Guide to Loop Engineering: How ‘autoresearch’ and ‘Bilevel Autoresearch’ Turn AI Agents Into Autonomous Machine Learning ML Research Loops

Loop engineering enables AI agents to autonomously conduct research and optimization by repeatedly proposing changes, evaluating results, and deciding whether to keep or discard modifications until a goal is met. A functional loop requires three components: a verifier that grades each attempt using an objective metric, state that records what was tried and what remains, and a stop condition that prevents excessive costs. The pattern matters because humans typically exhaust after roughly a dozen experiments, while loops can run continuously, and early implementations have demonstrated concrete improvements such as reducing model training time by double-digit percentages. More advanced "bilevel" loops add an outer loop that watches and modifies the inner loop's search strategy itself, achieving significantly larger gains by breaking the patterns that cause searches to stall.