AI agents are beginning to run in continuous self-improving loops, a shift Boris Cherny says is as significant as the move from hand-written code to AI-generated code—but the token costs could be steep.

What happened

Claude Code creator Boris Cherny explained at Meta's @Scale conference that AI agents are now moving beyond writing code on demand to running in loops where they continuously prompt other agents to improve code architecture, unify duplicated abstractions, and submit pull requests—never stopping as long as compute is available.

What to watch

Cherny framed loops as part of a broader trend in 'test-time compute'—the idea that models can solve nearly any problem if you throw enough computing power at them. For hill-climbing problems like code improvement, loops can keep making incremental advances until reaching a threshold or until compute runs out, making the financial sustainability of continuous agentic loops a key question for businesses beyond token-selling companies like Anthropic.