Abstract: Traditional agent cooperation model emphasizes the high degree of autonomy of agent, and its spontaneous collaborative process is completely out of internal “selfish” motivation. No macro guidance, free from external control, it is difficult to meet the reliability requirements of current software system. The computational complexity of the formal theory and no effective mechanism for conflict resolution are also obstacles for agent technology to be applied in the real software systems. In this paper, using organization and policy metaphor, we build an organized and policy-oriented agent collaborative model. Through the organization and policies, macro-guidance and external control are given to the agents to enhance system dependability. While using extended defeasible logical framework, we establish a formal theory for this collaborative process which has non-monotonic reasoning ability and linear time computational complexity. The built-in priority manner of this logical framework provides an effective mechanism for conflict resolution. Based on this model, a policy-oriented cooperation process in agent organization is formally described with four cooperation stages including organizational state updating, organizational goals creating, agent obligation distributing and agent action executing. And then we analyze computational complexity and prove the consistency of the model, as well as some other simple and interesting nature. Finally, the main characteristics of this model are verified by an example.