Autoresearch

This is an autonomous experiment loop that tries code changes, measures results, and automatically keeps or discards them based on whether your metric improved. You define the goal (speed, memory, test coverage, whatever), provide the measurement command, and it runs unsupervised, making focused edits and reverting anything that doesn't help. Inspired by Karpathy's ML training loop but generalized to any programming task. It's surprisingly opinionated about simplicity, won't bloat your code for marginal gains, and requires git since every experiment is a commit that gets kept or reset. Best for iterative optimization where you have a clear metric. Completely wrong for one-off fixes or anything you can't measure numerically.

skills.sh installs: 3.4K
GitHub stars: 41K
last updated: Aug 21, 2026

View on GitHub

Install

$ npx -y skills add github/awesome-copilot --skill autoresearch --agent claude-code

Installs into .claude/skills of the current project.

Files

Autoresearch: Autonomous Iterative Experimentation

An autonomous experimentation loop for any programming task. You define the goal and how to measure it; the agent iterates autonomously -- modifying code, running experiments, measuring results, and keeping or discarding changes -- until interrupted.

This skill is inspired by Karpathy's autoresearch, generalized from ML training to any programming task with a measurable outcome.