The problem
How it works
He packaged autoresearch, in which a coding agent edits a small GPT training script, runs five-minute experiments and keeps or reverts each change, into a container built on Red Hat AI base images. Claude Code drove the loop on an OpenShift cluster with H100 GPUs (A100s were also tried) during a 24-hour unsupervised run.
Results and limitations
He reports 198 experiments, 29 kept, and a 2.3% validation-loss improvement, with the agent going more than 20 hours in a single session. Its findings included that smaller batches and a wider MLP beat deeper networks, while weight tying and label smoothing made results much worse.
Before you try it
Vendor post about an early-access platform version, based on one run with no variance reported. A CUDA driver mismatch needed a library-path fix, and the exact Claude model version is not clearly stated.
Tools and repositories
- Tools
- autoresearch · Claude Code · Red Hat OpenShift AI · NVIDIA H100
- Prerequisites
- Needs GPU access and a coding-agent subscription or API key that can run unattended for many hours.
Try it yourself: Run a bounded autonomous experiment loop
Fill the fields to adapt the prompt. Empty fields stay as visible [placeholders].
You are running an experiment loop on [Training script] in [Repository path]. The only metric is [Metric to optimize] as reported by [Command that runs and scores an experiment]; lower is better. For each experiment, state a one-line hypothesis, make one small change, run it with a fixed budget of [Time budget per experiment], record the result in [Results log], and keep the change only if the metric improves; otherwise revert it with git. Never edit the evaluation code, the data loader or the metric definition. Stop after [Experiment cap] experiments or after three crashes in a row, then summarize what helped and what hurt. Ask before installing packages or using more than [Compute limit].Lock the metric and evaluation code so the agent cannot game them, and confirm kept changes on a longer run before trusting them.
Related videos
About anthropics/claude-code
The SECRET Claude Code Upgrade That Has Everyone Going Nuts (Full Guide) (opens in a new tab)
The title names the project, the description links the project website.
Watch on YouTube (opens in a new tab)About anthropics/claude-code
The only Claude Code guide you'll ever need (Opus 4.5) (opens in a new tab)
The title names the project, the description links the project website.
Watch on YouTube (opens in a new tab)Independent creators. Listing a video implies no endorsement, sponsorship or partnership.
Sources
- Running Karpathy's autoresearch on Red Hat OpenShift AI: 198 experiments, zero intervention (opens in a new tab)
Vendor post · developers.redhat.com · Published 7 Apr 2026 · Source checked 9 Nov 2026, 22:19 UTC
What this source supports
- autoresearch containerized on OpenShift AI with H100 GPUs
- Claude Code drove the experiment loop
- 198 experiments, 29 kept, 2.3% validation-loss improvement over 24 hours
- CUDA driver mismatch fixed with an LD_LIBRARY_PATH change