An agent logically has all the capabilities necessary to do what the agent should be able to do. The sum of capabilities of an agent indicates "the worst that can happen" if the agent is malicious. It makes sense that if a network service is malicious, all networking activities can be subverted. Still, the storage activities shouldn't be subverted, and of course the network service wouldn't have the storage service capability. However, if a user is malicious, anything could go wrong that the user is normally trusted to not make go wrong. Correspondingly, the user must have an expansive sum of capabilities.

Capabilities are themselves simple, but that is the mechanism perspective. Access control policy is an entirely different beast, and any mechanism at best minimizes the risks.

However, were were talking about "sudo" and this is purely interactive tool (I am sure some people run sudo non-interactively, but we can all agree it's a bad security practice). So, how does this nice capabilities vision mesh with the idea of "administrator user"?

From what I see, it's not much different. Instead of "sudo" which verifies user identity+permissions and then gives out "root" or some other user, we have some other hypothetical tool ("become-admin"?) which verifies user identity+permission and gives out powerful system-changing capabilities. Frankly, I don't see much difference - it's the same security model either way.

(A related argument I've seen is "you should not need sudo for common tasks like update software or configure network" - but on modern Linuxes, you don't need sudo for that already, this is done via dbus + polkit calls. No "proper microkernels" required, that is already deployed everywhere.)