Loops, explained without the hype.

A loop means giving an AI a goal instead of step-by-step instructions. You describe what done looks like, and the AI keeps working, checking its own progress and trying again until it gets there (or you stop it). Like a robot vacuum — you don't steer it around the room, you just ask for a clean floor.

Loops have been steadily increasing in hype over the past couple of months, so we thought it worth talking about. By the end you might not want to use them yet, but at least you'll know what they are.

On 2 June, Boris Cherny, the creator of Claude Code, told an audience that he doesn't prompt Claude anymore. Loops prompt it for him, and his job now is writing the loops.

Five days later Peter Steinberger, who built the open-source agent OpenClaw and was then hired by OpenAI, tweeted that you shouldn't be prompting coding agents at all — you should be designing loops that prompt them.

The tweet cleared 2.2 million views in a day, the replies filled up with people arguing about what it meant, and the inevitable wave of "loops will 100x your output" content followed.

Underneath the noise there is a real and useful idea. It is also simpler, older, and more expensive than the hype suggests.

A loop is a set of instructions that prompts an AI agent for you, reads what came back, decides whether the goal has been met, and if it hasn't, prompts again. You stop being the person typing prompts and become the person who wrote the loop.

It needs two ingredients — a trigger (a schedule, an event, or you kicking it off) and a goal that can be checked (the tests pass, the feature works, the ranking moved).

A thermostat is the everyday version. You don't tell it "run the heating for 14 minutes". You give it a goal — 21 degrees — and it keeps checking and adjusting until the room gets there.

Sceptics point out that computers have run tasks on timers since the 1970s, and they are half right. The new part sits in the middle.

A scheduled task runs the same fixed steps every time. A loop hands the decision to the AI, which looks at the current state, chooses what to do next, does it, and checks whether it worked.

A loop is a timer plus a decision-maker, and everything interesting is in how you stop that decision-maker running off a cliff.

Partly. Developers have been running simple loops for a while.

In 2025 Geoffrey Huntley published the ralph loop, a one-line script that feeds the same prompt into an agent over and over until a job is done, wiping the AI's memory back to a few key files on every pass so it doesn't drift. He built an entire programming language with it for about $297, and this year Claude Code and Codex turned the pattern into built-in /goal and /loop commands.

What Cherny and Steinberger are describing goes a step further. Their loops don't finish one job and stop — they are standing systems with ongoing responsibilities, like "keep my open work healthy".

They run on a schedule rather than waiting for a human to start them, they can supervise other loops, and they save their progress as they go, so a crash or a closed laptop doesn't lose the thread.

The single-task loop is old news among developers. Loops managing loops, unattended, is the new layer.

Cherny has published his own receipts, and they are striking. Over one recent 30-day stretch, 100% of his contributions to Claude Code were written by Claude Code itself — 259 changes shipped — and he deleted his code editor in November.

His starter example is one line, telling /loop to babysit his open work, fix problems, and respond to feedback as it arrives.

The idea stretches beyond coding to anything with a scoreboard. One founder runs a monthly SEO loop that reads his Google Search Console data (which reports exactly where his site ranks on Google), makes improvements, and checks a month later whether rankings moved.

He reports pages climbing from page three to page two over a couple of months at roughly $5 of AI usage per run. The same shape works for ad copy — try variants, read the results, cut the losers.

The productivity claims are marketing. Cherny's numbers are real, but he built the tool, works at the AI lab that makes the model, and has effectively unlimited usage. Nobody promising "95% automation" is publishing measurements.

The cost has moved from writing code to running the loop. Steinberger has shown monthly AI usage of around $1.3 million, employer-funded. Down on planet Earth, Uber capped its engineers at $1,500 per person per tool per month after burning through its annual AI budget in four months.

The failure mode is the loop that never stops. Give a loop a vague goal like "make this better" and it will happily go round forever, spending money and declaring victory when it shouldn't. Serious setups all use the same three guardrails — a maximum number of rounds, a check for whether progress is actually being made, and a hard spending cap.

A loop is only as trustworthy as its self-checking. An agent working fast with no feedback is a machine for generating confident mistakes. The verification step is what makes any of this work, and it's the part the hype skips.

Don't attempt a self-building software factory.

If you're curious, start with a single small loop where the result is easy to measure and each run costs pennies, then give it the three guardrails above, have it message you after each run, and check its work. If the numbers move, expand. If they don't, you've spent a few dollars finding out.

I (Kieran) have tested this myself. For around six months I had my Hermes agent run a weekly SEO audit, reading my Search Console results and making improvements to my site.

I never saw any measurable impact, and eventually I stopped the loop.

The concept is hugely appealing, and I think it has gone viral partly because coders love anything that promises less work. But like a lot of these things it's a bit too good to be true right now, and the risks and potential costs are high. One to experiment with on a small scale, not one to fully adopt just yet.