On the Regulatory Potential of User Interfaces for AI Agent Governance