Cancer Diagnostics in the Age of Agentic AI: Diagnosing the Diagnosticians

Agentic Diagnostic Systems (ADS) are emerging as a new paradigm in computational oncology, in which collaborating autonomous agents deliver structured diagnostic reasoning. However, their rapid deployment has outpaced rigorous evaluation. Most pipelines are still assessed on synthetic or simulated data and under reporting standards that were not designed for multi-agent architectures. This creates a reproducibility gap, with hallucinations, error propagation, fragile drug-drug interaction reasoning, and poor robustness under real-world distribution shifts, raising substantive clinical and regulatory risk. We argue that agentic AI for cancer diagnostics must be rebuilt around reproducibility, transparency, and governance. Existing standards (TRIPOD -AI, DECIDE-AI, CONSORT-/SPIRIT-AI, and STARD-AI) are necessary but incomplete, as they primarily target single models rather than agentic workflows. We propose a layered ADS design in which (i) a Governance Layer enforces interpretability-first reasoning, auditability, and protocol adherence, and (ii) a Federated Validation Checklist mandates multi-site evaluation on real patient cohorts, overseen by certified clinical bodies under privacy-preserving data use agreements. Our reference pipeline comprises agents specialised in general practice triage, urology, radiology, and pathology, coordinated by a tumour-board agent that synthesises outputs into a multidisciplinary diagnostic recommendation. This framework recasts ADS from loosely coupled agents into regulated, auditable systems that prioritise patient safety. By embedding governance and federated validation into their core, we outline a path to clinically trustworthy ADS that are reproducible, interpretable, and deployable across institutions, and caution that without such standards, agentic cancer diagnostics risk devolving into unsafe black-box infrastructure.