In the professional community of artificial intelligence developers, an event has occurred that has attracted close attention from analysts. An internal document from the company Anthropic, dedicated to the so-called "loop engineering" — a methodology for building AI agent work cycles — has leaked online. Despite the fact that the leak has not been officially confirmed, this guide is already being called the most valuable AI manual of the past year.

The Five Pillars of an Effective Cycle

Analysis of the document reveals five key principles on which each such cycle is built. The first is detection: the agent must independently find tasks for itself by analyzing tests, open issues, and commits. The second is isolation: each task receives a separate working branch in the version control system, preventing chaos.

The third principle is the most important, according to experts: separation of generator and evaluator. The code generator should never check its own work. For this purpose, a second agent is introduced — a skeptical "judge" who operates under the presumption that the code is broken. It runs tests, clicks buttons, and takes screenshots, ensuring objective verification.

The fourth principle is resilience: mandatory state recording to disk to preserve progress between runs. The fifth is scheduling: the system operates on a timer, including while the user sleeps, ensuring round-the-clock productivity.

Risks and Cost Optimization

The loop engineering methodology is not without risks. Among the main ones are "verification debt," loss of understanding of one's own codebase, and, most importantly, "cognitive capitulation" of the developer, who stops thinking, relying on the fact that "the cycle will handle it itself."

To control costs, a pragmatic "80/20" approach is proposed. For the 20% of the most complex tasks requiring maximum intelligence, expensive models (e.g., Opus) should be used. The remaining 80% of routine work can be delegated to cheap open-source models within the Claude Code toolkit.

Karpathy's Personal Configuration File

In the context of this leak, another insight is also of interest. It is reported that renowned AI developer Andrej Karpathy, having joined Anthropic five weeks ago, uses a personal configuration file Claude.md. According to engineers familiar with his work, after implementing this file, the model completely changed its response style — it stopped giving generic templates and began thinking in the user's own logic.

My professional opinion: The leak of the Anthropic guide is not just another piece of technical documentation. It is a signal that the industry is moving from makeshift prompts to engineering discipline. The separation of "generator-evaluator" roles and the "80/20" strategy for optimizing token costs are practices that will soon become the standard for any serious development based on AI agents. Ignoring these principles will mean technological lag.