I'm not sure how it's implemented in Graph, but the best thing you can do is to abstract out an execution to a general object.

From there, you can create all sorts of adapters and the "graph" engine just farms out the execution to the appropriate adapter. This could be anything from a job daemon on a batch farm, or, say, a typed ThreadPoolExecutor on a machine (which can either execute native clojure/java, or any sort of other script language.