The AI world is getting ‘loopy’: Continuous AI agents are the next big thing

Speaking at Meta's @Scale conference on Friday, Boris Cherny, the creator of Claude Code, was asked whether AI loops are just a hype cycle or a real phenomenon. His answer was a definitive yes—they are real. Cherny compared the transition to agentic loops to the shift from writing source code by hand to having agents write code. He noted that two years ago we wrote code manually, then agents took over, and now we're moving to agents prompting other agents to write code. In his talk, Cherny explained that he runs loops where one agent constantly seeks ways to improve code architecture, and another looks for duplicated abstractions to unify. These agents submit pull requests like human coders, and because the code is always changing, they never stop.

Cherny’s endorsement gives weight to the idea that agentic loops are a major step forward. The concept involves authorizing a swarm of agents to work continuously in the background, endlessly. While this requires a lot of trust in AI, improving model capabilities make it plausible as a way to handle real work.

The article notes that recursive loops are not entirely new—they are a staple of introductory computer science. However, AI loops operate with non-deterministic logic: a sub-agent decides when to stop the loop instead of a clear condition. A popular technique is the Ralph Loop (named after Ralph Wiggum), which sums up the model's work and asks whether the goal is complete, effectively bouncing the model back and forth until done.

Loops can also be seen as part of the push for more test-time compute. OpenAI researcher Noam Brown recently observed that modern models can solve nearly any problem if given enough compute. Thus, one approach is to keep throwing compute at a problem until it’s solved—for example, making incremental improvements to a codebase indefinitely, as long as resources allow.

However, this comes at a cost. Agentic loops burn through tokens much faster than simple chatbots, and because they run continuously, there’s no spending cap. This is good for token sellers like Anthropic but expensive for others. Still, if properly set up with oversight for token spend and drift, the benefits could be staggering enough to outweigh the costs.