What is an Agent Permissions?

Permissions sit one level more granular than authorization scope: a scope might describe the broad area an agent operates in (customer support, say), while permissions are the specific allow/deny rules within it — can this agent read this table, call this API, write to this field. In agent systems, permissions need to be checked at the moment of tool execution, not just assumed from the agent's role, because a single agent session might call dozens of different tools and each call is a separate point where the wrong permission could let something unintended through. Permission models for agents typically separate three concerns cleanly: what the agent's own service identity is allowed to do, what the delegated user's account is allowed to do, and what the specific task's declared scope restricts it to — with the tightest of the three winning. Collapsing these into one flag is a common source of over-permissioned agents.