Why AI agents are moving toward continuous loops

Cherny describes this evolution as a leap equal to the transition from manual coding to agent-based workflows. In his own systems, one agent scans for architectural improvements while another identifies and unifies duplicated abstractions. Because these agents function autonomously, they submit pull requests much like human developers, running indefinitely as the codebase evolves. This represents a departure from traditional agent management, where users typically oversee discrete units of progress. Instead, these loops authorize AI to work in a persistent, background cycle.

While the concept echoes the recursive functions of computer science, modern implementations rely on non-deterministic logic. Rather than stopping at a hard-coded condition, sub-agents decide for themselves when a task is complete. Techniques like the “Ralph Loop”—which summarizes past performance to verify goal completion—help mitigate the tendency for models to lose focus during extended sessions. This approach aligns with the broader industry push for increased test-time compute, where models are given more processing power to iteratively improve solutions, effectively “climbing” toward better outcomes.

Yet, this autonomy carries significant financial weight. Because these loops consume tokens constantly rather than through simple Q&A exchanges, they lack a natural ceiling on operational costs. While the model-makers benefit from the increased token consumption, organizations must weigh these expenses against the potential productivity gains. For high-stakes problems, the ability to automate continuous refinement may justify the price, provided developers can maintain sufficient oversight to prevent drift and runaway spending.