Loop Engineering: AI Agents Automate ML Research

A new guide explains loop engineering, a method that turns AI agents into autonomous machine learning research loops. This approach replaces manual, back-and-forth interactions with a continuous process. Here's the thing: instead of giving one instruction at a time, a loop sets a goal that the model pursues until it's met. The model plans, acts, checks its own results, and then repeats the process. This allows for iteration without constant human input. What's interesting is that a reliable loop needs three components: a verifier to grade each attempt, a state to record what's been tried, and a stop condition to prevent runaway costs. One example is Andrej Karpathy’s 'autoresearch' repository. This open-source project, released on March 7, 2026, quickly gained popularity. It features a design where an AI agent edits training code, runs experiments, and either keeps or rolls back changes based on a scoring metric. The reported outcomes are concrete. Karpathy's 'autoresearch' cut GPT-2 training time by 11% by finding 20 genuine improvements over two days. Separately, Shopify CEO Tobi Lütke reported a 19% improvement on an internal model using 'autoresearch'. The bottom line is that if you have an objective metric, these loops can significantly automate and improve machine learning research.