The caveat working that way is that you can end up working towards a local optimum, and lose sight of the forest for the trees when it comes to making the bigger algorithmic changes.