Agents operating in the real world must be able to interact with other agents. Game theory provides guidelines as to how agents should behave in strategic interactions with other agents. Until recently, however, algorithms from game theory were unable to deal with large-scale games. A few years ago, I worked on a system called Gala that made it very easy to describe interesting strategic situations, and also implemented state-of-the-art game-theoretic algorithms. As it turns out, even these advanced algorithms cannot solve really large games. An example is full-scale poker: there are simply too many possible hands to deal with all of them explicitly. My colleagues and I have recently started studying approximation methods for dealing with really large games. We have developed a theory of abstractions for imperfect information games. We are also working on an implementation of a poker player using our ideas, and hope to present results in the next few months.