I was also surprised by the file abstraction, since the article starts off with protobuf-based messages. I thought (excitedly) it would be messages all the way down. Protobufs have some nice backward-compatible properties as well.

time() is an interesting problem... but are there lots of those? My mind keeps going back to I/O and other operations where the messaging overhead is probably reasonable... or maybe even better than "reasonable" because it might replace similar existing and necessary abstractions. Time feels like it is in some odd space between computation itself and the system interface. What else lives there?

This would probably fall apart in practice, but if there were a small number of these cases one could imagine catching the message invocation and turning it into something more static. Though I realize this also opens up the question of "who am I talking to?" which seems like one of the major necessary building blocks of a message-based OS.