Andrej Karpathy released "autoresearch," an open-source repository under the MIT license on March 7, 2026, which implements loop engineering, a pattern where a model pursues a goal until it arrives, planning, acting, and checking its own result. Every reliable loop has three components: a verifier that grades each attempt, state that records what was tried and failed, and a stop condition that halts when the goal is met or after N attempts. The repository ships three core files and about 630 lines of code, and now sits near 90,000 GitHub stars. In testing, Karpathy pointed it at his nanochat GPT-2 training code, where it ran for two days and completed about 700 experiments, keeping 20 genuine improvements that cut GPT-2-quality training time by 11%, from 2.02 to 1.80 hours. Separately, Shopify CEO Tobi Lütke ran autoresearch overnight on an internal model and reported a 19% improvement after 37 experiments.