ATLAS: Agentic Taxonomy of Large-Scale Software Ecosystems

ATLAS is a novel framework for automatically constructing a hierarchical taxonomy of software repositories on GitHub, addressing the limitations of the existing flat organizational structure. It integrates a Designer Agent for proposing splitting dimensions and a Classifier Agent for project classification, utilizing a self-corrective loop to refine dimensions based on classification failures. Evaluated on 54,387 repositories, ATLAS achieves a Taxonomy Quality F-score of 83.13%, significantly outperforming existing methods, and enhances repository discovery with a precision at rank one (P@1) of 85.71%, providing critical insights into ecosystem trends that are obscured by traditional tagging methods.