Barbara Tillett noted back in 2003 that in FRBR, the "work" and "expression" entities may collapse into an "abstraction" and I think Bibframe is doing this.
Quote: "Using this model, one could even collapse both 'work' and 'expression' into something called an 'abstraction' when that was found to be useful for a particular application. It should be noted however, that FRBR kept them separate." http://www.loc.gov/catdir/cpso/frbreng.pdf
My conclusion is, it will be up to the Bibframe implementor to add more concrete entities that can distinguish a bf:Work for better contextualization. Maybe by using FRBR entities, maybe by other. It's for sure a set of bf:Work must be post-processed for many aspects, not only for search by author/title clustering, but also by building groups for text, sound, image, language etc. like we know from FRBR.
Another consequence is that Bibframe packagers will have to be a little cautious because there will be many properties for creating and linking bf:Work classes, since they will subsume the FRBR expression level. Some may package all properties in one bf:Work, while others may create multiple bf:Work's and link them.