You built AI agents to be autonomous. That was the point. But autonomy drifts, and in a multi-agent system it drifts at the group level, where every agent still passes its own check and nothing flags a problem. At OpenAI it ran that way for two months. Multi-agent AI governance is a system-level job: reduce the conditions that start drift, put a human only where being wrong is expensive and watch for organization itself appearing between your agents. Most monitoring today wat