Overview
AutoResearch is Andrej Karpathy's minimal autonomous ML research loop: an AI coding agent edits a compact training repo, runs short experiments, compares a validation metric, and keeps only improvements.
Overview
AutoResearch is an open-source experiment by Andrej Karpathy for autonomous machine learning research. The core idea is deliberately small: give an AI coding agent a real but compact language-model training setup, let it edit the training code, run a fixed short experiment, evaluate whether the result improved, and keep or discard the change.
The official repository describes the project as a way to let an AI agent experiment overnight on a simplified single-GPU implementation of nanochat. Humans are not meant to directly tune the Python loop in the ordinary way. Instead, the human programs the research process through program.md, which gives the agent its operating instructions and research direction.
See also Loop Architecture, Running AutoResearch, and Limitations and Implications.
Why it matters
AutoResearch is interesting less because it is a finished research platform and more because it compresses an ML research workflow into a tight agent loop:
- A human defines the objective and constraints.
- The agent proposes code changes.
- The system runs the experiment under a fixed time budget.
- A metric decides whether the change survives.
- Git history becomes the experiment log.
This makes the project a concrete example of agentic research: not just asking an LLM for ideas, but giving it a measurable environment where ideas can be tried, scored, and ratcheted forward.
Sources
- Official repository: https://github.com/karpathy/autoresearch
- Official README: https://github.com/karpathy/autoresearch/blob/master/README.md
- DataCamp explainer, March 23, 2026: https://www.datacamp.com/tutorial/guide-to-autoresearch
- VentureBeat coverage, March 9, 2026: https://venturebeat.com/technology/andrej-karpathys-new-open-source-autoresearch-lets-you-run-hundreds-of-ai