With the continuous development of urbanization, the past decades have witnessed the rapid growth of the ridesharing systems as a new business mode. Autonomous vehicles (AVs) are expected to be introduced into ridesharing systems for its advantages in central coordination. We aim to optimize the real-time operation of a ridesharing platform transiting from traditional ridesharing networks to fully automated mobility-on-demand systems, where conventional vehicles (CVs) and AVs coexist. The operator can directly control all the AVs and thus make centralized AV dispatching decisions. Human drivers focusing on their own monetary return, however, tend to gather at hot spots and lead to unbalanced supply and demand. Incorporating distinct characteristics of AVs and CVs, we propose a two-sided multi-agent reinforcement learning based framework to dynamically coordinate mixed fleets in ridesharing systems. Specifically, we model the operator and driver decision making procedure as Markov decision processes (MDP), where the operator and drivers interact with the environment at the same time. The operator makes centralized decisions on AV fleet dispatching and driver commission rates, while drivers make relocation decisions in a decentralized manner. The proposed model is validated using a case study in New York City using real taxi trip dataset. Results demonstrate that our algorithm significantly improve system performance. In particular, we also show that charging dynamic commission fee makes both the platform and the drivers better off, especially in scenarios with higher demand volume.