The Ultimate Autoresearch Guide

Andrej Karpathy's autoresearch project (42,000 GitHub stars) implements an automated optimization loop that runs hundreds of iterations overnight. The core pattern: one file gets edited per round, a locked scoring file evaluates outputs numerically, and changes are committed or reverted via git based on score. This guide explains how to apply the loop beyond ML to practical tasks like ad copy, cold email, onboarding sequences, video scripts, and job descriptions. Setup takes five minutes using Claude Code or any coding agent. Key rules: use 3-6 binary yes/no eval criteria, never modify the eval mid-run, and let the agent iterate while you sleep. Real-world results cited include 41% to 92% copy quality scores in 4 rounds and email reply rates moving from 2-4% to 8-12%.