The Karpathy Loop — AI Agents Running Autonomous Training Experiments

Narrative

In March 2026, Andrej Karpathy—independent researcher and former OpenAI/Tesla executive—released autoresearch, catalyzing mainstream coverage of what the AI community terms 'the Karpathy loop': an autonomous agent that iteratively reads training code, proposes modifications (learning rate, architecture depth, hyperparameters), executes time-boxed experiments (typically 5 minutes on GPU), evaluates outcomes, and repeats. Karpathy's proof-of-concept ran 700 experiments in 2 days on a small language model, discovering 20 optimizations that yielded ~11% training speedup when applied to larger models. Critically, Karpathy disclosed a personal inflection point: since December 2025, he has stopped writing code directly, instead directing AI agents for 16 hours daily—a shift he frames as entry into a 'loopy era' where autonomous systems conduct research without humans in the loop. The concept proved immediately portable: Shopify CEO Tobias Lütke ran autoresearch overnight on internal data and achieved 19% performance gains from 37 automated experiments. Fortune, The New Stack, Latent Space, and NextBigFuture all positioned autoresearch as foundational to how frontier ML labs will conduct research, though concrete adoption announcements from official labs (OpenAI, Google DeepMind, Anthropic, Meta) have not yet appeared in public coverage.