BrowseFull catalogOutcomesSolve a specific problemRolesStack by teamTrustFilter by risk tier
← Back to the Claude Observatory

AutoResearch (Karpathy)

Workflow Data & Analytics Usable
Works inClaude Code Claude Cowork Claude Chat
Usable Reviewed

Built by Andrej Karpathy (OpenAI founding team, former Tesla AI director) — a strong credibility signal.

Autonomous ML research loop from Andrej Karpathy: give an AI agent a small but real LLM training setup (a simplified single-GPU version of nanochat) and let it experiment overnight. The agent edits train.py — architecture, optimizer, hyperparameters — trains for a fixed 5-minute budget, keeps changes that improve validation bits-per-byte, and repeats. You steer the research org by editing program.md, not the Python files.

93,783 starsFreeDeveloper setup
Usable rating — This tool is functional but has notable gaps. Review the evaluation notes below before deploying.

Built by Andrej Karpathy (OpenAI founding team, former Tesla AI director) — a strong credibility signal. It is a minimal working template for agent-driven ML experimentation: the human programs the research org in Markdown while the agent iterates the training code against an objective metric.

ML researchers and practitioners with an NVIDIA GPU who want to experiment with agent-driven training-loop optimization and autonomous overnight experimentation.

Claude Code Claude Cowork Claude Chat

https://github.com/karpathy/autoresearch

By Andrej Karpathy

How to Get It

This is a methodology or approach. Paste the instructions below into a Claude conversation to get started.

Command
git clone https://github.com/karpathy/autoresearch && cd autoresearch && uv sync && uv run prepare.py  # one-time data prep (~2 min); test with: uv run train.py; then point your agent at program.md

Tip: Paste this into a Claude Code conversation. Verify command matches your Claude Code version.

First thing to try

After installing, paste this into Claude:

Help me run overnight autonomous experiments where an agent iterates on a small GPT training loop
PrerequisitesA single NVIDIA GPU (tested on H100); Python 3.10+; uv package managerCostFree

Trust Signals Reviewed

Stars93,783Contributors9Last updated2026-03-26Known CVEsNone foundSources: GitHub Advisory Database + OSV.dev · Scanned 2026-08-13 · scanner v1

Community Pulse Active

Discussed on Hacker News, Reddit

3 mentions across 2 sources

Reviewer notes

Reviewed review. These are observations, not a security certification.

catalog_hygiene stale-eval refresh: Scored from trust signals (evidence-eval-v1): 90,169 GitHub stars; 9 contributors; last commit 108d ago; license no license.

2026-05-10: Strong credibility signal from the author, but 2 production mentions against 71k stars tells you this is more reference architecture than battle-tested tooling—the community is watching, not shipping. In practice it's useful for prototyping multi-step research pipelines: web search, paper pull, synthesis in sequence without manual handoffs. The gap to close before team deployment is observability and failure handling; those patterns aren't mature here yet. If you need something in production now, LangChain's research agent templates or AutoGPT variants have more operational mileage. Worth a close read if you're designing your own agent workflow from scratch.

Things to check

  • Recently discovered. Looks promising but has limited community feedback so far. We added it early so you can evaluate it before the crowd finds it.

How to evaluate tools before deploying →

Data shown here comes from public APIs and automated scanning. Reviewer notes reflect one person's experience. This is not a security certification or legal recommendation. Always evaluate tools according to your own organization's policies.

Evaluation

Ease of Use
2/5
Versatility
5/5
Reliability
4/5
Security
3/5
Overall score3.40 / 5.00 UsableEvaluatedJul 2026
catalog_hygiene stale-eval refresh: Scored from trust signals (evidence-eval-v1): 90,169 GitHub stars; 9 contributors; last commit 108d ago; license no license.

← Back to the Claude Observatory

Rolling Claude out in your org? Let's talk.

Start a conversation →