We present tools and techniques for the design of an interactive motion data structure to drive virtual avatars. Our approach automatically analyzes a traditional labeled motion graph and constructs a condensed, coarser graph referred to as a Behavior Finite State Machine (BFSM). Each node of the BFSM represents a single behavior, and each weighted edge represents a valid transition and its likelihood. We present a semi-automated mechanism for user-editing of the graph. We also introduce an interface for user-in-the-loop design of transitions and blend spaces to generate controllable avatars from a BFSM. We present the results of our approach using two dynamic interactive examples, locomotion and martial arts, and discuss the strengths and weaknesses of our technique.