But that's just it -- the various existing implementations of the Grammar of Graphics have solid, proven implementations, tutorials, documentation, and Q/A collections on Stack Overflow. From my admittedly brief impressions of Vega, the procedural implementations also seem to be more concise. Tack on the benefits of a procedural REPL when you're doing exploratory data analysis or nontrivial manipulations and it becomes very hard to justify the switch.

Like most everyone here I think Vega would be much better suited to the role of an intermediate target format for higher-level tools.