NVIDIA Introduces ACES for Agentic Skill Evaluation in Enterprises

Enterprise AI agent programs are moving from prototypes to production, requiring robust evaluation of reusable skills and tools. Traditional review processes often focus on structure and security, but fail to assess whether a capability package actually helps a live agent complete real-world tasks within a given environment. To address this, NVIDIA has developed ACES (Agentic Continuous Evaluation of Skills), a repository-native framework. ACES conducts paired live trials, comparing agent performance with and without a specific skill. It normalizes trajectories, grades six runtime metrics, and reports "Skill Lift," which quantifies the added value of a skill for a fixed task, harness, and scoring policy. Testing ACES on 145 real enterprise skills showed that structural scans provide complementary but distinct insights compared to LLM-judged evaluations. Across 947 scored cases, ACES found a mean composite Skill Lift of 0.2134, with positive lift in 72.8% of cases. The framework revealed significant gains in skill execution, behavior checks, and efficiency, offering insights into discovery, routing, and tool use that static scans cannot capture. An open-source implementation is available.