autoresearch

Autonomous agent-driven optimization loop inspired by Karpathy's autoresearch. Sets up and runs an iterative hill-climbing harness where subagents modify an artifact, evaluate against a single scalar metric, and keep improvements. Use this skill whenever the user wants to "optimize something iteratively", "run an autoresearch loop", "hill-climb on performance", "auto-optimize", "iterate and improve automatically", "run experiments autonomously", "autonomous optimization", or mentions "autoresearch" in any context. Also triggers when the user describes a workflow like "try variations and measure which is best", "keep tweaking until it's faster", "optimize this config", "find the best prompt", "tune hyperparameters", "benchmark variations", or any scenario where they want an agent to autonomously explore a search space against a measurable objective. Works with any domain — code performance, prompt engineering, config tuning, SQL optimization, CSS optimization, model training, build flags, or anything with a measurable outcome.