On Friday, Claude Code creator Boris Cherny spoke at Meta’s @Scale conference, and surprisingly, the first question from the audience was about loops.
“Are languages the next hype,” asked the questioner, “or are they real?”
Cherny’s response was “yes, they really are.”
“Two years ago, we wrote the source code by hand. We started the transition so that agents wrote the code. Now we’re moving to the point where agents push the agents who write the code,” he said. “The bigger the step from the source code to the agents, the more important and bigger the loop is.”
Later in the speech (about the 32:00 mark in the YouTube video posted above), Cherny got specific about the twists and turns he’s been taking in his work. One agent continuously looks for ways to improve the code architecture, while the other looks for repeatable abstractions that can be combined. They submit pull requests like any other coder and never stop working because the code is constantly changing.
It’s a powerful idea, especially with a figure as important as Cherny behind it. With the agency transition to artificial intelligence, the focus for most users has been to control their agents as much as possible: set clear goals, check discrete units of progress, and don’t let them stray too far from instruction. Loop takes it a step further by allowing a bunch of agents to run continuously in the background. Relying on AI is a big deal – but as models improve rapidly, this could be the next step for AI to handle real work.
The first thing to recognize is that this is not entirely new. Recursive loops—functions that call themselves to repeat an action, with a condition that stops the loop—are a staple of introductory computer science courses. These loops follow non-deterministic logic—that is, it’s a sub-agent that chooses when to stop the loop instead of an explicit condition—but the same basic approach works. As programmers began to use AI to perform tasks, some version of a recursive loop with AI controlling the AI was bound to emerge.
Unlike classical computing, agent loops can be insanely simple. It is one of the most popular tricks Ralph Loop (named for Ralph Wiggum), in which he summarizes all the work the model has done and asks if he has achieved his goal. This is one way to deal with the loss of AI models when they run too long – bouncing the model back and forth until the task is complete.
Another way to think of loops is as part of the overall push for more test-time computing. As OpenAI researcher Noam Brown observed earlier this monthmodern models can solve almost any problem if you throw enough computing at them. This means that one way to ensure that a problem is solved is to keep throwing calculations at it until it is finished. This is especially true for hill-climbing problems such as codebase upgrades, where the model can continue to make incremental upgrades until a given threshold is reached. Or, as in Cherny’s case, it can just keep making incremental improvements as long as there’s computing to spend.
If this sounds expensive, it should be. Like previous agent AI, AI loops burn tokens much faster than simple Q&A chatbots – and since the goal is to keep the loop running all the time, there’s no ceiling on how much you can spend. This is fine for Anthropic, which is ultimately in the token sale business, but for anyone else, it can be an expensive way to operate.
Still, depending on the problem the agent loop is trying to solve and the proper setup that allows for control of token spending, drift, and other classic AI problems, the benefits can be amazing enough to outweigh the costs.
When you purchase through links in our articles, we may earn a small commission. This does not affect our editorial independence.
