It deals with lists and strings well, which is decent: Arc's only other sequence-like type is the table (I don't think you'd ever want to treat symbols as a sequence of 1-character symbols; you'd just use a string). Tables would work better if they were properly coerced, cf. the comment above tablist and listtab in arc.arc.
The more I think about it, the more I like this model. Conceptually, it seems that map should behave like
(def map (f seq)
(map-as (type seq) f seq))
even if it's not implemented like that -- all the coercions would surely be slow. (Tangential: map would also need to handle multiple sequences.) But it makes more sense for map and coerce to at least have compatible behavior. Plus, map's current behavior is a degenerate case of the coerce-compatible map: