autoresearch
AI agents running research on single-GPU nanochat training automatically
Autoresearch is an agent skill that enables AI assistants to autonomously conduct machine learning research experiments on single-GPU nanochat training setups. The skill provides a structured workflow where agents iteratively modify training code, run 5-minute experiments, evaluate results, and decide whether to keep or discard changes. Developers configure the autonomous research process by editing program.md instructions while agents optimize train.py containing the GPT model, optimizer, and training loop. The skill is designed for overnight autonomous experimentation, potentially running ~100 iterations while unattended.
Key Features
Use Cases
- 01Autonomous hyperparameter tuning for GPT models overnight without manual intervention
- 02Exploring neural network architecture variations through iterative agent-driven experimentation
- 03Optimizing training loops and optimizer configurations across dozens of experiments while unattended
- 04Conducting reproducible ML research comparisons using fixed-time budgets on single-GPU setups
- 05Delegating repetitive model experimentation to AI agents while researchers focus on high-level strategy
- 06Learning neural network training dynamics by reviewing agent experiment logs and decisions
Related Skills
View moresuperpowers
An agentic skills framework & software development methodology that works.
skills
Skills for Real Engineers. Straight from my .agents directory.
skills
Public repository for Agent Skills
ponytail
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
autoresearch — FAQ
What is autoresearch and what does it do?+
Autoresearch is an agent skill that enables AI assistants like Claude or Codex to autonomously run machine learning experiments on a simplified GPT training setup. The agent iteratively modifies training code, runs 5-minute experiments, checks for improvements in validation loss, and decides whether to keep or discard changes.
How do I install and set up autoresearch?+
Install the uv package manager, clone the repository, run 'uv sync' to install dependencies, then 'uv run prepare.py' to download training data and build the tokenizer. After setup, point your AI agent (Claude, Codex, etc.) to the program.md file and prompt it to begin experiments.
Which AI clients work with autoresearch?+
Autoresearch works with any AI coding assistant that can read instructions and modify files in a repository, including Claude (via Claude Code or API), Codex, and similar agents. You should disable unnecessary permissions when running the agent in the repository.
What are the hardware and software prerequisites?+
You need a single NVIDIA GPU (tested on H100), Python 3.10 or later, and the uv package manager. The default configuration requires substantial GPU memory, though community forks exist for MacOS, Windows RTX, and AMD platforms with smaller memory requirements.
Is autoresearch free to use?+
Yes, autoresearch is released under the MIT license and is free to use. However, you'll need access to an AI coding assistant (which may have its own costs) and appropriate GPU hardware to run the experiments.
How long does each experiment take?+
Each training experiment runs for exactly 5 minutes of wall-clock time (excluding startup and compilation), allowing approximately 12 experiments per hour or around 100 experiments during an 8-hour overnight run.
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.