Agentic Digital Twins for Precision Oncology: A Survey of Multimodal Foundation Models and Causal Reasoning for Personalized Cancer Care

Cancer treatment is increasingly guided by patient-specific molecular, radiological, and clinical evidence, yet most decision-support tools remain static, single-modality, and correlational. Three technological trends are converging to change this picture: (a) digital twins, which maintain continuously updated virtual replicas of a patient's tumor and physiology; (b) multimodal foundation models, which learn shared representations across imaging, pathology, genomics, and text; and (c) causal inference, which moves prediction beyond association toward estimating what would happen under an untried treatment. A fourth, more recent trend agentic artificial intelligence, in which large language model (LLM)-based agents plan, retrieve evidence, invoke tools, and act with limited autonomy offers a mechanism for orchestrating these components into a coherent clinical workflow. This paper surveys the state of the art in each of these four areas and synthesizes them into a conceptual framework for an agentic digital twin for precision oncology. We review recent literature on cancer digital twins, multimodal oncology foundation models, causal machine learning for treatment effect estimation, and agentic AI in clinical medicine, and we outline the architectural layers, applications, open challenges, and research directions relevant to building such systems responsibly.