Precision oncology: from large language models to multi-agent systems

With the rapid growth of electronic health records, medical imaging, and high-throughput omics data, precision oncology faces increasing demands for cross-modal information integration and complex clinical decision support. In recent years, large language models (LLMs) and their multimodal extensions have opened new technological avenues for addressing these challenges and have shown considerable promise across a range of applications. This review provides a structured narrative overview of the current applications of these technologies across the precision oncology care continuum, encompassing key stages such as cancer screening, diagnosis, staging, treatment recommendation, and clinical documentation. However, single-model approaches still have clear limitations in constructing complex clinical reasoning pathways, ensuring the traceability and verifiability of decision processes, and integrating deeply with established clinical workflows. Against this backdrop, AI agents with autonomous planning and coordination capabilities, together with multi-agent systems (MAS), have emerged as an important new direction in precision oncology research. Building on this development, we further propose an analytical framework centered on task-architecture alignment, emphasizing that foundation models, single-agent systems, and multi-agent systems should be selected according to the complexity and risk profile of the clinical task. Such a framework may provide a useful basis for the design, evaluation, and clinical translation of AI systems in precision oncology.