AI loops are the new way developers run AI: leave the bots running on a task, in a loop, and let them keep going until the work is done. The technique is taking off inside Anthropic, Meta and OpenAI, and it changes what AI can actually deliver.
The Gist
- AI loops let agents work autonomously in the background for hours, without a human watching
- One known recipe is the Ralph Loop: the bot lists what it did, then asks itself if the goal is reached
- The approach is expensive in tokens, and the cost is the main brake on wider adoption
Have ChatGPT Recap This Article
ChatGPTWhat an AI loop really is
An AI loop is a simple idea with a strong twist. You give an AI agent a goal, then you let it run on its own, again and again, until the goal is reached or the time runs out. An AI agent is just a model wired up to take actions, not only answer questions.
The agent acts, looks at the result, decides what to do next, and acts again. That cycle is the loop. Each turn the bot checks where it stands and what is left to do, then keeps moving. Most loops end either when the model thinks the goal is met or when a token budget runs out.
Boris Cherny, who works on Claude inside Anthropic, runs loops that continuously improve code architecture. The agents look for duplicated patterns, simplify them, and open pull requests on their own. A pull request is the standard way of suggesting a code change to a project.
One known recipe got a name. The Ralph Loop. The bot lists the work it just did, then asks itself a single question. Is the goal achieved. If no, it keeps going. If yes, it stops. The genius is that the model becomes its own supervisor, no human required between the steps.
The loops keep changing code while developers sleep, eat or take a break. By morning, the project has new suggestions, and the human picks what to keep. The work that used to require a focused session can now happen as background processing.
Why companies are starting to take it seriously
The promise is real value. Tasks that used to need a human present for hours can run with almost no oversight. For software work especially, a loop running overnight can move a feature forward without occupying a developer’s day.
Anthropic, Meta and OpenAI are all exploring AI loops, in different ways. The shift is comparable to the move from coding by hand to coding with AI suggestions. That move took two years to spread. The AI loops wave is moving faster because the infrastructure is already in place.
AI loops are also showing up in customer support, in writing, and in research. Anywhere a task is long, repetitive, and can be split into small steps, an agent can stay in the loop until the task is done.
The catch is the cost. A loop that runs all night burns through tokens, the unit of measure for how much an AI model processes. For a single developer, this can mean hundreds of dollars in a single night. For a company running many loops in parallel, the bill is much bigger. OpenAI’s own 34 billion dollar spending shows how fast model usage can pile up.
For now, only companies and motivated individuals run real loops. The tools and the wallet are not yet ready for a casual user to push a button and let an agent work for 8 hours straight.
Keep learning on AI Noobies:
- iOS 27 AI Features: 8 Things Your iPhone Will Do
- Signal Boss: AI Chatbots Are Not Your Friends
- AI Chatbot News: 1 in 10 Use Them, Trust Lags
How AI loops show up in the apps you already use
In the short term, you may not run a loop yourself, but the work of people around you starts looking different. A developer who was reviewing your idea at 5 pm could be reading the loop’s output the next morning. Decisions move from a single sprint to a continuous background process.
In the medium term, the loops show up inside the products you already use. Imagine your favorite writing app finishing a chapter while you sleep, or your project tool reorganizing tasks overnight. The technique fits any tool where progress can be checked against a clear goal.
There are risks. The biggest one is drift, the slow movement away from what you actually wanted. An agent in a loop can spend hours optimizing the wrong thing, and the cost only shows up at the end. Without checkpoints, a loop becomes an expensive way to be wrong.
Cost is the other limit. Loops eat tokens by the thousand, and the bill is hard to predict. The current generation of loops is for people who can pay the difference between a great idea and an experiment that did not pan out.
The main takeaway is simple. AI loops are not a hype trick, they are a real shift in how work gets done with AI. Knowing the word, the idea, and the limits puts you ahead of most people at the next dinner conversation about what AI actually does today.
Follow the story on AI Noobies.
About the Author
