The AI world is getting ‘loopy’

Boris Cherny, creator of Claude Code, showcased the potential of "loops" in AI-driven software development at Meta's @Scale conference. He highlighted that loops allow agents to continuously work on improving code architecture and eliminating duplicated abstractions by submitting pull requests. This method represents a significant step forward in AI-assisted coding, enabling autonomous, ongoing enhancements. Cherny's example illustrates how one agent can focus on code architecture improvements while another identifies and unifies duplicated abstractions.

The concept of loops aligns with the broader trend of agentic AI, where AI systems manage and interact with each other to achieve complex tasks. This development is particularly noteworthy given the growing emphasis on test-time compute, as researcher Noam Brown from OpenAI noted that contemporary models can solve nearly any problem with sufficient computational resources. The agentic loop approach takes this idea further by enabling continuous, autonomous problem-solving.

The implications of loops are substantial, but they also come with significant costs. Since loops continuously consume tokens, they can be expensive to operate. For instance, Anthropic, the company behind Claude, is in the token-selling business and may benefit from this approach. However, for other users, the costs may outweigh the benefits unless properly managed. Effective oversight of token spend, drift, and other classic AI issues will be crucial to harnessing the potential of loops.