The AI Daily Brief: Artificial Intelligence News and Analysis · Nathaniel Whittemore

Autoresearch, Agent Loops and the Future of Work

March 9, 2026·26 min·2 clips

Andrej Karpathy's Auto Research hands the entire machine learning research loop to an AI agent, running experiments every five minutes.

The AI Daily Brief, hosted by Nathaniel Whittemore, dedicates this episode entirely to Andrej Karpathy’s Auto Research project, released on a Saturday as a weekend GitHub repository. Karpathy—OpenAI founding team member, former Tesla AI director, and the person who coined “vibe coding” in February 2024 and more recently declared the era of “agentic engineering”—published a minimal self-contained system of approximately 630 lines of code that trains a small language model entirely autonomously. The system runs in a loop: the human writes a high-level prompt file specifying the research goal; an AI agent then iteratively modifies the training code on a Git feature branch, running five-minute training experiments, reading the results, committing improvements, and continuing indefinitely without further human intervention. Each dot in the accompanying visualization represents a complete LLM training run. Karpathy framed the project with deliberately unsettling science-fiction flavor, writing as though from a future where frontier AI research is conducted by autonomous swarms across compute megastructures, with the repo itself described as the story of how it all began. Whittemore uses Auto Research to expand on his earlier discussion of a software development system called Ralph (named after Simpsons character Ralph Wiggum) that similarly runs persistent, iterative loops to build software. Together, he argues, these systems point toward a new work primitive: the autonomous agentic loop, a basic building block of work so fundamental it will appear across roles and industries the way spreadsheets, email, or version control eventually did. Work primitives of this type are rare—they emerge perhaps once a decade—and their introduction tends to restructure what categories of work humans do rather than just speeding existing workflows. The episode discusses what kinds of tasks are best suited to agentic loop structures (long-horizon, well-defined objective, iterative, fast feedback loops), what tasks resist them, and how this changes the human role from executor to director of research intent. Tone is thoughtful and speculative, pitched at AI practitioners, technical founders, and anyone thinking about the future of knowledge work.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from The AI Daily Brief: Artificial Intelligence News and Analysis