The multimodal featuring fusion for natural human-computer interaction involves complex intelligent architectures facing unexpected errors and mistakes made by users. These architectures should react to events that occur simultaneously with eventual redundancy from different input media. In this paper, intelligent agent based generic architectures for multimedia multimodal dialog protocols are proposed. Global agents are decomposed into their relevant components. Each element is modeled separately using timed Colored Petri networks. The elementary models are then linked together to obtain the full architecture. The generic components of the application are then monitored by a based agent expert system to perform dynamic changes in reconfiguration, adaptation and evolution at the architectural level. For validation purpose, the proposed multi-agent architectures and its dynamic reconfiguration are respectively applied on practical examples.