source: techcrunch ai: the ai world is getting ‘loopy’
level: technical
at meta's @scale conference, claude code creator boris cherny said loops are the next major shift in ai development. he described a transition from writing code by hand to agents writing code, and now to agents prompting agents that write code. cherny runs loops where one agent improves code architecture and another unifies duplicated abstractions, submitting pull requests like human coders. because the codebase changes constantly, these agents never stop running.
loops build on recursive programming concepts but use non-deterministic logic, where a sub-agent decides when to stop. a simple example is the ralph loop, which repeatedly checks if a task is complete to keep models on track. this approach aligns with the push for more test-time compute, as noted by openai researcher noam brown, who observed that throwing enough compute at a problem can solve it. for tasks like incremental code improvement, loops can keep refining until a threshold is met or compute runs out.
the main drawback is cost, since loops burn tokens far faster than simple chatbots and can run indefinitely. while this benefits token-selling companies like anthropic, others may find it expensive. however, with proper oversight of token spend and drift, the productivity gains could justify the expense for complex, ongoing tasks.
why it matters: agentic loops could automate continuous code improvement and other persistent tasks, but managing compute costs and oversight will be critical for practical use.