Everyday human information processing requires the use and interaction of many cognitive centers of the brain. In different areas of the brain, information of different modalities is processed and understood, which are interconnected by neural connections to achieve multimodal integration. The paper proposes a multi-agent neurocognitive approach to the problems of integrating multimodal data obtained from various afferent sources of an intelligent agent. This approach is based on the computational abstraction of multi-agent neurocognitive systems that model the architectural correspondences of neural connections in the human brain, which allows us to develop a model that can independently learn, recognize and understand data flows using existing knowledge, context and experience. The architecture of an intelligent agent based on the multi-agent neurocognitive approach includes a sensor module that receives data from several modalities and transmits them to the cognitive centers of attention, situation modeling, goal setting and control, since the use and interaction of these systems is required for full integration and understanding of incoming information.