Confidence in agentic AI: Why eval infrastructure must come first