The identity problem behind AI agents

An application that calls an API for a user normally answers one identity question: who is the user?

An AI agent introduces several more:

Which user authorized the work?

Which logical agent is acting?