AI agents fail differently from traditional software. A normal stack trace can tell you where code crashed, but not why an agent selected a tool, skipped a guardrail, retried the wrong step, or ...