How to stop AI agents going rogue

Agentic AI is taking decisions and acting on behalf of users, but how to stop that going wrong?