SkillTrace: New Framework for Auditing LLM-Agent Skill Reuse

A recent study has unveiled SKILLTRACE, a framework for multi-trace provenance auditing aimed at identifying the reuse of skills within LLM-agent ecosystems. As these ecosystems expand, reusable skills—comprising metadata, natural-language instructions, code, tools, references, and workflows—have emerged as valuable marketplace items. The authors contend that monitoring the reuse of these skills differs from typical code clone detection, as current tools primarily target single-modality source code or overall package similarity, which may overlook instances where only one component of a skill is reused. SKILLTRACE identifies three provenance traces: Expression, Implementation, and Operational. The Operational Trace is depicted as a Skill Operational Graph (SOG), detailing activation, procedure, and resource-flow structure, with an LLM aiding solely in the extraction of the Operational trace. The paper can be found on arXiv with the identifier 2608.05204.