You wouldn't do that for everything. I'd reserve it for work with higher uncertainty, where you're not sure which path is best. Different model families can make very different choices.