Abstract: To address the inadequate decision reliability of large language models in oncology multidisciplinary team (MDT) consultations caused by the lack of collaboration mechanisms and evidence-based quality control, a multi-agent framework named OncMDT with dynamic orchestration and reflection coordination is proposed. The framework dynamically instantiates a subset of specialist agents based on patient clinical features and constructs a multi-stage collaborative reasoning workflow mapping the actual clinical agenda. An iterative retrieval strategy with planning and verification feedback is introduced to achieve role-aware guideline evidence addressing. Before consensus generation, a reflection and coordination mechanism is designed to structurally review and correct inter-disciplinary conflicts, evidence deviations, and missing intervention dimensions, combined with a case memory bank utilizing dual-layer manifold retrieval for empirical reference. Evaluated on a benchmark dataset comprising 70 cases with varying complexity across four cancer types, OncMDT achieves a decision consistency rate of 85.7%, a completeness rate of 88.3%, and a citation accuracy of 82.6%, significantly outperforming baseline methods. The system effectively replicates the multidisciplinary collaboration logic, enhancing the evidence-based rigor and clinical safety of complex oncology decisions.