# Autoresearch (Karpathy)

> Karpathy's loop where an agent edits training, measures, and keeps or discards

- Category: Agents & dev (`agents`)
- Tags: `open-source`, `free`
- Official: https://github.com/karpathy/autoresearch
- Wunderlust: https://wunderlust.xyz/r/autoresearch
- Image: https://wunderlust.xyz/media/autoresearch-poster.jpg
- Video preview: https://wunderlust.xyz/media/autoresearch.webm
- Added: 2026-11-05

## What it is

Andrej Karpathy's repository in which an AI agent edits a model's training code, trains for 5 minutes, compares the validation result and keeps or discards the change.

## What it’s for

A conceptual basis for improving something iteratively and measurably, such as skills, with one reversible change at a time.

## Benefits

- Three files: a fixed prepare.py, a train.py edited by the agent and a program.md with the instructions
- A fixed budget of 5 minutes per experiment, which makes results comparable (about 12 per hour)
- A single metric, val_bpb, where lower is better

## Caveats

- MIT license; the original requires an NVIDIA GPU (tested on an H100) and Python 3.10+
- The video uses a Claude Autoresearch plugin, not an official product; the exact repository wasn't legible in the narration
- No automatic improvement is guaranteed; the example's metrics are proxies

## Similar on Wunderlust

- [Claude Autoresearch (plugin)](https://wunderlust.xyz/r/claude-autoresearch-plugin.md): Plugin that improves a skill in measured cycles, with a metric and Git rollback
- [Unsloth Studio](https://wunderlust.xyz/r/unsloth-studio.md): App for running, chatting with and training AI models locally
- [OpenViking](https://wunderlust.xyz/r/openviking.md): Context database for agents that organizes memory and skills like a file system
- [MLX](https://wunderlust.xyz/r/mlx.md): Stack for running local models on Macs
- [Jev Ultrafast](https://wunderlust.xyz/r/jev-ultrafast.md): A Python browser agent that picks indexed actions over the DOM
- [Personal Context Portfolio](https://wunderlust.xyz/r/personal-context-portfolio.md): Ten markdown files of personal context for AI agents, with a guided interview