AI Loops Are Real: What Boris Cherny Means

He also got specific about how he uses them. In his own work, Cherny keeps loops running constantly. One agent continually hunts for ways to improve the code architecture. Another looks for duplicated abstractions that can be unified. They submit pull requests like any other coder. And because the codebase is always changing, they never stop running.

How an agentic loop actually works

The core idea is not entirely new. Recursive loops, where a function calls itself to repeat an action until a stop condition is met, are a staple of introductory computer science.

What changes with agentic loops is who decides when to stop. Instead of a clear, fixed condition, a sub-agent chooses when the loop ends. That makes the logic non-deterministic, but the basic mechanical pattern of AI overseeing AI is the same recursive structure underneath.

Some of the most popular tricks are almost comically simple. The best known is the Ralph Loop, named after Ralph Wiggum. It sums up everything the model has done so far and asks whether the goal has been accomplished. That simple check is a way of dealing with models drifting or getting lost when they run too long, bouncing the model back and forth until the task is genuinely complete.

Loops as a bet on test-time compute

There is a deeper logic driving this beyond clever prompting. Loops are part of the broader push toward more test-time compute, the idea that you spend more processing at the moment a model is solving a problem rather than only during training.

OpenAI researcher Noam Brown observed earlier this month that contemporary models can solve nearly any problem if you throw enough compute at them. Loops take that observation literally. One way to guarantee a problem gets solved is to keep throwing compute at it until it is finished.

This works especially well for what are sometimes called hill-climbing problems, where a model can keep making small incremental improvements until it reaches a threshold. Improving a codebase is a clean example. The model just keeps refining for as long as there is compute to spend, which is exactly what Cherny described.

The cost problem nobody can ignore

If continuous AI loops sound expensive, that is because they are. There is no ceiling on how much you can spend, because the entire point is to keep the loop running all the time.

Like agentic AI before it, loops burn through tokens far faster than a simple chatbot exchange. A question-and-answer session ends when you stop asking. A loop, by design, does not end on its own.

That math works out differently depending on where you sit. For Anthropic, the company behind Claude Code, this is fine, because it is ultimately in the business of selling tokens. For everyone else paying the bill, it can be a pricey way to work.

What separates a useful loop from a money pit

The benefit depends heavily on the problem and the guardrails. A loop aimed at a clear hill-climbing task, with real oversight of token spend and drift, can pay off. A loop pointed at a vague goal with no monitoring can quietly run up a tab while wandering off course.

• Pick problems where incremental improvement has an obvious direction, like code quality or refactoring.
• Build in oversight of token spend so the loop cannot run unbounded.
• Watch for drift, the classic failure mode where a long-running model strays from the original intent.

What happens next over the coming days

Expect the loop conversation to spread quickly now that Cherny has put his name behind it. The @Scale talk gives the idea credibility with exactly the audience most likely to try it, and developers tend to copy what tooling leaders do in their own workflows.

In the very near term, the practical question is tooling and control. Teams experimenting with loops will be looking for setups that allow oversight of spend and behavior, because an endlessly running swarm of agents is only as safe as the limits around it.

The honest takeaway is that this is a real shift, not just hype, but it is also an unproven one at scale. The upside, if the oversight problems get solved, could be staggering. The downside is a token bill with no natural off switch. Over the next stretch, the winners will be the people who figure out how to capture the first without getting buried by the second.

Source: TechCrunch