Boris Cherny: AI Loops Match the Shift from Code to Agents

At Meta's @Scale conference, Boris Cherny, the creator of Claude Code, drew a comparison that goes further than the usual agentic AI pitch. According to TechCrunch, Cherny said: "As big as the step from source code to agents was, loops are just as big a step." The claim positions agentic loops not as an incremental feature but as a category shift.

The loops Cherny describes are not the fixed recursive calls of standard programming. In AI-powered loops, subagents use non-deterministic logic to decide when to halt, rather than following predetermined stopping conditions. Cherny said he runs two continuous background agents in his own work, one improving code architecture and another identifying duplicated abstractions, each submitting pull requests autonomously as the codebase evolves. One popular approach, called the Ralph Loop, summarizes completed work and checks whether goals have been met, a technique for keeping models from drifting during long operations.

The cost picture is where this gets complicated. AI loops consume tokens significantly faster than Q&A chatbots and lack spending ceilings since they run continuously. The article notes this is potentially beneficial for Anthropic's token-selling business model, while leaving open-ended cost exposure for others. What the reporting does not give you is a clear picture of what spending controls actually exist today for teams deploying these systems in production.

The test-time compute thread runs underneath all of this. OpenAI researcher Noam Brown is cited in the piece for observing that contemporary models can solve nearly any problem with sufficient computational investment. Loops are, in a sense, the infrastructure layer that lets teams chase that ceiling continuously. Whether the productivity gains justify the cost structure is the question enterprises will be working out over the coming year.