The AI world is getting ‘loopy’

Boris Cherny, the creator of Claude Code, told an audience at Meta’s @Scale conference that agentic AI loops represent a shift as significant as the move from hand-written source code to agent-generated code [1]. Cherny was responding to a direct question about whether loops are a genuine advance or merely the next hype cycle. “Yes, they’re for real,” he said, describing a progression over the past two years from manual coding to agents writing code, and now to agents prompting other agents to write code [1]. He characterized the step from source code to agents and the step to loops as equally important [1]. During his presentation, Cherny detailed loops he runs continuously in his own work. One agent searches for architectural improvements while another identifies duplicated abstractions that can be unified. Both submit pull requests like human developers and, because the codebase is constantly changing, the loops never stop running [1]. The concept extends a familiar programming construct. Recursive loops—functions that call themselves with a termination condition—are a staple of introductory computer science. Agentic loops apply a non-deterministic version of that logic, where a sub-agent decides when to stop rather than relying on a fixed condition [1]. A widely used example is the Ralph Loop, named for the character Ralph Wiggum, which summarizes all completed work and checks whether the goal has been met, effectively bouncing the model back and forth until the task is complete [1]. Cherny’s framing aligns with broader research on recursive AI processes. A 2025 study that interviewed 25 leading researchers from labs including Anthropic, Meta, and OpenAI found that 20 of them identified automating AI research as one of the most severe and urgent AI risks, with participants predicting a gradual transition from AI assistants to autonomous AI developers [6]. The same study noted that 17 of 25 participants expected advanced coding and R&D systems to be increasingly reserved for internal use at AI companies or governments, unseen by the public [6]. Loops also intersect with the push for greater test-time compute. OpenAI researcher Noam Brown recently observed that contemporary models can solve nearly any problem given enough compute, making continuous loops a natural fit for hill-climbing problems such as incrementally improving a codebase until a threshold is reached—or for as long as compute is available [1]. That approach carries a cost. Agentic loops burn tokens far faster than simple chatbots, and because the goal is to keep the loop running continuously, there is no spending ceiling [1]. For Anthropic, which operates a token-selling business, that dynamic is favorable. For other organizations, the expense may be a barrier, though the benefits could outweigh the costs depending on the problem and the oversight mechanisms in place to manage token spend and model drift [1].