How Agents Ask for Permission: User Permissions for AI Agents, from Interfaces to Enforcement

This paper surveys 21 proposals for user-permission systems in AI agents and develops a taxonomy spanning policy specification at the interface, internal policy representation, policy derivation from user input, and runtime enforcement. It also compares five prominent commercial agents with the research literature. The authors identify recurring design themes and gaps, especially around expressing individual preferences, translating natural-language intent into enforceable policies, and ensuring that sensitive actions remain authorized at execution time.