The intention is to completely re-implement translation, though (1) the hope is to accomplish it…
Paul Christiano

To put it another way, a human translator has learned a lot of valuable information, much of it inaccessible to introspection, because they’re stored as synapse weights in the middle layers of a giant neural network. I don’t see how you take advantage of such information (i.e., have it enter your computation in a useful way) except by activating the giant neural network with the kinds of inputs that it was trained on, i.e., phrases, sentences, or even larger units of natural language. Using such an agent as meta-executioner is unlikely to be competitive with using them directly, just from an information theoretic perspective.