This is such a great and insightful piece of feedback that I now feel inspired to write something up on the Axum blog to elaborate on our thoughts on these topics. I was hoping to not work this weekend, but it will be hard to keep myself from writing something. Please look for something later, possibly before Monday morning.
I suggest installing the bits and trying some of the samples, which are still meant to be educational rather than realistic.
The messaging overhead is comparable to Erlang, while the memory pressure per agent is close, but not yet close enough to make me happy. Compared threads, the cost per agent is about 0.001 of what a thread requires, as long as you place receives in asynchronous methods (see the WebFetcher sample).
And, unlike some other technoligies that may have disappointed you in the past, we're actively looking for your input. That's why we're getting Axum out there before it's completely baked.