How should permissions work when AI agents start delegating tasks to each other?

’ve been thinking about a problem that seems to get more complicated as AI workflows become more autonomous. A lot of discussions around agent security focus on things like prompt injection and whether an agent can be manipulated into doing something it shouldn’t. But even if we put that aside, there’s another question I find interesting: how much access should an agent actually have in the first place? An agent might be able to interact with files, browse websites, call tools, or pass work to another agent. But does having access to something mean it should be able to use it at every step of a workflow? The delegation part seems especially interesting to me. If Agent A asks Agent B to handle a task, should Agent B automatically get the same permissions as Agent A? Or should the delegated task come with a much narrower set of permissions? I’m starting to think that agent permissions may need to be more granular than traditional user permissions, especially once workflows involve multiple agents. For those building or experimenting with agentic workflows, how are you thinking about permissions and delegation? Do you prefer broad permissions with safeguards, strict least-privilege access, or something in between?