A note on building reliability infrastructure for AI agents and why post-incident debugging matters more than pre-flight validation.