As we are moving towards the new era of 5G supported verticals and their services, there is a surge in the mobile network investments and equipment deployments. The traditional monolithic mobile network radio access equipment, generally called as the base station (BS), houses radio units (RU), processors and accelerators along with proprietary software. With the advances in cloud technologies, disaggregated Radio Access Network (RAN) architectures emerged in the past decade, where Network Functions (NFs) are ripped away from specialized hardware and placed on bare metal servers. Although the concept of RAN evolution has been around for some time starting from Cloud Radio Access Networks (C-RANs), continuing with Virtual Radio Access Networks (vRANs) and most recently Open Radio Access Networks (ORANs), we are seeing commercial implementations and tests of disaggregated, virtualized and open RANs only since the beginning of 2020. One unique aspect that motivates this paper is the availability of new opportunities that arise from applying machine learning to optimize the RAN in closed-loop, i.e. without human intervention, where the complexity of disaggregation and virtualization makes well-known Self-Organized Networking (SON) solutions inadequate. In particular, Multi-agent systems (MASs) is expected to play an essential role in the control and coordination of near-real-time and nonreal-time RAN Intelligent Controller (RIC). In this visionary article, we first present the state-of-the-art research in multiagent systems and team learning, then we provide an overview of the landscape in RAN disaggregation, virtualization, as well as ORAN which emphasizes the open interfaces introduced by the O-RAN Alliance. We present a case study for ML-based disaggregation, and finally, we identify challenges and open issues to provide a roadmap for researchers.