The point of the actor model, at least in OTP, is not to have the actor model as a unit of computation (as was imagined by Hewitt, et, al) but as a failure domain. Erlang was not designed to be an actor system, it was designed to be a robust system and they later discovered that what they had built was more or less the actor model (if an impure version) and it was easy to communicate the architecture's computational capabilities and mental model as actors. Actually I think if anything it's led to a fair share of misunderstandings, and truly misses the important and unique aspects of Erlang processes.

In the sense of being failure domains, OTP "actors", or processes, ARE composable. With very little boilerplate, that is built in as primitive BIFs in the standard library, I can bundle failure domains together and orchestrate them using links, monitors, and supervision strategies, with the ultimate failure domain being a single erlang VM operating inside a cluster of erlang VMs, all supported without third party libraries as a part of the way of doing things.