MAS-SL is a multi-agent scheduling engine for prefabricated buildings that integrates a self-learning mechanism to achieve dynamic task coordination and scheduling optimisation in this study. The three parts of the engine are state awareness, self-learning scheduling and feedback correction, and they form a closed-loop strategy evolution mechanism. Conflict prediction mechanism and real-time resource regulation for stability under multi-task concurrency. Based on BIM simulations and about 1,000-1,200 valid scheduling samples, experiments have shown that MAS-SL performs stably in terms of scheduling accuracy (around 90%), average response delay (about 1.8 s), and system stability (about 86%), and is moderately better than the baseline model. Ablation studies show that removing the state encoding module causes a drop of about 4-5 per cent in scheduling accuracy, and omitting the feedback mechanism extends the response delay by approximately 2.4-2.5 s; thus, all these modules have demonstrated their functions. As shown in the above results, the engine is generalizable and practically feasible in complex construction environments.
