Supervised agentic AI for precision oncology

Abstract Agentic artificial intelligence could transform precision oncology, particularly in molecularly heterogeneous diseases like breast cancer, by integrating multimodal data and rapidly evolving evidence, but probabilistic models clash with the deterministic requirements of cancer care. We describe a Supervisory Reference Architecture that separates specialized worker agents from an oversight meta-agent, enforces traceable reasoning, and enables auditability and regulatory accountability. This framework reframes evaluation around reasoning traces and supervisory interventions. We present it as a conceptual reference architecture rather than a validated system: no implementation, simulation, or empirical evaluation is reported.