An agentic ML engineer that delegates training to Claude

What it does Karpathy is a thin orchestration layer that spins up a local sandbox and launches the Google ADK web interface. The agent inside can tap a library of 130+ Scientific Agent Skills to run data workflows and train models, dropping all outputs into an isolated directory. It is essentially glue code demonstrating how a standard LLM agent can behave like a domain-specific machine-learning engineer.

The interesting bit The project treats agentic ML as a skill-composition problem rather than a model-architecture problem. By offloading scientific workflows to reusable, upstream skills, it keeps the local codebase small and turns a chat session into a reproducible experiment.

Key highlights

• Built on the Claude Agent SDK and Google ADK; exposes a browser-based UI at localhost:8000
• Integrates 130+ reusable Scientific Agent Skills for research, analysis, and ML tasks
• Sandboxed execution keeps dependencies, code, and artifacts in one directory
• Compatible with Cursor, Claude Code, Codex, and Google ADK
• Explicitly unaffiliated with Andrej Karpathy; the name is a tribute

Caveats

• Requires Python 3.13, the uv package manager, Claude Code authentication, and an OpenRouter API key, so the setup is picky
• Described by its own authors as a “simple implementation” that leans heavily on the upstream Scientific Agent Skills repo
• Datasets and custom scripts must be manually copied into the sandbox before the agent can see them

Verdict Useful if you are prototyping agentic research workflows or evaluating the Agent Skills standard. Look elsewhere if you want a batteries-included AutoML suite—this is scaffolding with a clear upgrade path to K-Dense’s commercial platform.