In arid and semi-arid regions like Tensift Al Haouz in central Morocco, optimizing irrigation strategies is critical due to increasing water scarcity and the high costs of field experimentation. Crop simulation models such as AquaCrop have proven valuable for evaluating water use and crop productivity, particularly for winter wheat. In this study, we develop Deep Reinforcement Learning (DRL) agents using the Proximal Policy Optimization (PPO) algorithm to learn profit-oriented irrigation policies, trained entirely within calibrated AquaCrop environments. The model is configured using local crop, soil, and weather data collected from the Tensift Al Haouz region during the 2002–2004 growing seasons and further calibrated with field measurements from nearby test sites. This simulation-based methodology enables the training of adaptive irrigation strategies without the logistical and financial constraints of real-world trials. Preliminary results show encouraging learning progress in both models, where the agents’ performance is comparable to that of experienced human irrigators. The integration of DRL with biophysical crop models demonstrates a promising path toward scalable, data-driven irrigation management in water-limited contexts.