The design and implementation of these programs is about as different as possible, because the refinement requires tighter integration between Phenix and Rosetta.  phenix.rosetta_refine is using a generic API for distributing similar jobs across a number of supported clusters, which makes it much easier for us to add support for queuing systems in the individual programs in Phenix, and easier for users to invoke this support, but also requires specific code to handle the different software.  So it is slightly less flexible than what MR-Rosetta (or AutoBuild, or LigandFit) is doing internally.

-Nat