The only AutoResearch tutorial you’ll ever need

AutoResearch is an open-source project by Andrej Karpathy that enables AI agents to autonomously run experiments, keep improvements, and discard failures in a continuous loop. The core architecture uses three files: program.md (goals and constraints), train.py (the one file the agent can modify), and prepare.py (the immutable evaluation metric). The loop runs experiments, commits successful ones to git, and resets failed ones. Beyond ML model optimization, the pattern applies to any domain with a measurable outcome: trading strategies, marketing A/B tests, prompt engineering, and website performance. Three conditions are required for success: a clear scalar metric, automated evaluation without human involvement, and a single modifiable file. A practical demo shows using Claude Code and Codex CLI to build an auto-research loop that reduces a portfolio website's load time from 50ms to 25ms in under 4 minutes.