Boris Cherny, the creator of Claude Code, addressed the topic of AI loops during a recent appearance at Meta’s conference. When asked if loops were the next hype cycle or something real, Cherny responded affirmatively, stating, 'Yes, they’re for real.' He explained that the shift from writing source code by hand to using agents to write code has now evolved to agents prompting other agents to write code. Cherny emphasized that this transition is as significant as the step from manual coding to agent-based coding.
Cherny provided specific examples of loops in action, describing how agents in his workflow continuously improve code architecture and unify duplicated abstractions. These agents submit pull requests like traditional coders and operate without interruption as the code evolves. This approach represents a significant shift in how AI can handle complex tasks, with agents working in the background continuously. Cherny noted that this level of trust in AI is substantial, but with rapid model improvements, it could be the next step in enabling AI to manage real-world work.
The concept of AI loops is not entirely new. Recursive loops, where functions call themselves to repeat actions with a stopping condition, are standard in introductory computer science. However, agentic loops introduce non-deterministic logic, where a subagent decides when to stop the loop instead of a clear condition. As AI has become more integrated into programming, the use of recursive loops with AI overseeing AI has become inevitable. This approach allows for continuous problem-solving, particularly in tasks like improving codebases, where models can make incremental improvements until a threshold is met. Cherny’s example illustrates how agentic loops can keep making improvements as long as compute resources are available.
Source: techcrunch
