AutoResearch GitHub Guide: What It Is and Where to Start

What Is AutoResearch?

AutoResearch is an open-source experiment in autonomous ML research. It gives an AI coding agent a compact training loop and asks it to improve the setup by editing code, running short experiments, and checking whether the result got better.

The repo is small on purpose. That makes it easier to inspect than a large research stack and easier for agents to work with. It also means the files that matter most are easier to understand for people who are just trying to decide whether the project is worth digging into.

Where Is the Official GitHub Repo?

The official source is github.com/karpathy/autoresearch. This site is not the official home of the project, and it should be used as a guide rather than a replacement for the repository itself.

Which Files Matter Most?

• prepare.py for data prep, tokenizer setup, and utilities
• train.py for the main training logic the agent is expected to change
• program.md for the human-written instructions that guide the agent

A beginner can get value by reading these in order because they explain both the mechanics of the run and the boundary between what the human specifies and what the agent is allowed to change.

How to Read the Repo Quickly

The fastest path is to scan the README for the overall loop, inspect program.md to understand the intended behavior, and then open train.py to see what the agent can modify. After that, browsing issues and forks can help you understand what other users are trying to run and where the setup starts to bend.