Karpathy Wants Your Agent to Run Its Own Experiments, and Anthropic Just Loosened the Safety Rails

Andrej Karpathy released autoresearch on March 8 — a minimalist, open-source Python tool (~630 lines) designed for AI agents to autonomously run machine learning experiments on a single GPU. The setup is deliberate: a human writes a research prompt in Markdown, and the AI agent iterates on the training code in Python. Each iteration is a complete LLM training run lasting exactly five minutes. The agent works in an autonomous loop on a git feature branch, accumulating commits as it finds lower validation loss through changes to architecture, optimizer, and hyperparameters.

The design philosophy is instructive. Rather than building a complex research orchestration system, Karpathy stripped nanochat's LLM training core to its minimal form and exposed it as an agent loop. The goal is not to replace researchers but to create a legible benchmark for comparing agent research productivity — different prompts, different agents, different strategies, all measured on the same task. Early community adoption suggests the framework's value lies less in the specific experiments it runs and more in the paradigm it normalizes: treating research iteration as an agent-native workflow.