SKILLFOUNDRY: Building Self-Evolving Agent Skill Libraries from Heterogeneous Scientific Resources

Source: arXiv Authors: Shuaike Shen, Wenduo Cheng, Mingqian Ma, Alistair Turcan, Martin Jinye Zhang, Jian Ma Date: 2026-04-05 Primary category: cs.AI All categories: cs.AI

Abstract

SkillFoundry turns repositories, APIs, scripts, notebooks, documentation, databases, and papers into validated agent skill packages with provenance and tests. It organizes a domain knowledge tree, mines high-value branches, compiles operational contracts into skills, and then expands, repairs, merges, or prunes the library in a closed loop. This is one of the strongest examples of external procedural artifacts becoming the actual learning surface.