The Shift to Continuous Agentic Loops

AI development is transitioning from discrete, prompt-based tasks to continuous, autonomous loops. Boris Cherny, creator of Claude Code, argues that this shift is as significant as the move from manual coding to initial agentic workflows. In this model, agents do not just execute a single task; they run indefinitely in the background, constantly monitoring and improving systems. For example, one agent might focus on architectural improvements while another identifies and refactors duplicated code, submitting pull requests automatically as they find optimizations.

Mechanisms of Agentic Loops

These loops function as non-deterministic recursive processes where sub-agents determine when a task is complete rather than relying on fixed logic. Two common patterns are emerging:

• The Ralph Loop: A technique where the model periodically summarizes its progress against a goal to prevent it from getting lost during long-running tasks. It forces the model to re-evaluate its state before proceeding.
• Test-Time Compute Scaling: Loops serve as a practical implementation of increasing test-time compute. For hill-climbing problems—such as codebase optimization—models can achieve better results by simply spending more compute to make incremental improvements until a specific threshold is met.

The Trade-offs of Autonomous Execution

While agentic loops offer the potential for staggering productivity gains, they introduce significant risks and costs. Because these systems are designed to run continuously, they can burn through tokens at an unpredictable rate, creating a high financial ceiling. Developers must implement rigorous oversight to manage token spend, prevent model drift, and ensure the agents remain aligned with their original objectives. The effectiveness of this approach depends heavily on the specific problem domain and the ability to build guardrails around the autonomous process.