WecoAI/awesome-autoresearch

Implements an agent-driven optimization loop where LLMs (Claude, Codex) iteratively edit a target file, execute it against a metric within a fixed time window, and commit or revert changes based on improvement—originally designed for LLM training but generalized across domains (GPU kernels, template engines, tabular ML, prompt engineering). Each entry includes full optimization traces showing the agent's exploration trajectory, not just final results. Covers implementations for Apple Silicon (MLX), Windows (RTX), distributed multi-agent swarms, and framework integrations (Claude Skills, Pi extensions).

460 stars.