We are already internally tinkering with a 'genetic algorithm' type approach using the best structures generated, crudely similar to what you describe. We have not yet come up with a good way to parallellize this yet though, as the important things to consider are that we don't want the client to have to contact the server to get work, and we don't want it to have to download updated, refined maps of conformational space every few hours, which is unfeasible for most people.
We will hopefully come up with something in the next month or two that will work well in parallel and improve on the current brute force approach.