looking a bit at a protocol buffer actionscript implementation and google's python implementation -- it's interesting to note that they both first generate native descriptions of the classes; then they provide generic code that walks those descriptions to de/serialize buffers.
probably no one gave it much thought ( likely the c implementation is the same ) but i wonder why they went that route.
alternatively: they could have just generated the de/serialization code directly. for python and actionscript that could be especially interesting, because you wouldn't even need the class header. the serialize could just build the class by setting attributes dynamically. i'm not convinced that a standalone description would be smaller than the generated code, and i'd bet that generated code would be much faster.
interesting side note: looks like there are two projects for actionscript on google code: the one started by an adobe evangelist has no code; the other seems to work and was the inspiration for this post; is that an indication that adobe needs be more aware of what else is going on out in the open source world?