autoresearch
AI agents running research on single-GPU nanochat training automatically
Autoresearch is an agent skill that enables autonomous AI-driven experimentation on language model training by iteratively modifying code, running short training cycles, and evaluating results. Designed for single-GPU nanochat training, it allows researchers to set up overnight autonomous experiments where an AI agent edits training code, tests modifications in fixed 5-minute windows, and logs progression toward better validation metrics. Developers configure agent behavior through a Markdown instruction file while the agent optimizes model architecture, hyperparameters, and training loops without manual intervention.
Key Features
Use Cases
- 01Running unattended neural network hyperparameter optimization experiments overnight
- 02Exploring model architecture variations through autonomous agent-driven iteration
- 03Discovering training optimizations within fixed compute budgets
- 04Learning AI-assisted research workflows for language model development
- 05Benchmarking different GPT architectures under consistent time constraints
- 06Prototyping autonomous research organization structures through instruction file tuning
Related Skills
View moreawesome
😎 Awesome lists about all kinds of interesting topics
superpowers
An agentic skills framework & software development methodology that works.
skills
Skills for Real Engineers. Straight from my .agents directory.
andrej-karpathy-skills
A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.
autoresearch — FAQ
What is autoresearch and what does it do?+
Autoresearch is an agent skill that runs autonomous AI research experiments on language model training. It gives an AI agent control over training code modifications, runs quick 5-minute training cycles, evaluates results, and iterates automatically to find performance improvements.
How do I install and set up autoresearch?+
Install the uv package manager, then run 'uv sync' to install dependencies and 'uv run prepare.py' to download training data and train the tokenizer. After setup, point your AI agent (Claude, Codex, etc.) to the program.md file in the repository to begin autonomous experiments.
Which AI clients work with autoresearch?+
Autoresearch works with coding-capable AI agents like Claude (including Claude Code), Codex, Cursor, and similar clients that can read instructions from Markdown files and edit Python code. You run the agent in the repository directory with appropriate permissions disabled.
What are the hardware and software requirements?+
The default implementation requires a single NVIDIA GPU (tested on H100), Python 3.10 or higher, and the uv package manager. Community forks exist for MacOS, Windows RTX, and AMD platforms with modified requirements.
Is autoresearch free and open source?+
Yes, autoresearch is released under the MIT license and is completely free to use and modify. The code is open source and available on GitHub.
Do I need API keys to use autoresearch?+
You need access to an AI coding agent (like Claude or Codex) which may require API keys or subscriptions depending on the service. The autoresearch skill itself does not require additional API keys beyond what your chosen AI agent needs.
How do I install autoresearch?+
Open the source repository on GitHub and follow its README. autoresearch is a skill — MCP Agents Market links you directly to the official repo.
Is autoresearch free?+
autoresearch is an open-source project hosted on GitHub. Check the repository for its license and any usage requirements.