Autoresearch

An open-source agentic framework that automates the machine learning research cycle through a recursive self-improvement loop.

Autoresearch is a minimalist, 630-line Python framework developed by Andrej Karpathy to delegate the scientific method to AI agents. The system operates on a Recursive Research Loop (RRL) where an agent reads a research strategy, modifies training code, executes experiments, and analyzes results to improve a target metric (such as validation loss). By keeping the entire codebase small enough to fit within an LLM context window, it eliminates structural hallucinations and allows for high-velocity experimentation. In one benchmark, the tool completed 126 experiments overnight on a single GPU, achieving an 11% efficiency gain on an already optimized nanoGPT model by identifying architectural oversights that human researchers missed.

https://github.com/karpathy/autoresearch