Tangible agent (TA) is a new medium that communicates the senses of sight, hearing, touch, smell, and taste of human to computer. Intelligent behavior is a key property to realize the agent because it must interact with human to grasp all the information. We have adopted a behavior-based approach for high-level behaviors such as navigation of office, conversation with human, and cleaning of room. Behavior-based method can control unexpected situation without prior knowledge and generate high-level behavior with behavior selection. However, TA requires improvements of behavior selection architecture for better communication with human. In this paper, we propose an intelligent behavior selection architecture that contains interferencing, learning and planning capability to TA. In this paper, overview of technical details and experimental results on physical device (Khepera robot) are presented. Preliminary results show the possibility of the proposed behavior selection for TA.