This project aims to develop and evaluate a Multi-Agent Reinforcement Learning (MARL) framework integrated with an ontology layer to optimize simulated steel-processing operations, including rolling, quenching, and levelling. A fast surrogate environment will be used to safely train and test agents, with the ontology providing domain knowledge for improved coordination, safety, and reward shaping. Results will help assess the benefits of knowledge-guided MARL in industrial control, using synthetic data only.