Tag
AgentAudit is an open, extensible framework for evaluating the full lifecycle of AI agents across capability, grounding, security, and behavioral dimensions, enabling precise failure attribution and highlighting trustworthiness differences among various language models.
The paper presents ACES, a framework for continuous evaluation of AI agent skills through live trials, measuring Skill Lift to quantify added value, and demonstrating its effectiveness on enterprise repositories compared to scan-only gates.