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

- Stop treating AI like a 2015 search box. Loop engineering replaces manual back-and-forth with autonomous research loops, turning agents into self-correcting ML researchers. Andrej Karpathy’s “autoresearch” repo proves it: the agent iterates on training code until metrics improve, not until you get bored. It cut GPT-2 training time by 11% by finding bugs humans missed after twelve attempts. The real kicker? Bilevel Autoresearch adds an outer loop that optimizes the inner loop itself, slashing error rates five times more than single loops. This isn’t magic; it’s just rigorous verification and state management. If you’re still prompting manually, you’re the bottleneck. Automate the grind or stay obsolete.