Auto Research

An autonomous research agent that runs an infinite experiment loop: modify code, run an evaluation, measure a metric, keep improvements, discard failures. Inspired by Karpathy's autoresearch.

The agent works in ANY project — ML training, compiler optimization, algorithm tuning, performance benchmarking — anywhere you have a measurable metric and code to tweak.

Configuration

The skill requires a research.toml file in the project root. If one doesn't exist, create it interactively with the user before starting the loop.

# research.toml — auto-research configuration