AI loops are for real, says Claude Code creator Boris Cherny

Boris Cherny, creator of Claude Code, stole the show at Meta's @Scale conference by emphatically confirming that AI loops are 'for real.' He argued that just as the transition from hand-writing code to agents writing code was transformative, the shift to loops—where agents prompt other agents that write code—is equally significant. Cherny runs specific loops in his own work: one agent continuously seeks code architecture improvements, another looks for duplicated abstractions to unify. Both submit pull requests like human coders, and since the code constantly changes, the loops never stop.

The concept builds on classic recursive loops but with non-deterministic stopping—a subagent decides when to halt instead of a fixed condition. Popular implementations like the Ralph Loop (named after Ralph Wiggum) summarize work done and check goal completion, bouncing the model back until done. This aligns with OpenAI researcher Noam Brown's observation that throwing enough compute at models can solve nearly any problem. However, the cost is steep: unlike simple Q&A chatbots, these loops burn tokens continuously with no ceiling. For Anthropic (token seller), this is fine, but for users, it's a pricey way to work—though the benefits of autonomous, constant improvement could outweigh the costs for certain problems.