The world of artificial intelligence has become “complex”

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

The questioner asked: “Are the episodes the next hype cycle, or are they real?”

He said Cherney’s answer was emphatic: “Yes, it’s real.”

“Two years ago, we wrote source code by hand. We started moving to agents writing the code. Now we’re moving to the point where agents are asking agents to write the code next,” he continued. “Despite the size of the step from source code to agents, the loops are no less important and a big step.”

Later in the talk (around the 32:00 mark in the YouTube video posted above), Cherny spoke specifically about the loops he continues to run in his work. One agent is constantly looking for ways to improve the code structure, while another is looking for redundant abstractions that can be standardized. They make pull requests like any other programmer, and since the code is constantly changing, they never stop working.

It’s a powerful idea, especially with an important character like Czerny behind it. With the shift to agent AI, the focus of most users has been on managing their agents as best as possible: setting clear goals, checking discrete blocks of progress, and not letting them stray too far from the prompt. The loop takes it a step further by allowing a group of agents to run continuously in the background, indefinitely. There is a lot of confidence in AI, but with models improving rapidly, this may be the next step in making AI handle real work.

The first thing to realize is that this is not entirely new. Recursive loops — functions that call themselves to repeat an action, along with a condition that stops the loop — are a mainstay of introductory computer science courses. These loops follow non-deterministic logic — that is, they are a sub-agent choosing when to stop the loop rather than an explicit condition — but the same basic approach works. Once programmers started using AI to complete tasks, some version of the recursive loop was bound to emerge, with the AI ​​supervising the AI.

Unlike classical computing, proxy loops can be maddeningly simple. One of the most popular tricks is the Ralph Loop (named after Ralph Wiggum), which basically summarizes all the work the model has done and asks if they achieved their goal. It’s a way to deal with AI models getting lost when they’re running for too long – essentially causing the model to bounce back and forth until the task is complete.

Another way to think of loops is as part of the general push to further compute test time. As Noam Brown, a researcher at OpenAI, noted earlier this month, contemporary models can solve almost any problem if you use enough computing for them. This means that one way to ensure that the problem is solved is to keep performing calculations on it until it is finished. This is especially true for hill-climbing problems such as code base optimization, where the model can continue to make incremental improvements until it reaches a certain threshold. Or, as in Cherny’s example, it can continue to make incremental improvements as long as there is an account to spend it on.

If this sounds expensive, it should. Like proxy AI before it, AI rings burn through tokens much faster than simple Q&A chatbots — and because the goal is to keep the ring running all the time, there’s no cap on the amount you can spend. This is a good thing for Anthropic, which is ultimately in the token selling business, but for everyone else, it could be an expensive way to operate.

However, depending on the problem the proxy loop is trying to solve, and the right setup that allows for oversight of token spending, drift and other classic AI problems, the benefits can be amazing enough to outweigh the costs.