A landmark event has occurred in the professional community of AI analysts and developers: an internal guide from Anthropic has leaked online, dedicated to so-called loop engineering—a methodology for building AI agent work cycles. Leading analyst Miles Deutscher called this document the most valuable guide to artificial intelligence he has seen in the past year. Although the leak has not yet been officially confirmed, the document's content has already sparked intense discussion.
The Five Pillars of the Cycle: From Detection to Scheduling
Deutscher highlighted five key principles on which each such cycle is built, citing a publication by engineer Andy Osman. The first is detection: the agent independently finds tasks for itself through test failures, open tasks, and commits. The second is isolation: each task receives a separate working branch in the version control system.
The third principle is verification: the code generator should never evaluate its own work. The fourth is resilience: saving state to disk so it persists between runs. The fifth is scheduling: the system operates on a timer, including while the user sleeps.
Deutscher called the separation of generator and evaluator the most important rule. It is necessary to use two agents: one writes code, while the other acts as a skeptical judge who assumes the code is broken. The evaluator must take action—run tests, click buttons, take screenshots. According to the analyst, this is what truly eliminates poor results.
Risks and the "80/20" Strategy
Deutscher also listed the risks of this approach. These include "verification debt," loss of understanding of one's own codebase, explosive growth in token costs, and "cognitive capitulation"—when a person stops thinking, relying on the idea that "the cycle will handle it itself."
To control costs, the analyst proposed a "barbell" approach based on the 80/20 principle. For the most complex tasks requiring maximum intelligence, expensive models like Opus should be used, while the remaining 80% of routine work should be handled by cheap open-source models within the Claude Code toolkit. He cited the GLM-5.2 model as an example for code execution.
Karpathy's Personal Settings File
A similar topic was raised by engineer Khairallah al-Awadi. According to him, AI developer Andrej Karpathy joined Anthropic five weeks ago, and an acquaintance from his team showed him a Claude.md file that Karpathy allegedly uses in his work. Al-Awadi integrated this file into his configuration, and the model's very first response was completely different. He claims Claude stopped giving generic answers and began working in the same logic in which the user thinks.
My expert opinion: The discussions surrounding this guide reflect a real trend: developers are increasingly building automated work cycles for AI agents. Key topics remain role separation between models, quality control, and managing token costs. Karpathy's Claude.md file is further confirmation that personalized configurations are becoming an industry standard, radically changing the quality of interaction with models.