This paper describes a design method of a robot system which consists on mechanically linked plural agents, so as to realize autonomous robots which can adaptively behave in the real world. A difficulty in designing distributed autonomous systems, is how to embed a dynamics for self organizing environment-oriented-action-rules in the systems. So, we propose a robot system which controlled by a decision making method that has oscillator and learning method based on same object functions, and which has a structure of mechanically linked agents that are identically designed, in order that each agent actively interacts with other agents and the outer world. For verifying an usefulness of the system, behavioral acquiring tests in target approaching and obstacle avoidance were implemented by using Distributed Autonomous Swimming Robot that was constructed by the proposed way. Moreover, for learning efficiency in complex tasks such as obstacle avoidance, we proposed Switching Q-learning in which previously obtained action rules were effectively used As a result, the robot acquired simple obstacle avoidance behavior, and time-sequential connections between acquired action rules were observed by interaction-based learning in a distributed autonomous system. That is, it was verified that the proposed system design method was one of the solutions of adaptive systems to complex environments.