A new framework, CARGO, is proposed to evaluate agentic AI in production by treating retrieved references as procedural exemplars and grounding factual judgments in the live instance's observed context. CARGO eliminates false penalties in reference-based evaluation and retains high contradiction recall.