It is also a testament to solid engineering and attention to good security practices in general. These still work, also against fancy new AI attackers.
When sophisticated attacks become cheaper to run, maybe it will (finally) be cheaper to do more solid engineering instead of doing it quick and dirty and ending up in indefinite bug-squashing mode.
How does this change the point that is being made in the article? Your agent is also only taking one of the existing roles that humans today occupy (e.g. the software operator or developer)
Varlink is based on much more conventional UNIX technology than Dbus, which is decades old: You connect to a named UNIX socket through its socket file in the filesystem (man page: unix(7)).
This is an old mechanism and it is known to work well. It does not require a broker service, it works right at system startup, and it does not require a working user database for permission checks (which would be a circular dependency for systemd in some configurations). If at all, I am surprised that systemd didn't use that earlier.
The main thing that Varlink standardizes on top of that is a JSON-based serialization format for a series of request/response pairs. But that seems like a lightweight addition.
It also does not require kernel support to work, the kernel support is already there. He mentioned in the talk that he'd like to be able to "tag" UNIX sockets that speak varlink as such, with kernel support. But that is not a prerequisite to use this at all. The service discovery -- and he said that in the talk as well -- is simply done by listing socket files in the file system, and by having a convention for where they are created.
Etoile had its own Smalltalk dialect back in the day, Pragmatic Smalltalk. This was a Smalltalk based on the Objective-C runtime, based on an OMeta implementation and a LLVM backend. David Chisnall, who created it at the time, ended up getting involved more in LLVM in the long run, I believe.