andrej karpathy shipped a repo that optimizes any measurable process. it's called autoresearch. the agent edits the code, runs it, keeps what improves the metric, reverts what doesn't, and repeats while you sleep. why it applies broadly: you don't write code, you write program.md, what to explore, what must never change, when to stop. every run is scored by one number, so the loop needs no human judgment mid-run. improvements get committed, failures get reverted, you wake up to the diff. the most cited proof point: shopify ceo tobi lütke pointed a variant called pi-autoresearch (built on pi, an open-source agent, not claude code) at liquid, shopify's 20-year-old templating engine. 93 commits from roughly 120 experiments, parse-plus-render time down 53%, allocations down 61%, all 974 unit tests passing. worth knowing honestly: lütke himself called it "probably somewhat overfit," and the pr remains unmerged, a benchmark result, not a shipped change yet. the pattern still generalizes well beyond that one case, gpu kernel tuning, prompt optimization, even bitcoin price modeling, anywhere you have a codebase and a number to improve. grab it: http://github.com/karpathy/autoresearch(CyrilXBT)