Flexible manufacturing systems require rapid adaptation to process changes. However, in industrial environments, BPMN modeling relies heavily on manual design, which is both inefficient and costly. To address these challenges, this paper presents H2A-BPMN, a dual-agent hierarchical reasoning framework to support the automatic BPMN modeling from natural language process descriptions. First, a lightweight intermediate representation, iBPMN, is designed. It preserves BPMN semantics to reduce the reasoning load of LLMs. Second, the analysis agent analyzes user requirements into rules template. It then generates an analysis report to guide the modeling task. Third, the modeling agent adopts the H-CoT to achieve the modeling task in a hierarchical manner. In each layer, H-CoT supports dynamic interaction with the environment. It includes knowledge-base retrieval and invocation of the checking tool to execute the progressive reasoning “Generate-Check-Correct”. Finally, experiments conducted on Camunda-BPMN and TPD-BPMN demonstrate that our method outperforms the baseline approaches in multi-role scenarios. It has achieved a 13–37% improvement in similarity evaluation, along with higher scores in understandability, correctness, completeness, and role clarity.