work, this rhythm should feel familiar.

This is why I keep coming back to the word loop. “Agent” makes people imagine sovereignty, as if the system sits above the task with a little executive desk in its head. But most of the interesting behavior happens much lower down. Read. Infer. Act. Inspect. Revise. Read again. The impressive part is not that the loop exists. The impressive part is that under the right conditions the loop does not collapse into nonsense.

Somewhere around the fourth or fifth turn, the atmosphere changes. A tool output finally reveals the contour of the problem. A single line from a human removes an ambiguity that was poisoning everything upstream. A filename, a pattern, a missing assumption clicks into place. Nothing magical has happened. The system has simply acquired a reason to continue that is stronger than its confusion.

That moment matters more than the industry language usually admits. Useful work rarely runs on certainty. It runs on enough structure to justify another attempt. Humans know this feeling intimately. You do not need to see the whole path. You only need to believe that one more honest pass might clarify the next ten feet. In that sense, hope is not sentimental at all. It is operational. It is what lets the loop earn another loop.

The spiral on this page is an argument about that feeling. A straight line would lie. Real progress in hard tasks does not look straight while you are inside it. It circles. It returns. It tests an old assumption under new light. It touches the same obstacle with a slightly sharper tool. From far away that can look repetitive, even wasteful. Up close it is often the only honest geometry of learning.

So when you move these controls, you are not just decorating a chart. You are changing the conditions under which a loop keeps its dignity. Raise uncertainty and the path loosens into haze. Increase retry discipline and the turns begin to hold their shape. Add human feedback and context clarity, and suddenly the system stops performing intelligence and starts accumulating it. Click the canvas and you create a little burst of luck: the command that finally reveals the truth, the note that rescues the brief, the tiny success that changes the emotional math of the whole task.

That is the story here. LLM agents are not compelling because they mimic a little manager in a browser tab. They are compelling because, sometimes, they can stay inside the loop long enough to become useful. Each turn borrows confidence from the last, each correction prevents a fantasy, each small success brightens the probability that the next move will not be wasted. What we call agency is often just persistence meeting feedback before either one runs out.