The Future Internet promises to pervade our everyday lives by interweaving an increased variety of access technologies (cellular broadband, wireless hotspots, short range radios). Cognitive network management is a promising approach to cope with such access diversity and to enable dynamic adaptation of the network configurations and parameters. This article reveals the insights and conclusions of a unique real life implementation of a cognitive architecture, comprising software agents and artificial intelligence algorithms, and its deployment within a heterogeneous access network composed of a Broadband WiMAX base station and WiFi access points. We discuss the architectural challenges that this approach poses, and present the distributed software architecture we implemented. We detail the experimental access network we deployed and the algorithms that enable channel reselection and vertical assisted handover. We conclude with our findings and recommendations for the deployment of cognitive network management architectures and technologies in the Future Internet.