On Friday, Claude Code creator Boris Cherny made an appearance at Meta’s @Scale conference and, surprisingly, the first question from the audience was about loops.

“Are loops the next hype cycle,” the questioner asked, “or are they for real?”

Cherny’s answer was an emphatic, “Yes, they’re for real.”

“Two years ago, we wrote source code by hand. We started to transition so agents write the code. And now we’re transitioning to the point where agents are prompting agents that then write the code,” he continued. “As big as the step from source code to agents was, loops are just as important and as big a step.”

Later in the talk (around the 32:00 mark in the YouTube video posted above), Cherny got specific about the loops he keeps running in his own work. One agent is continually looking for ways to improve the code architecture, while another looks for duplicated abstractions that can be unified. They submit pull requests like any other coder, and since the code is constantly changing, they never stop running.

It’s a powerful idea, particularly with a figure as significant as Cherny behind it. With the shift to agentic AI, the focus for most users has been managing their agents as well as possible: establish clear goals, check in on discrete units of progress, and don’t let them stray too far beyond the prompt. The loop takes it a step further by authorizing a swarm of agents to work continuously in the background, endlessly. It’s a lot of trust to place in AI — but with models getting better fast, it could be the next step in getting AI to handle real work.

The first thing to recognize is that this isn’t entirely new. Recursive loops — functions that call themselves in order to repeat an action, along with a condition that stops the loop — are a mainstay of intro computer science courses. These loops are following a non-deterministic logic — that is, it’s a subagent that chooses when to stop the loop instead of a clear condition — but the same basic approach is at work. As soon as programmers started using AI to complete tasks, some version of the recursive loop, with AI overseeing AI, was bound to come up.

Unlike classic computing, agentic loops can be maddeningly simple. One of the most popular tricks is the Ralph Loop (named for Ralph Wiggum), which basically sums up all the work that the model has done and asks if it’s accomplished its goal. It’s a way of dealing with AI models getting lost as they run for too long — essentially bouncing the model back and forth until the task is complete. 

Another way to think of loops is as part of the general push for more test-time compute. As OpenAI researcher Noam Brown observed earlier this month, contemporary models can solve nearly any problem if you throw enough compute at them. That means one way to ensure a problem gets solved is to just keep throwing compute at it until it’s finished. That’s particularly true for hill-climbing problems like improving a code base, where the model can just keep making incremental improvements until it reaches a given threshold. Or, as in Cherny’s example, it can just keep making incremental improvements for as long as there’s compute to spend on it.

If that sounds expensive, it should. Like agentic AI before it, AI loops burn through tokens a lot faster than simple Q&A chatbots — and because the point is to keep the loop running all the time, there’s no ceiling to how much you can spend. That’s fine for Anthropic, which is ultimately in the token-selling business, but for everyone else, it may be a pricey way to work.

Still, depending on the problem the agentic loop is trying to solve, and the right setup that allows for oversight of token spend, drift, and other classic AI issues, the benefits could be staggering enough to outweigh the costs.

View Bio