The experimentation in system's programming still exists, it just takes a different form (possibly because the cost of committing a design to code is a bit more). Jeff Dean says that the hallmark of a good systems programmer is the ability to mentally simulate the effects of a number of different designs so that you can quickly throw out ones that won't work. You don't sit in a corner trying to perfect your design, you go out and talk to many potential users of your system to make sure you understand its requirements, then you come up with many possible designs that might fulfill those requirements, then you prune that down to one design that you actually implement.
I've been mulling this a lot lately. I'm going to try to remember this when I get too deep into 'designing' a solution.