Recently robosoccer has proved to be one of the most successful test beds for multi-agent behavior. In robosoccer, agents take actions based on the feedback they receive from the environment. In this paper, we propose a novel concept to select high level actions in a multi-agent environment. We conduce the whole problem to that of a defense strategy of goalie as it has got more scope for behavior recognition and action selection. We first characterize the various behaviors and actions so as to suit the HMM model. We then describe how these twin HMMs (Behavior Hidden Markov Model and Action Hidden Markov Model) are used in two different phases to recognize behaviors and select actions. Finally we present an algorithm that would implement the complete defense strategy of a goalie in key game situations.