WecoAI/awesome-autoresearch

What it solves

AutoResearch solves the challenge of manually optimizing code or models by automating the trial-and-error process. Instead of a human developer tweaking parameters or code to improve performance, an AI agent runs a continuous loop of testing, measuring, and refining to find the best possible outcome.

How it works

At its core, AutoResearch uses a specific prompt that instructs a coding agent (like Claude Code) to follow an iterative workflow. The agent modifies a target file, runs it for a set time on a GPU, and checks the results against a specific metric. If the change improves the metric, the agent keeps it; if not, it reverts the change. This loop repeats indefinitely, allowing the agent to explore and optimize complex tasks autonomously.

Who it’s for

This project is for developers, researchers, and engineers who want to automate performance tuning, prompt engineering, or architectural search. It is particularly useful for those who want to see the full "optimization trajectory"—a record of every attempt the agent made—rather than just the final result.

Highlights

• A curated collection of real-world use cases, ranging from GPU kernel optimization and LLM training to game AI and scientific research.
• Includes links to optimization traces, allowing you to inspect exactly how the agent arrived at its improvements.
• Features a variety of open-source implementations and forks, including versions for Apple Silicon, Windows, and distributed multi-agent swarms.
• Provides access to benchmarks and evaluation tools to measure the effectiveness of research agents.