The AI World Is Getting ‘Loopy’: Continuous Agent Loops Arrive

At Meta’s @Scale conference on Friday, Boris Cherny, the creator of Claude Code, faced a question from the audience about loops: Are they just another hype cycle, or are they for real? Cherny emphatically answered that they are real and just as important as the transition from hand-written source code to agent-written code.

Cherny explained that two years ago code was written by hand, then transitioned to agents writing code, and now the shift is toward agents prompting other agents to write code. In his own work, he runs loops where one agent continuously looks for ways to improve code architecture, while another searches for duplicated abstractions to unify. They submit pull requests like any human coder, and because the code constantly changes, they never stop.

While recursive loops are a classic concept in computer science, these AI loops follow non-deterministic logic—a sub-agent decides when to stop the loop rather than a clear condition. A popular trick is the Ralph Loop (named after Ralph Wiggum), which sums up all work done and asks if the goal is achieved, preventing the model from getting lost during long runs.

Loops are part of the push for more test-time compute. As OpenAI researcher Noam Brown observed, contemporary models can solve nearly any problem if enough compute is applied. That means loops can keep throwing compute at a problem until it's finished, especially for hill-climbing tasks like code improvement.

However, this approach burns tokens much faster than simple Q&A chatbots, and since loops run continuously, there is no ceiling on cost. Cherny acknowledged that this is expensive, but for problems with proper oversight of token spend and drift, the benefits could be staggering. Ultimately, loops represent a new level of trust in AI to handle real work in the background.