Chapter 2: Multi-Agent Systems in Oncology |
A Plain-Language Summary |
This chapter explores a new generation of artificial intelligence systems called multi-agent AI, which are transforming how cancer care is planned and delivered. Unlike earlier AI tools that tried to answer complex medical questions with a single model, multi-agent systems divide the work among specialized digital assistants that collaborate, cross-check each other's findings, and ground their reasoning in clinical guidelines. The chapter begins with a detailed look at TrustedMDT, a pioneering system developed at the University of Oxford that supports cancer tumor boards, then surveys how similar approaches are being applied across drug discovery, clinical trial analysis, pathology, radiology, and insurance claims processing. Each application illustrates a common theme: breaking a hard problem into smaller, verifiable tasks handled by focused AI agents produces more reliable and transparent results than asking one AI to do everything at once. |
|
1. Introduction: Why One AI Model Is Not Enough for Cancer Care |
Cancer is not a single disease. It is hundreds of diseases, each with its own biology, staging criteria, treatment guidelines, and patterns of response to therapy. A patient's cancer journey involves radiologists, pathologists, surgeons, medical oncologists, radiation oncologists, nurses, genetic counselors, and sometimes more specialists, all contributing fragments of a larger picture. |
For decades, the standard way to assemble this picture has been the multidisciplinary tumor board meeting. In these meetings, experts from different specialties gather to review a patient's diagnostic results and agree on a treatment plan. In the United Kingdom, these meetings are called Multidisciplinary Team meetings, or MDTs, and they represent the gold standard for cancer treatment planning . |
MDTs work well, but they are under enormous strain. Rising cancer caseloads mean that expert teams often have less than two minutes of discussion time per patient. A review by Cancer Research UK found that critical information gaps lead to postponements in seven percent of cases. These constraints cause treatment delays, missed opportunities to enroll patients in clinical trials, and significant clinician burnout . |
In this environment, the appeal of artificial intelligence is obvious. If AI could summarize patient records, stage cancers accurately, and draft evidence-based treatment recommendations before or during the MDT, clinicians could spend their limited time on the hardest decisions rather than on information gathering. |
But here is the problem: early attempts to use AI for cancer care relied on single-model chatbots. A single large language model, no matter how sophisticated, struggles with the high-stakes complexity of oncology. It may hallucinate facts, miss critical details from a patient's history, or apply guidelines inconsistently. The risk of what researchers call 'guessing' is simply too high when treatment decisions affect survival . |
This is where multi-agent systems enter the picture. Instead of asking one AI to do everything, a multi-agent system divides the work among specialized agents, each with its own area of expertise, its own data sources, and its own tools. These agents work in concert, and their collaboration is coordinated by an orchestrator. The result is a system that can reason through guidelines more carefully, cross-check its findings, and show its work in ways that clinicians and regulators can audit . |
The rest of this chapter examines multi-agent systems in oncology through concrete examples. We begin with TrustedMDT, the Oxford system that represents one of the most clinically realistic deployments of this technology to date. From there, we survey how similar architectures are being applied across the cancer care continuum, from drug discovery to clinical trial analysis to insurance claims processing. |
|
2. TrustedMDT: A Closer Look at the Oxford Multi-Agent System |
2.1 The Problem TrustedMDT Addresses |
TrustedMDT was developed by researchers at the University of Oxford, led by Dr. Andrew Soltan, an oncologist and engineer who designed the system based on challenges he encountered in his own clinical practice at Oxford University Hospitals . The system is designed to support MDT meetings by preparing and organizing information before and during the meeting, allowing clinicians to focus on judgment and discussion rather than on assembling facts. |
The system is integrated into Microsoft Teams through a collaboration with Microsoft, using Microsoft's Healthcare Agent Orchestrator. This integration is significant because it means TrustedMDT fits into the workflow that MDT teams already use, rather than requiring them to adopt a separate application or interface . |
2.2 The Three Agents |
TrustedMDT comprises three specialized AI agents working together. |
The first is the Clinical Summarization Agent. This agent analyzes electronic health records, including radiology reports, pathology results, and biomarker tests. It produces concise, tumor-specific summaries that capture a patient's current status in their cancer journey. Instead of a clinician spending twenty minutes reading through a stack of records, the agent produces a digest that can be reviewed in a fraction of the time . |
The second is the Cancer Staging Agent. This agent applies international standards for cancer staging, specifically the AJCC and UICC systems, to determine the stage of a patient's disease. Staging is foundational to treatment planning: a Stage II breast cancer is treated very differently from a Stage IV breast cancer, and getting the stage right matters enormously . |
The third is the Treatment Planning Agent. This agent drafts evidence-based treatment recommendations aligned with professional guidelines, drawing on sources such as the NCCN and ESMO guidelines. The recommendations it produces are draft plans for the MDT to review, refine, or reject . |
2.3 Why the Architecture Matters |
Dr. Soltan has explained the rationale for the multi-agent design in clear terms: 'Because standard chatbots struggle with the high-stakes complexity of oncology, we developed a hierarchical multi-agent system. In this architecture, each agent contains a dedicated team of sub-agents grounded in specific data with access to tools. This granular approach reduces the risk of 'guessing' because the system is required to reason through guidelines and explicitly cross-check its work against the patient's history' . |
The key insight here is that each agent is constrained in what it can do and what it can access. The Clinical Summarization Agent does not make treatment recommendations. The Staging Agent does not summarize records. The Treatment Planning Agent does not determine stage on its own. Each agent operates within its lane, and the orchestrator ensures that their outputs are combined coherently . |
This design also produces transparency. Because each agent's reasoning is separate, clinicians can inspect how a recommendation was reached, which data sources were consulted, and where disagreements or uncertainties remain . |
2.4 The Evaluation Plan |
TrustedMDT is undergoing a two-phase evaluation at Oxford University Hospitals. In Phase I, the system's accuracy is validated using anonymous cancer cases, with AI outputs benchmarked against expert decisions and physician preferences. In Phase II, the system is deployed in simulated MDT meetings with clinicians, assessing how effectively it summarizes information, supports discussion, and drafts treatment plans in a realistic clinical workflow . |
The study is supported by clinical specialists including Dr. Sajan Patel and Dr. Jaya Sharma, who contribute clinical expertise and quality assurance . The study protocol has received approval from the NHS Health Research Authority and a favorable ethical opinion from the University of Oxford . |
2.5 What TrustedMDT Represents |
TrustedMDT is notable not because it is the first AI system in cancer care, but because of how it is deployed and evaluated. It is one of the earliest deployments of 'agentic AI' within a clinically realistic tumor board setting, using real patient data under proper governance . It is not a laboratory experiment or a retrospective analysis. It is a pilot study designed to generate the evidence needed for larger-scale trials. |
The system also represents a philosophical commitment: the human remains the final decision-maker. As the Oxford team has emphasized, TrustedMDT is a 'digital collaborator,' not a replacement for clinical judgment. The AI prepares, the clinician decides . |
|
3. The Broader Landscape of Multi-Agent Systems in Oncology |
TrustedMDT is a flagship example, but it is part of a much larger movement. In 2025 and 2026, multi-agent AI systems have moved from theoretical proposals to real-world studies across cancer care and biomedical research. An editorial in a leading medical journal described this as 'the rise of the agentic AI colleague,' noting that multi-agent frameworks are now being applied from transplant selection committees to post-surgical patient management to data-driven drug discovery . |
3.1 Drug Discovery and Target Identification |
One of the most ambitious applications of multi-agent AI in oncology is in drug discovery. A system called the Virtual Biotech, developed by researchers at Stanford University, was designed to mirror the structure of a human biotechnology company. It comprises as many as 37,000 AI agents, led by a Chief Scientific Officer agent that delegates tasks to specialized scientist agents with expertise in areas such as statistical genetics, functional genomics, chemoinformatics, and clinical data . |
The Virtual Biotech was tested on a massive scale, analyzing outcomes from nearly 56,000 clinical trials to identify features of drug targets associated with trial success. One notable finding was that drugs targeting genes active in specific cell types were significantly more likely to progress through clinical development and reach the market, with lower rates of adverse events . |
The system was also asked to evaluate a specific protein called B7-H3 as a potential target for lung cancer. Integrating data from multiple sources, the agents proposed an antibody-drug conjugate strategy while identifying potential liabilities and opportunities for differentiation. According to external reviewers, this was a promising avenue for further investigation . |
3.2 Clinical Trial Design and Analysis |
Multi-agent systems are also being applied to the challenge of making clinical trials more efficient. The Virtual Biotech's analysis of tens of thousands of trials produced insights that could inform trial design, such as the association between cell-type-specific gene targeting and improved success rates . |
More broadly, agentic AI systems are being developed to analyze terminated trials to understand why they failed and to propose biomarker-guided enrollment strategies that could address precision medicine gaps. This kind of post-mortem analysis, which requires integrating data from genomics, clinical outcomes, and trial design, is well-suited to a multi-agent approach where different agents specialize in different data modalities . |
3.3 Pathology and Diagnostic Support |
While TrustedMDT focuses on treatment planning, other multi-agent systems are being developed for diagnostic support. A full-body AI agent framework has been proposed that would coordinate agents operating at different biological scales: molecule, organelle, cell, tissue, organ, and organ system. The goal is to enable cross-scale reasoning that connects molecular alterations to systemic phenotypes, which is particularly relevant for understanding cancer metastasis and for evaluating therapeutic responses . |
In a more immediate application, automated medical coding systems are using multi-agent architectures to assign ICD codes to clinical documentation. Coding is essential for billing, research, and quality reporting, but it is time-consuming and error-prone when done manually. Multi-agent systems that decompose the coding task into specialized subtasks have shown improvements in accuracy and consistency . |
3.4 Clinical Decision Support Beyond Tumor Boards |
Agentic AI systems are being applied to other high-stakes clinical decisions. Multi-agent AI committees have been shown to accurately identify patients eligible for liver transplantation, with improvements in health equity and consistency compared to traditional assessments . In ophthalmology, theoretical workflows have been proposed for agentic systems that autonomously manage diagnostic and care processes . |
What these applications share is a recognition that complex clinical decisions benefit from the same division of labor that human teams use. A transplant selection committee, like a tumor board, brings together different perspectives and expertise. Multi-agent AI systems replicate this structure in digital form. |
|
4. Applications Across Industries: Multi-Agent Systems Beyond Oncology |
The multi-agent architecture pioneered in systems like TrustedMDT is not limited to oncology or even to healthcare. The same principles---specialization, orchestration, cross-checking, and transparency---are being applied across industries where complex decisions require integrating diverse evidence. |
4.1 Healthcare Insurance and Claims Processing |
The healthcare insurance industry processes billions of claims annually, and the complexity of medical coding, policy interpretation, and regulatory compliance makes it a natural fit for multi-agent AI. A comprehensive framework for agentic AI in healthcare insurance has been proposed that integrates machine learning, natural language processing, and distributed computing to create autonomous agents capable of performing complex tasks in claims processing, fraud detection, and risk evaluation . |
Testing on real-world insurance datasets exceeding 500,000 claims demonstrated significant improvements: 94.2 percent accuracy in claims validation, a 78 percent reduction in processing time, and a 15.3 percent increase in detection of fraudulent activities compared to traditional systems . |
Another study focused specifically on medical claim documentation and coding, combining retrieval-augmented generation with structured agent reasoning. The results showed that incorporating reasoning substantially improved output fidelity, with the best-performing configuration achieving a semantic similarity score of 98.9 percent . |
These systems are not simply automating a single task. They are orchestrating multiple specialized agents: one for admissibility verification, one for financial adjudication against policy terms, one for fraud detection, and one for synthesizing findings into a report. The multi-agent structure allows the system to handle the diversity and complexity of insurance claims while maintaining audit trails for regulatory scrutiny . |
4.2 Drug Development and Pharmaceuticals |
Beyond the Virtual Biotech, multi-agent systems are being developed across the pharmaceutical research pipeline. A review of AI agents in drug discovery describes a progression from computational tools to autonomous systems capable of hypothesis generation and experimental design. Current industry data indicates 25 to 30 percent improvements in clinical trial success rates and 30 to 40 percent reductions in preclinical costs associated with AI agent adoption . |
The concept of a 'self-driving laboratory' extends the multi-agent paradigm to the physical world. In these closed-loop facilities, AI designs experiments, robotic systems execute them, and AI analyzes the results to decide on next steps. Each component can be viewed as an agent, and the orchestration layer coordinates their activities . |
4.3 Precision Medicine and Genomic Analysis |
Multi-agent systems are being applied to precision medicine, where treatment decisions depend on integrating genomic data, clinical history, imaging, and other patient-specific factors. Agentic AI enables adaptive, patient-specific approaches by continuously integrating demographic covariates, therapeutic drug monitoring data, biomarkers, and clinical response signals . |
The Full-Body AI Agent framework mentioned earlier represents a conceptual blueprint for how multi-agent systems could organize biomedical knowledge across scales. By decomposing questions into level-specific tasks and integrating outputs through iterative feedback, such systems could support more coherent reasoning about complex disease processes . |
4.4 Beyond Healthcare: A General Pattern |
The pattern that emerges across these applications is consistent. A complex, high-stakes decision requires integrating multiple types of evidence and expertise. A single AI model, even a powerful one, struggles because it lacks the structure to reason systematically across all these dimensions. A multi-agent system succeeds by dividing the problem into specialized sub-problems, assigning each to an agent with the right data and tools, and orchestrating their collaboration with human oversight. |
This pattern is applicable far beyond healthcare. It is relevant wherever decisions are complex, evidence is fragmented, and accountability matters. The specific agents and orchestrations will differ, but the architectural principles are general. |
|
5. Challenges and Limitations |
5.1 Hallucination and Reliability |
Multi-agent systems are not immune to the fundamental problem of AI hallucination. A study testing six large language models on physician-validated clinical vignettes containing fabricated details found hallucination rates ranging from 50 to 82 percent, with prompt-based mitigation reducing them to 44 percent . The multi-agent architecture helps by constraining each agent's scope and requiring cross-checking, but it does not eliminate the risk entirely. |
5.2 Data Fragmentation |
Agentic AI systems depend on access to data, and healthcare data remains notoriously fragmented. The lack of standardized data and the persistence of information silos make it difficult for multi-agent systems to function seamlessly across institutions and specialties . |
5.3 Accountability and Oversight |
When AI agents participate in clinical decisions, clarifying responsibility becomes more complex. If a Treatment Planning Agent drafts a recommendation that is accepted by an MDT, who is accountable for the outcomeThe physician who accepted it, the developer who built the agent, or the institution that deployed itThese questions are not fully resolved . |
5.4 Equity and Bias |
Multi-agent systems can still perpetuate algorithmic bias. The promise of greater equity through consistent, guideline-based reasoning has not yet been fully realized. Bias can enter through training data, through the design of agent interactions, or through the choice of guidelines and evidence sources . |
5.5 The Gap Between Simulation and Reality |
Most evaluations of multi-agent systems in healthcare have been conducted in simulated or retrospective settings. Prospective validation in real clinical workflows, with real patient outcomes, is still rare. The TrustedMDT pilot is notable precisely because it is one of the first to move into a clinically realistic setting with proper governance . |
|
6. Future Trajectories |
6.1 From Copilot to Colleague |
The trajectory of multi-agent AI in oncology and beyond points toward systems that function less like tools and more like colleagues. The current generation of systems prepares information, drafts recommendations, and awaits human review. Future systems may take on more proactive roles, flagging issues, suggesting alternatives, and participating in discussions. |
6.2 Integration with Clinical Workflows |
The success of TrustedMDT's Teams integration suggests that the future of clinical AI is not a separate application but an embedded capability within the tools clinicians already use. This approach reduces friction and increases the likelihood of adoption . |
6.3 Regulatory Evolution |
The regulatory landscape for AI in healthcare is evolving rapidly. By early 2026, the FDA had authorized more than 1,350 AI-enabled devices, roughly double the number from 2022. However, analyses have found that only a small fraction of these devices underwent prospective testing or included human-in-the-loop evaluation . The gap between regulatory authorization and robust evidence of clinical benefit remains a concern. |
6.4 The Path Forward |
The future of multi-agent AI in oncology and other domains depends on several factors: the development of open, interoperable data standards with appropriate safeguards; continuous bias assessment and transparent reporting of model limitations; and the involvement of multidisciplinary teams including clinicians, engineers, ethicists, and patients as co-designers . |
|
7. Conclusion: A Detailed Summary |
This chapter has examined multi-agent systems in oncology and beyond, using TrustedMDT as the central example and branching out to applications across drug discovery, clinical trials, insurance, and precision medicine. |
The core insight of the multi-agent approach is that complex, high-stakes decisions benefit from specialization and orchestration. A single AI model, no matter how capable, struggles with the complexity of oncology because it must simultaneously summarize records, determine stage, apply guidelines, and integrate evidence from multiple sources. A multi-agent system divides these tasks among specialized agents, each grounded in specific data and tools, and orchestrates their collaboration. |
TrustedMDT, developed at the University of Oxford and integrated into Microsoft Teams, comprises three agents: a Clinical Summarization Agent that analyzes electronic health records, a Cancer Staging Agent that applies international standards, and a Treatment Planning Agent that drafts evidence-based recommendations. The system is being evaluated in a two-phase study at Oxford University Hospitals, with Phase I validating accuracy against expert decisions and Phase II deploying in simulated MDTs with clinicians. This pilot represents one of the earliest deployments of agentic AI in a clinically realistic tumor board setting . |
The broader landscape of multi-agent systems in oncology and life sciences is expanding rapidly. The Virtual Biotech, with its thousands of agents led by a Chief Scientific Officer agent, has demonstrated the ability to analyze tens of thousands of clinical trials and identify features of drug targets associated with success. Multi-agent systems are also being applied to clinical trial design, pathology, medical coding, and transplantation decisions . |
Beyond healthcare, the same architectural principles are being applied to insurance claims processing, where multi-agent systems have achieved significant improvements in accuracy, speed, and fraud detection . The pharmaceutical industry is investing in agentic systems for drug discovery and development, with reported improvements in trial success rates and reductions in preclinical costs . |
Yet significant challenges remain. Hallucination is not eliminated, only mitigated. Data fragmentation limits the ability of systems to function across institutions. Questions of accountability and oversight are unresolved. Equity and bias concerns persist. And the evidence base, while growing, still relies heavily on simulated and retrospective studies rather than prospective, real-world validation . |
The trajectory of multi-agent AI in oncology and beyond points toward systems that function as colleagues rather than tools, embedded in the workflows clinicians already use, subject to evolving regulatory frameworks, and developed with input from diverse stakeholders. Whether these systems ultimately improve patient outcomes depends on addressing the challenges of reliability, transparency, accountability, and equity. The early evidence is promising, but the hard work of rigorous evaluation and responsible deployment is just beginning. |
cs@easiersoft.com