Can We Make AI Agents Coordinate Safely?