NVIDIA’s experimental SkillEvaluator tool provides a three-stage method for evaluating the safety, specification compliance, overlap, and task performance of AI agent skills. Developers can run it from the command line, integrate it into continuous integration workflows, and add SkillEvaluator skills to Claude Code or Codex. The final stage compares agent performance with and without skills enabled in a Harbor sandbox, and the project is open source under the Apache License 2.0.
Source:Global Economic Intelligence · Read original
Leave a Reply