My understanding of Kay's OOP model is that it's akin to each object running as a separate process. Each process can only see its own memory, but RPC enables objects to "message" each other.
This message passing is more like Objective C's messaging than RPC, in the sense that any message can go to any object. But there's no guarantee the receiver will know what to do with that message.
In this model, objects are isolated from each other. If they're isolated, their code can be run in parallel without explicit thread safety. And objects can replace each other at runtime as long as they respond to the same messages. Well, technically even if they don't.