Cancer remains a leading cause of morbidity and mortality. Global Cancer Observatory (GLOBOCAN) 2022 estimated that there were nearly 20 million new cancer cases and 9.7 million cancer-related deaths worldwide in that year [1]. Treatment delays are associated with higher mortality across several cancer indications [2]. Timely diagnostic pathways are therefore important, but they must also account for genomic, histopathological, microenvironmental, and temporal heterogeneity [3].

Modern oncology diagnosis is intrinsically multimodal. Clinicians combine imaging, pathology, laboratory results, molecular assays, and longitudinal electronic records, often across separate specialties and information systems. This fragmentation creates cognitive and coordination burdens that motivate integrated workflows [4,5]. Task-specific AI has advanced lesion detection, pathology classification, molecular modeling, and risk prediction [4,6-8]. However, most deployed or evaluated models address a single bounded input-output task rather than the full diagnostic process.

Large language models (LLMs), multimodal foundation models, and tool-use frameworks create a technical basis for coordinating these tasks [9,10]. However, a model that accepts several modalities is not automatically an agent. Similarly, retrieval-augmented generation (RAG) adds external evidence to generation, and fixed workflow orchestration links predetermined modules, but neither property alone establishes runtime planning or feedback-driven control.

This Viewpoint examines AI agents as auditable orchestration layers for multimodal oncology diagnostics. We organized the field around 4 connected stages: data collection, preprocessing, multimodal fusion and representation learning, and diagnostic decision support. Our central position is deliberately bounded: agentic orchestration may improve coordination and auditability, but clinical benefit and safety remain deployment-specific questions.

We used an operational definition of agentic behavior. In a conventional fixed pipeline, an output is generated through a predetermined composition, y=f_n(...f₁(x)), and the sequence of functions does not change at runtime. An AI agent instead maintains a state s_t and selects an action a_t = π(g, s_t, A_t) from an allowed tool set A_t. For a clinical goal g, it observes each tool’s result and updates its state before choosing the next action. The practical distinction therefore lies in conditional planning, tool selection, observation, and revision within a governed loop.

Four adjacent concepts should remain separate. A multimodal foundation model supplies reusable representations or generation across data types. RAG retrieves external sources to ground an answer. Fixed workflow orchestration executes a prescribed sequence of services. A feedback-driven AI agent can change the sequence, request missing evidence, retry or switch tools, abstain, and escalate to a clinician according to the evolving state. Long-term memory is optional and should be limited to an authorized, versioned, patient-specific or institutional state rather than treated as an unrestricted capability.

First, Figure 1 contrasts a conventional specialist synthesis with an agent-coordinated 4-stage workflow. Second, Figure 2 operationalizes the agentic layer as a plan-act-observe-update-check loop. Act coordinates the 4 stages in Figure 1B; update records state and provenance; and check evaluates evidence sufficiency, cross-modal consistency, and whether further action remains within the authorized scope. If these criteria are met, the synthesis proceeds to clinician review. Remediable gaps return the workflow to planning; gaps that cannot be supplemented trigger clinician escalation. The agent-coordinated elements in both figures are prospective design propositions and not validated clinical systems or performance claims.

We applied 3 evidence labels throughout this Viewpoint. Agent-level evidence evaluates a complete loop that plans, uses tools, and produces an integrated output, as in the external oncology evaluation by Ferber et al [11]. Component or infrastructure evidence evaluates a model, retrieval method, data pipeline, or governance mechanism that an agent could invoke [12]. Prospective propositions describe plausible clinical workflows that still require system-level and prospective validation. This hierarchy prevents strong component performance from being misread as proof of a clinically validated agent.

The first stage is the governed assembly of patient-specific evidence. Relevant data may be distributed across a hospital information system (HIS), laboratory information system (LIS), picture archiving and communication system (PACS), electronic medical record (EMR), pathology archives, and genomic platforms. These systems differ in format, ontology, temporal granularity, and access control.

Component-level infrastructure studies show that integration must be engineered. A single-center oncology data supply chain linked clinical, genomic, and imaging data for more than 170,000 patients across 11 cancer types and more than 800 features per case [13]. Its extract-transform-load (ETL), natural language processing (NLP), and quality control procedures support the feasibility of a data layer but not the validation of an agent. Standards such as Health Level Seven Fast Healthcare Interoperability Resources (HL7 FHIR) and Digital Imaging and Communications in Medicine (DICOM) can reduce site-specific translation [14]. Local mapping and governance nevertheless remain necessary.

An agent could use secure interfaces to identify, retrieve, and temporally align data for a specific diagnostic question. It could connect a new scan to prior pathology, biomarker trajectories, sequencing results, and notes [9,15]. Unstructured text and patient-reported or wearable data may add context in selected settings [16]. Each input should carry its source, acquisition and update times, missingness status, and freshness threshold. Contradictory or stale records should trigger clarification rather than silent fusion. Once evidence is assembled, the next risk is whether every modality is technically fit for analysis.

Multimodal preprocessing converts heterogeneous inputs into representations that downstream tools can use. Imaging varies by scanner and acquisition protocol. Digital pathology depends on tissue preparation, staining, and slide digitization [6]. Laboratory and genomic data may be incomplete or generated on different platforms, while clinical text often contains duplication and institution-specific shorthand.

A fixed workflow applies the same preprocessing sequence. An agentic workflow could select modality-specific tools according to file type, quality, and clinical purpose. It might invoke imaging quality control and segmentation or pathology stain normalization, tiling, tissue filtering, and feature extraction. Pathology foundation models provide reusable component representations [17,18], while self-supervised medical imaging models illustrate robustness and data-efficiency strategies [19]. Attribution tools such as gradient-weighted class activation mapping can support region-level inspection, although an attribution map is not itself a clinical explanation [20].

Every action should be recorded with the input source, time stamp, parser or model version, quality control result, and review status. If optical character recognition or parsing fails, the system should not fabricate a normalized record. It should retry within a bounded policy, switch to a validated alternate parser, return a partial result with an explicit missingness flag, or request human review. This provenance-aware preprocessing creates the conditions for fusion without concealing upstream uncertainty.

Fusion is not the mere presence of multiple data types. Imaging describes macroscopic phenotype, pathology reveals tissue architecture, molecular assays characterize biological drivers, and longitudinal records provide clinical context. Multimodal models can combine these views for characterization and risk estimation [7], but many assume that all modalities are harmonized and simultaneously available.

At the molecular level, machine learning can integrate complementary omics layers and prioritize regulatory interactions. Omidi combined RNA sequencing and microRNA sequencing profiles from 2 public cohorts [21]. The analysis included 78 adrenocortical carcinoma samples from The Cancer Genome Atlas adrenocortical carcinoma (TCGA-ACC) and 250 normal adrenal samples from the Genotype-Tissue Expression (GTEx) project, reconstructing condition-specific competing endogenous RNA networks. Random forest models performed best for predicting microRNA-messenger RNA associations, and predicted interactions were cross-referenced with TargetScan and miRTarBase. This study is an in silico systems biology module that an agent could invoke. Batch effects, cohort imbalance, cross-cohort heterogeneity, and the absence of independent or experimental validation prevent these findings from being interpreted as validation of a multimodal agent.

An agent could select a fusion strategy according to the clinical question, available modalities, and uncertainty. If pathology and imaging are available while genomic testing is pending, it could use a modality-robust model, mark molecular evidence as unavailable, and revise the synthesis later. Explainable multimodal real-world models have identified patient-level marker contributions in large cohorts and undergone independent validation [22]. These results support component-level representation and explanation but not autonomous clinical orchestration. Fusion outputs must therefore pass an evidence-adequacy check before decision support. The evidence-adequacy gate and its return-or-escalate logic are shown in Figure 2.

The final stage translates analyzed evidence into a structured, reviewable diagnostic summary. The purpose is not autonomous diagnosis or treatment selection. It is to state the supported findings, unresolved conflicts, missing evidence, uncertainty, and provenance in a form that clinicians can audit.

A recent feasibility study evaluated Gemini 2.5 Pro (Alphabet Inc) and ChatGPT 4o (OpenAI) on 80 Turkish-language prostate-specific membrane antigen positron emission tomography–computed tomography (PSMA PET-CT) reports [23]. The prompts incorporated the American Joint Committee on Cancer (AJCC) staging, the Chemohormonal Therapy Versus Androgen Ablation Randomized Trial for Extensive Disease (CHAARTED) criteria, and few-shot examples. The models assigned T, N, and M categories and disease volume classes, with overall task accuracies of 93.8% and 91.3%, respectively. Errors clustered around equivocal findings and different interpretive thresholds, including overstaging of ambiguous lesions. The single-center design, 1 expert reference standard, and exploratory sample of 80 reports limit generalizability. The study supports an LLM-based report-to-staging component that requires expert oversight; it does not validate a complete AI agent.

Other component studies illustrate evaluable end points. Multimodal oncology chatbots have been tested on image-containing clinical cases [24]. Explanation types can alter physician trust and diagnostic performance [25]. On the contrary, fabricated details can induce hallucinated clinical elaboration despite simple mitigation attempts [26]. Agent evaluation should therefore include tool selection, evidence citation, contradiction handling, safe refusal, calibration, workflow burden, and final conclusion quality rather than fluency alone.

In practice, an agent might prepare a multidisciplinary team (MDT) case by linking a suspicious lung lesion to pathology, molecular markers, smoking history, and prior imaging. It could rank competing diagnoses, identify a missing confirmatory test, and provide an evidence-linked summary. Proposed MDT assistance and randomized source-attribution studies support case preparation and verifiable retrieval as useful components [27,28]. Clinicians must retain authority over the final judgment, with explicit responsibility assigned across users, institutions, and manufacturers [29].