CADUCEUS: Pressure-Testing Agentic AI on the Highest-Stakes Domain There Is

A molecular tumor board is the multidisciplinary review every cancer centre runs weekly: pathology, radiology, genomics, labs and clinical notes fused into one defensible plan. It is the textbook case for multi-agent AI — many specialists, real concurrency, a decision that lives in the connections across modalities. It is also the textbook case for getting AI in medicine wrong. I built CADUCEUS to see whether the agentic patterns I care about hold up under the highest stakes there are.

Try the live demo →

Why this is the right place to pressure-test

Three things sink clinical AI, and they are precisely the three things that sink agentic AI everywhere else:

1. No provenance — a confident claim no one can trace back to a source is a liability, in oncology fatally so.
2. Silent consensus — a committee of agents that all agree is just one opinion wearing many hats; nobody argued the other side.
3. Automating a decision that must keep a human on the hook — legally and ethically, a clinician has to dispose.

If the patterns survive here, they survive in a bank. CADUCEUS is built around fixing exactly those three.

55+ agents, 7 layers, genuine concurrency

Given a full case, CADUCEUS convenes a hierarchical board: a supervisor (Board Chair) over departmental sub-graphs — Intake, a Diagnostic department (Pathology, Radiology, Molecular, Labs), Therapeutic plus Pharmacy, Evidence/Trials/Guidelines, Supportive, and Synthesis. The data spine is a medallion: Bronze raw → Silver FHIR-normalised → a Gold unified Patient Case object the agents consume.

The concurrency is real, not decorative. The four diagnostic departments have no mutual dependency, so they execute as one parallel wave. Then the map-reduce move: Variant Actionability, Patient-Trial Matcher and Abstract Screening each fan out one ephemeral worker per item and a reduce step aggregates — which is what drives peak concurrency well past 40, to roughly 150 with the bundled cases, streamed live to the UI over SSE so you can watch the graph light up and the concurrency meter climb. Dependency gates keep it honest: therapeutics wait for diagnosis, pharmacy waits for the proposed regimen.

Provenance and dissent, by construction

Two safety layers are the whole point. A Citation Validator rejects any recommendation lacking a retrievable source — the audit log targets ~100% citation grounding, and an unsourced claim is blocked before a human ever sees it. This is the same instinct as RegRadar's programmatic citation verifier: "please cite your sources" is a prompt, not a control; a validator that checks the claim against the knowledge base is.

In oncology, a confident-but-wrong claim is the failure mode. No source, no statement.

And a dedicated Dissent / Red-Team agent exists solely to attack the consensus, alongside a Contradiction Checker that surfaces modality conflicts. Silent agreement is the failure mode of committee AI; CADUCEUS makes the minority view a first-class output rather than something smoothed over. It is the same idea I later gave an objection-scoring formula in DELPHI — build the skeptic in.

Real data, keyless, honestly labeled

CADUCEUS runs two ways. Offline, three synthetic cases exercise distinct reasoning paths — an EGFR exon-19-deleted NSCLC with a solitary brain met, a HER2+/HR+ metastatic breast case, and a BRAF V600E MSI-high colorectal board deliberation. Live, it runs on a real, de-identified TCGA patient assembled in real time from the NCI Genomic Data Commons open tier, with variants annotated against CIViC, drug safety from openFDA + FAERS, trials from ClinicalTrials.gov v2, and literature from Europe PMC — six keyless, free public APIs, with open-tier data gaps marked honestly rather than fabricated. On synthetic cases, guideline concordance scores 95–97% against NCI PDQ for matched cancer types; on live cases that is shown honestly as evidence-grade strength, because PDQ has no machine-readable API to grade against.

Crucially: decision support, not decision making. Human-in-the-loop checkpoints clear after diagnosis and before finalisation. CADUCEUS drafts; a clinician decides. It is a research prototype, explicitly not an FDA-cleared medical device.

What this taught me

1. The patterns are domain-agnostic. Provenance enforced in code, a dedicated adversary, human gates, a deterministic backbone with the LLM on top — they are the same whether the output is a treatment plan or a credit memo.
2. Concurrency has to be earned, then bounded by dependencies. Map-reduce fan-out gets you genuine parallelism; dependency gates keep it from producing a regimen before a diagnosis.
3. Honesty about data gaps is a feature. Marking what the open tier does not give you, instead of hallucinating it, is exactly what makes a risk-conscious reviewer trust the rest.

I do not plan to work in oncology — I plan to build agentic AI for finance. But the fastest way to find out whether provenance, dissent, and human gating actually hold is to wire them into the highest-stakes board there is. CADUCEUS is that stress test, passed.

Try the live demo → · Source on GitHub → · Full case study →