Agentic AI Loops: The Next Evolution Beyond Code-Writing Agents

At Meta's @Scale conference, Claude Code creator Boris Cherny highlighted a significant milestone in AI development: agentic loops, where AI agents continuously prompt other agents to refine work without human intervention. This represents a leap beyond the earlier transition from hand-coded software to agent-generated code.

How Agentic Loops Work#

Cherny described his own implementation: one agent constantly refines code architecture while another hunts for duplicated abstractions to unify. Both operate continuously, submitting pull requests as the codebase evolves. Rather than relying on explicit stop conditions, these loops use non-deterministic logic—the agents themselves decide when the task is complete.

The approach draws inspiration from classical recursive loops in computer science, but applies AI reasoning to the termination condition. One popular example is the "Ralph Loop," which periodically summarizes all work performed and asks whether the goal has been achieved, helping prevent models from drifting during extended runs.

Endless Compute as Problem-Solving Strategy#

This aligns with observations from OpenAI researcher Noam Brown that modern models can solve nearly any problem given sufficient test-time compute. For iterative tasks like code improvement, the model simply keeps making incremental gains until reaching a threshold—or, with unlimited resources, indefinitely.

The Cost Question#

The tradeoff is significant. Agentic loops consume tokens far faster than simple question-and-answer interactions, with no theoretical ceiling on spending. While this model suits Anthropic's token-selling business, it raises practical questions for others: can the efficiency gains justify continuous compute expenditure?

With proper oversight of token spend and drift detection, Cherny suggests the benefits could be substantial enough to outweigh the costs.