Almost all of our reasoning is *defeasible*:  that is, our inferences go through only other things equal, and there are always more of them -- the list of things that might go wrong is open-ended, and doesn't run out.   Most work on defeasibility (or non-monotonic reasoning, if you're in the AI world) is focused on how to *represent* it, but I want to take a step back, and ask why it's *there*. I will argue that defeasible inference is a hard-to-avoid design feature of certain kinds of boundedly rational agents, that the open-endedness is genuine, and that we need to understand defeasibility from an engineering -- rather than a formal -- perspective.