Abstract: A multi-agent cooperation mechanism based on dynamic infinite game model was proposed in a multi-agent platform. The infinitely repeated game structure was formed with multi-stage inviting and evaluating actions. Meanwhile， the trust benchmark was proposed to control agent blindness in selection phase and to make sure it could do the rational stage decision. Through the feedback， the multi-agent cooperation mechanism could not only dynamically adjust the income function of each game-agent， but also control cooperation priority. Also， it could achieve a closed-loop control. Additionally， multi-agent were assembled as a basic component unit of intelligent structure. Nine grids experiment was used to validate cooperation mechanism between multi-agent under multiple loading conditions. Moreover， the collaboration status of the agent components unit was also tested based on different trust benchmark. Experiments show that， trust benchmark can effectively adjust the trust level between agents， and promote cooperation in the high frequency agent system.

Key words: multi-agent, cooperation mechanism, dynamic game, infinity, trust benchmark