Fred Wilson (VC/blogger extraordinaire) wrote up an interesting analysis of his own investment career viewed from the perspective of startup nimbleness. His main conclusion is that the majority of startups that succeed in achieving superb returns at some point in their history "transform" by being flexible enough to experiment. I love pieces like this because they help to pop the myth of entrepreneur-as-visionary who just seems to know at every decision what needs to be done.
To Fred's point, all of the best entrepreneurs I've seen combine two key abilities: a set of lightweight planning capabilities (to set up measures for success and gates that determine failure), and an unrelenting ability to execute quickly over repeated iterations without the rest of the company getting completely demoralized. In the software space in particular, another key skill is being able to build enough platform assets that can be re-used across experiments. Even in the fast and cheap world of web development, this is a key part of being able to iterate quickly and with as little technical debt as possible, which is why getting some hackers with experience on board can mean the difference between Facebook (iterated well) and Friendster (crapped out).
Finally, the article is also interesting because it applies well outside the context of the traditional startup. I'd say that about 50% of what I've been focused on since the HP acquisition has been on setting up the right execution context for maintaining this nimble experimental capability. Some of that involves figuring out how to work around big corporate process that exist for very good reasons in other contexts, and some of it involves cultural changes that reset the definition of success on the part of key contributors. But most of it— the really challenging bit— involves managing the impedance mismatch between what a big company needs to be successful (predictable success), and what a little startup relies on to make a dent in the world (huge risk-taking and nimble execution that may or may not work).