Boris Cherny, the creator and head of Claude Code at Anthropic, asserted that "AI loops" represent a significant advancement in artificial intelligence, comparable in impact to the shift from manual coding to agent-driven code generation. Speaking at Meta’s @Scale conference, Cherny emphasized that these continuous, agent-driven processes are "for real" and are already dramatically enhancing developer productivity. This development signals a new era where AI agents not only write code but also continuously refine and improve it.
Cherny explained that AI loops involve agents prompting other agents to write code, and then continuously monitoring and improving that code. He detailed his own use of such loops, where one agent optimizes code architecture while another identifies and unifies duplicated abstractions, submitting pull requests like human developers. This continuous operation allows for constant improvement, pushing the boundaries of what AI can achieve autonomously.
The concept extends beyond traditional recursive loops in computer science, featuring non-deterministic logic where a subagent decides when to halt the process. While not entirely new, agentic loops, such as the "Ralph Loop" which validates an AI's progress against its goal, are proving effective in preventing models from drifting off task during extended operations. This approach is part of a broader trend towards increased "test-time compute," where models are given persistent computational resources to solve complex problems through iterative refinement.
The implications for productivity are substantial; Anthropic has seen a nearly 70% increase in productivity per engineer since the beginning of the year, despite the company tripling in size. Cherny noted that up to 90% of Claude Code itself is now written using Claude Code, demonstrating the technology's self-improving capabilities. However, this continuous, token-intensive operation comes with significant cost implications, making efficient oversight of token spend crucial for broader adoption.
Cherny, whose background includes significant roles at Meta before joining Anthropic to lead Claude Code, highlighted that the comparison for AI tools should be against the cost of engineering labor, not other software subscriptions. He cited examples of entire codebases being rewritten in days, a task that previously would have taken a year. This shift underscores a fundamental change in engineering workflows, moving engineers towards orchestrating AI agents rather than solely writing code.
