Two minutes. That's how long most tumor boards in the UK have to discuss each patient's cancer treatment plan. In that window, clinicians need to synthesize radiology reports, pathology results, biomarker tests, and treatment guidelines to make decisions that directly affect survival outcomes.

When information gaps appear (missing test results, incomplete staging data), the case gets postponed. This happens in 7% of cases. Meanwhile, caseloads keep rising and expert capacity stays fixed.

Dr. Andrew Soltan saw this operational reality from inside. He's a practicing oncologist at Oxford University Hospitals and Academic Clinical Lecturer in the Department of Oncology. So he built TrustedMDT, a multi-agent system that works directly in tumor board workflows through Microsoft Teams.

Soltan's work reveals something specific: when someone who actually practices oncology builds agents for their own workflow, the design constraints become visible. Explainability under time pressure. Audit trails in regulated environments. Integration without workflow disruption. These aren't theoretical requirements. They're what production-ready means when errors have immediate human consequences.

When you're making cancer treatment decisions in two minutes, "mostly accurate" isn't good enough.

Standard chatbots can't explain their reasoning or show their work. Soltan designed a hierarchical system where each agent contains dedicated sub-agents grounded in specific data with access to tools. This granular approach means the system must reason through guidelines step-by-step and explicitly cross-check its work against patient history. The architecture isn't elegant. It's necessary. In regulated environments, you need to show exactly how the system reached its recommendation.

The system uses three specialized agents:

• EHR analysis agent — Produces tumor-specific summaries from electronic health records
• Staging agent — Applies international cancer staging standards
• Treatment agent — Drafts recommendations aligned with professional guidelines

During actual tumor board meetings, clinicians can probe the recommendations and provide new information in real-time.

The production challenges here look different from typical web automation. The system must extract data from hospital electronic health records, navigating whatever integration methods the hospital provides. It must make sense of clinical notes where one oncologist writes "mets to liver" and another writes "hepatic metastases" and both mean the same thing. It must extract staging information when pathology reports use different formats across hospitals. And it must preserve the context that makes clinical data interpretable. "Stable disease" means something different for a patient on their first treatment versus their fifth.

Soltan's previous work taught him that deployment reality differs sharply from research prototypes. He built CURIAL, an AI screening test for COVID-19 that actually got used in clinical practice. That experience showed him the gap between "works in research" and "works in hospitals." TrustedMDT embeds into existing Teams workflows rather than forcing clinicians to adopt new tools.

TrustedMDT received regulatory approval from NHS Health Research Authority in December 2025 and will pilot at Oxford University Hospitals in Q1 2026. The pilot will measure whether the system actually reduces case postponements and helps tumor boards handle rising caseloads. The operational outcomes that matter.

Regulated environments require more than reliable execution. They need systems that can explain their reasoning, maintain audit trails, handle unstructured data while preserving context, and integrate with legacy systems without disrupting established workflows. The challenges center on making AI systems that humans can actually trust with high-stakes decisions.

Building web agent infrastructure at scale means encountering similar requirements. When domain experts build for their own operational reality, they surface what infrastructure must handle in production. Two minutes per patient isn't a design choice. It's operational reality. And when someone who actually practices oncology builds agents for tumor boards, you see what production-ready means in contexts where the stakes are immediate and human.