Karpathy left an agent running overnight on his own training code, the code he had already spent a long time optimizing himself. It came back with about 20 improvements he had missed. The setup is almost insultingly small. One file the agent is allowed to edit, one it cannot touch, one number to beat. The agent changes something, trains for five minutes, checks the score. Better, it keeps the commit. Worse, it throws the change away and tries again. That is the entire loop. Two days, roughly 700 experiments. Time to GPT-2 quality went from 2.02 hours down to 1.80. It also caught a missing multiplier in his attention code that he had walked past. The part worth sitting with: you never edit Python here. You edit a markdown file that tells the agent how to do research. Karpathy says that file is the real benchmark now. Repo is karpathy/autoresearch.