Initially, the direction we want is molecules that do stuff under some kind of external control. Lucent's DNA tweezers from five years ago are a good example. The fact that they're tweezers tells you that they are doing something: they don't merely have material properties, they have actual mechanical behaviors. The external control comes in the form of "fuel strands" and "anti-fuel strands" which, when added to the test tube, make the tweezers close and open respectively.
The tweezers were cool but they obviously didn't usher in a brave new world of full-blown nanotechnology. So it's got to be small, but it's also got to be capable. And because we want to tell it what to do, it needs to be controllable from the human scale.
Chris Schafmeister at the University of Pittsburgh is considerably further along. His lab has figured out a Lego set of little molecules that snap together rigidly into any shape. Schafmeister can control the assembly sequence and therefore control what shape he gets.
Peixuan Guo at Purdue is doing some similarly interesting stuff with RNA. This work has already resulted in more specific delivery systems for chemotherapeutic drugs used to treat cancer.
Another promising direction is to harness ribosomes, the widgets inside our cells that produce proteins. We can specify any DNA sequence we want, so we can instruct ribosomes to build any of a wide variety of novel human-designed proteins. We are already doing some of that; what we need to do more is to understand proteins better. Some of that is science, but a lot of it is just simulation horsepower.