How to test AI agent output guardrails means building a stack, not a single check: input filters, output validators, an eval ...