At last Friday's Meta @Scale conference, Boris Cherny, creator of Claude Code, ignited a debate on the future of artificial intelligence with a clear statement: agent loops are real and represent the next major evolutionary step. When asked whether loops were just another hype cycle, Cherny emphatically answered 'yes, they're real.' According to him, two years ago code was written by hand; today agents write it; tomorrow agents will prompt other agents to write code. This transition is as significant as the leap from human-written code to autonomous agents.
Cherny described how he already uses agent loops in his daily work. One agent continuously monitors the code architecture for improvements, while another looks for duplicated abstractions to unify. These agents submit pull requests like any human developer, and because the code is constantly changing, they never stop looping. The idea is powerful: instead of managing individual agents with discrete goals, you authorize a swarm of agents to work endlessly in the background, placing a great deal of trust in AI.
Sponsored Protocol
Recursive AI loops have roots in classical computer science
The concept isn't entirely new. Recursive loops, functions that call themselves to repeat an action along with a stop condition, are a staple of introductory computer science courses. However, in agentic loops the logic is non-deterministic: a sub-agent decides when to stop the loop, not a predefined condition. Since programmers first used AI to complete tasks, some version of recursive loop with AI overseeing AI was inevitable.
Unlike classical computing, agentic loops can be surprisingly simple. One popular trick is the Ralph Loop (named after Ralph Wiggum), which sums up all the work the model has done and asks if the goal has been achieved. It's a way to handle AI models getting lost when loops run too long, bouncing the model back and forth until the task is complete.
Sponsored Protocol
Loops and test-time compute: more compute power to solve complex problems
Loops are part of the broader push for more 'test-time compute.' As OpenAI researcher Noam Brown observed earlier this month, contemporary models can solve almost any problem if you throw enough compute at them. One way to ensure a problem gets solved is to keep spending compute until it's finished. This is especially true for hill-climbing problems like improving a codebase, where the model can make incremental improvements until it reaches a threshold. Or, as in Cherny's example, it can keep making improvements as long as compute is available.
Of course, if that sounds expensive, it's because it is. Like agentic AI before them, AI loops burn through tokens much faster than simple Q&A chatbots. Since the goal is to keep the loop running all the time, there is no ceiling on spending. That's fine for Anthropic, which sells tokens, but for everyone else it might be a costly way to work. However, depending on the problem to solve and the right setup to monitor token spend, drift, and other classic AI issues, the benefits could be staggering enough to outweigh the costs.
Sponsored Protocol
To understand the underlying technologies, consider how backend development environments are evolving, such as with Node.js Backend offering flexibility to integrate AI agents. Additionally, APIs are crucial touchpoints: REST API Node.js with versioning and OpenAPI help manage interactions between agents and existing systems. For a theoretical grounding in recursive loops, see Wikipedia on recursion.
Source: https://techcrunch.com/2026/06/22/the-ai-world-is-getting-loopy
