Creating complex and realistic group behaviors can be a difficult and time consuming task. Automated approaches for motion generation typically involve explicitly defining a set of possible agent behaviors, associating appropriate behaviors with all environmental events, and setting the priorities among various behaviors in every possible situation. Generally, such approaches are pre-tuned to particular situations and are difficult to adapt for other scenarios or for different sets of behaviors. In this paper, we investigate methods to facilitate the generation of complex group behaviors for applications such as games, virtual reality, robotics and biological/ecological simulation. Our general approach is to provide a framework that automatically combines simple composable behaviors into more complex behaviors. Adaptation to new environments and specialization for new tasks or new agent abilities is supported by a ``learning'' process through which agents select their current behavior based on their prior experiences. We illustrate how our framework can be applied to pursuit/evasion and laser tag scenarios.