In the rapidly evolving field of wireless communication, Multiple Input Multiple Output (MIMO) networks have emerged as a pivotal technology, offering enhanced data rates and spectral efficiency by leveraging multiple antennas at both the transmitter and receiver. The introduction of Open Radio Access Network (O-RAN) architecture has further revolutionized this domain, enabling greater flexibility, scalability, and interoperability through its open interfaces and software-defined approach. This paper presents DRAGON, a novel Deep Reinforcement Learning (DRL)-based framework for joint Layer and Modulation and Coding Scheme (MCS) selection, tailored for downlink single-user MIMO networks under the O-RAN framework. Our approach is designed to be highly scalable, capable of efficiently managing a large number of configuration options in one-shot prediction, including up to 25 MCS and 4 layer choices. The proposed solution has been rigorously evaluated using an O-RAN-based simulation environment, demonstrating up to an 18% performance improvement over the state-of-the-art (SOTA) methods and achieving a best throughput of 87.4% when compared to the collected ground-truth dataset. Furthermore, our method supports real-time prediction, making it viable for practical deployment. In addition to these advancements, we explore the potential integration of our DRL-based solution with real-world platforms and discuss the extension of our approach to handle multi-user (MU) MIMO scenarios, paving the way for broader applications in next-generation wireless networks.