Top 200 ReposEXPLORER
Browse agents, categories and pages
Back to results
Reported by sourceThe source says this happened. Top 200 Repos has not reproduced it.

Let a coding agent run overnight ML training experiments

Point a coding agent at autoresearch so it edits a small GPT training script, runs short timed experiments and keeps what improves.

Try the promptRead the original (opens in a new tab)

The problem

How it works

autoresearch fixes data preparation in one file and lets the agent edit only `train.py`, guided by instructions you write in `program.md`. Each experiment trains for a fixed five-minute budget and is judged on validation bits per byte; the agent keeps or discards each change and logs the results for you to review later.

Results and limitations

The README estimates about a dozen experiments per hour, or roughly a hundred overnight, on a single NVIDIA GPU.

Before you try it

It targets NVIDIA GPUs (tested on H100), results are only comparable on the same hardware, and the README runs the agent without permission prompts, so isolate the machine. GPU and agent costs add up. Results depend on your setup and were not independently reproduced by Top 200 Repos.

Tools and repositories

Tools
NVIDIA GPU · Python 3.10+ · uv · Claude Code or Codex
Prerequisites
A single NVIDIA GPU and a coding agent subscription or API key.

Try it yourself: Write a program.md for an overnight run

Fill the fields to adapt the prompt. Empty fields stay as visible [placeholders].

Preview
Help me write program.md for an autoresearch run on [GPU]. The research direction is [Research direction]. Rules for the agent: edit only train.py, never prepare.py; run each experiment within the fixed time budget; record every change, its val_bpb and a one-line rationale in [Log file]; revert any change that does not improve the metric. Stop after [Experiment cap] experiments or if three runs crash in a row. Do not install packages, access the network beyond what prepare.py already did, or touch files outside the repository. Show me the draft before I start the run.

Run unattended agents only on an isolated machine or container with no access to secrets.

Related videos

About anthropics/claude-code

About openai/codex

About anthropics/claude-code

Independent creators. Listing a video implies no endorsement, sponsorship or partnership.

Sources

  1. karpathy/autoresearch README (opens in a new tab)

    Project documentation · github.com · Undated · Source checked 9 Nov 2026, 22:25 UTC

    What this source supports
    • Agent edits only train.py, guided by program.md
    • Fixed five-minute training budget per experiment
    • val_bpb metric, lower is better
    • Roughly 12 experiments per hour, about 100 overnight
    • Single NVIDIA GPU; tested on H100
Report a problem

Reported by source

Answer technical questions by letting async agents run experiments

Simon Willison sends research questions to cloud coding agents such as Claude Code for web and Codex Cloud, which write and run experiments in a dedicated repo and open a PR.

simonwillison.net · 6 Nov 2025anthropics/claude-code, openai/codex
Reported by source

Run unattended ML training experiments with autoresearch

Red Hat's Roy Belio containerized Karpathy's autoresearch on OpenShift AI with H100 GPUs and let Claude Code run 198 training experiments over 24 hours.

developers.redhat.com · 7 Apr 2026karpathy/autoresearch, anthropics/claude-code
Planned, not done

Build a formula-driven Excel model with the Claude xlsx skill

Install Anthropic's xlsx skill in Claude Code so the agent writes live formulas, recalculates the workbook with LibreOffice and checks it for formula errors.

github.comanthropics/skills, anthropics/claude-code
Reported by source

Build an on-call investigation agent from your own alert history

Goutham Veeramachaneni connected Claude Code to Grafana's gcx CLI, his code and runbooks, then trained its skill on two weeks of past pages for his services.

gouthamve.dev · 21 Apr 2026anthropics/claude-code
Reported by source

Get a new OCR model running on unfamiliar GPU hardware with Claude Code in a sandbox

Simon Willison gave Claude Code full permissions inside a Docker container on an NVIDIA Spark and had it get DeepSeek-OCR working; it took four prompts and under 40 minutes, with one nudge from him.

simonwillison.net · 20 Oct 2025anthropics/claude-code, pytorch/pytorch, huggingface/transformers
Reported by source

Port a Python library to JavaScript with a coding agent

Simon Willison had Codex CLI translate the JustHTML parser into a dependency-free JavaScript library, steering it with about eight prompts over one afternoon.

simonwillison.net · 15 Dec 2025openai/codex

More with Claude Code

Reported by source

Review pull requests automatically with Claude in GitHub Actions

Install the Claude Code GitHub Action so Claude reviews each pull request with inline comments and responds to @claude requests.

code.claude.comanthropics/claude-code
Reported by source

Trace a disk-filling production bug from core dumps with an agent

B12 CTO Adam Marcus used Claude Code to analyze gigabyte core dumps filling web server disks and traced them to an RSS endpoint reading an endless radio stream.

blog.marcua.net · 28 Jan 2026anthropics/claude-code
Reported by source

Give Claude your Google Calendar and Gmail through a self-run MCP server

The open-source Google Workspace MCP server exposes Calendar, Gmail and other Workspace tools to Claude Desktop, Claude Code and other MCP clients, using your own OAuth client and optional read-only mode.

github.com

View all: Claude Code