Agent Loops, explained plainly.

Over the past few weeks, the same piece of advice keeps showing up in the tools I use and the threads I read: stop handing the AI a task list, give it a goal and build a loop around it. The major AI tools, Claude and ChatGPT among them, are all shifting from step-by-step prompts toward goals and loops.

I think this is directionally where things are headed. But it’s still early, and most of the articles I’ve seen in the last few weeks are the usual hype merchants giving you anxiety for prompting your agents directly instead of turning them loose to run on their own. Very little gets said about what these loops actually are, or the cautions worth heeding before you dive in.

This article’s goal is to ground you in the basics: what loops are, how they can be valuable to you, and what the downside is.

01 · The ideaWhat a loop is.

Imagine asking someone to tidy a room.

You could hand them a checklist:

Or you could give them a goal:

With the checklist, they do four things and stop, whether or not the room is actually clean. With the goal, they fall into a different rhythm. They look around, find something out of place, deal with it, and look again. They keep going until nothing is out of place.

That second pattern is a loop:

An AI loop works the same way. The agent looks at where things stand, decides on a move, makes it, checks the result, and goes around again:

A checklist can’t notice that you missed a sock under the bed. A loop can, because looking again is built into how it works.

02 · The key stepThe step that makes a loop work.

A loop only functions if, on every pass, the agent can answer two questions: did that last move help, and am I done yet? In the room, that’s the “look around again” step. Take it out and you don’t have a loop, you have someone tidying with their eyes closed, repeating actions with no idea whether they’re getting anywhere.

Picture a more useful job: building a board presentation. Imagine an assistant whose goal is a board-ready deck. It drafts a slide, then steps back and reads it the way the board will: does this number actually support the headline, does the story answer the question they’ll ask, would this survive the sharpest person in the room? That read-back is the evaluation step. It’s what tells the assistant whether to move on or rework the slide:

A good evaluation step is why loops can catch their own mistakes, and a weak one is how they can spin out of control.

03 · Why nowFrom instructions to goals.

For a couple of years, the way I used AI was one prompt at a time. You write a prompt, you get a response, you’re done:

That works right up until the task takes more than one pass. Preparing a board deck, researching a new market, planning a launch, writing the second draft after the first one taught you something. Real work doesn’t resolve in one step, and it changes while you’re doing it. Halfway through the deck, you realize the story is really about margin, not growth, and the whole thing needs to turn.

A loop is what lets an agent work that way. Instead of running a fixed script written before anything was known, it re-checks where things stand and decides the next move based on what’s actually true now. That’s why the language shifted from giving instructions to setting goals. The goal is the destination. The loop is what keeps the agent driving toward it and correcting as the road turns.

04 · The shapesThree ways to picture a loop.

You’ll see loops described in a few different shapes. They’re easiest to understand not as competing types but as one idea at rising levels of organization.

Start with the basic loop. Observe, think, act, repeat. Nothing fancy. Its value is that it adapts instead of marching through a fixed list. When the situation changes, the next pass takes the change into account.

Add a step, and you get OODA. Observe, Orient, Decide, Act, a decision framework that came out of military strategy. The addition is “orient,” which means understanding the context before moving. An assistant that sees a thin slide and immediately bolts on a chart is reacting. One that pauses to weigh what the board actually cares about this quarter, what last quarter’s deck promised, and what the numbers really say before it touches the slide is orienting. Same loop, better decisions.

Zoom out, and you get what engineers call a state machine. Bigger jobs have phases, and each phase runs its own loop. Think of how the board deck actually comes together:

Gathering is a loop. Drafting is a loop. Each has a different goal and a different sense of what “done” means, and the work hands off from one phase to the next. The value here is structure: a long job stays predictable because you always know which phase you’re in and what finishing it looks like.

So the three aren’t rival options. A basic loop is the unit, OODA is that unit made smarter, and a state machine is how you organize several of them across a job too big for one.

05 · The costWhat loops cost when they run.

Here’s the part you feel the first time a loop goes sideways.

A single AI request has a fixed cost. You know roughly what you’re paying before you send it. A loop doesn’t work that way, because the loop itself decides how many times to run. That open-endedness is what makes a loop powerful, and what makes it risky.

The failure mode looks like this:

This is the flip side of the evaluation step. A loop with a weak sense of whether it’s making progress can’t tell that it’s stuck, so it keeps going, rebuilding the same slide a dozen ways, pulling more data, second-guessing itself. It can run up real time and real money this way and hand you nothing you’d actually put in front of the board.

Which is why working systems don’t just turn an agent loose. They put boundaries around it: spending limits, step limits, and rules that hand the problem back to a person after a few failed attempts. None of it is glamorous, and all of it is the difference between a loop you can trust and one you have to babysit.

06 · Try itTry it yourself.

The fastest way to feel the difference is to stop handing your AI a task list and hand it a goal instead.

Next time you’d normally write something like this:

Try giving it the goal and the boundaries instead:

Then let it decide what to do next. Notice that those constraints are the stopping rules from the last section, written in plain language. “Check in after three tries” is the escape hatch. “Stop when each topic has a clear headline backed by a number” is giving it a real way to know it’s done.

What you’ll notice is that the work starts to feel less like prompting and more like handing a project to a capable assistant. That’s not a coincidence. That’s what’s happening.

07 · ClosingThe direction is steady.

It’s worth being honest that most of this is still early. The tools are improving fast, and a lot of what feels clever today will look clumsy in a year.

But the direction is steady and it’s simple: define a goal, give the agent a way to check its own work, set sensible limits, and let it work toward the outcome instead of down a list. The model does the thinking. The loop is what turns thinking into finished work.