The development of control unit functions is a complex process, often leading to high costs and less effective solutions. Reinforcement Learning (RL) offers a promising approach to autonomously train agents for complex control tasks with minimal human involvement. However, its application is typically confined to simulated environments due to the high costs of data generation and safety considerations. This research highlights the use of RL in function development, particularly focusing on accelerating RL agent training by combining Transfer Learning (TL) and X-in-the-Loop (XiL) simulation. In this study, targeting transient Exhaust Gas Re-circulation (EGR) control in internal combustion engines, Model-in-the-Loop (MiL) simulations are initially employed for hyperparameter tuning, and initial training of RL agents. These agents are then fine-tuned on Hardware-in-the-Loop (HiL) systems based on real hardware interactions through TL. Comparatively, agents trained with TL and XiL simulations showed training time reduction compared to those trained exclusively on HiL, for the same performance. The study underscores the effectiveness of training RL agents on real hardware, demonstrating the effectiveness of integrating TL and XiL simulations in RL training.