RareDx coordinates 12 Bedrock agents for rare-disease diagnosis support

AWS has demonstrated RareDx, a prototype that coordinates 12 specialist agents to support clinical geneticists from referral intake to a structured report. Bedrock AgentCore, Strands Agents and domain tools handle document extraction, phenotype coding, literature and variant work, while the physician remains responsible for clinical reasoning and the final diagnosis.

A prototype for the diagnostic administration burden

AWS has demonstrated RareDx, a multi-agent prototype intended to support physician geneticists through the long administrative and analytical journey of rare-disease diagnosis. The official source is dated 9 October. AWS cites more than 300 million people living with roughly 7,000 classified rare diseases and an average diagnostic journey of 4.8 years in Europe. Referrals may contain dozens of pages of specialist letters in several languages and low-quality scans.

A geneticist must extract symptoms, map them to Human Phenotype Ontology codes, identify gaps, consider images, interpret genomic variants, review current literature and produce a compliant report. RareDx brings those materials into one interface. Its stated purpose is to reduce switching and preparation work so the clinician can spend more time on reasoning and communication, not to make an autonomous diagnosis.

Twelve specialists instead of one general model

The system uses a clinical coordinator to orchestrate 12 specialist agents. AWS lists roles for patient-summary synthesis, phenotype analysis, biomedical research, terminology, pedigree work, clinical guidelines and report generation. Each has its own prompts, tools and model choice. The modular design allows one specialist to be tuned or replaced without rebuilding the whole application.

This decomposition reflects the problem: extracting a timeline from scanned letters is different from interpreting a variant or formatting an insurance-coded report. A single agent with every tool would carry a large context and broad authority. Specialists can narrow the evidence and permissions used at each stage, although the coordinator still needs clear rules for resolving conflicts and passing uncertainty forward.

AgentCore supplies runtime and connection points

RareDx is built with the Strands Agents SDK and runs on Amazon Bedrock AgentCore Runtime. AWS Amplify provides the front end, while the coordinator reaches memory and tools through AgentCore Gateway. Other AWS services support identity, storage and application state. The architecture gives each step a managed execution environment and allows tool access to be traced rather than hidden inside one opaque prompt.

Traceability is essential in clinical work. A final report should show which source letter, database or publication supports each material statement. The system must distinguish patient facts from generated interpretation and preserve model and tool versions. Without those records, a polished report can be difficult to audit when evidence changes or a clinician challenges a suggestion.

Human judgement is the release gate

A diagnostic assistant can summarise, code and retrieve, but it cannot assume responsibility for consent, phenotype interpretation or disclosure to a family. Rare diseases often have incomplete evidence and uncertain variants. The clinician needs to see confidence, alternative explanations and missing information, and must be able to reject an agent’s work without the rejection being silently overwritten later in the workflow.

Evaluation should therefore be task-specific. Document extraction can be scored against labelled records, ontology mapping against expert coding and literature retrieval against a known search set. Report generation needs factuality and completeness checks. A single overall model benchmark would conceal which stage creates risk.

Promising workflow, early evidence

RareDx is a prototype rather than a reported clinical deployment with patient outcomes. That is appropriate for a high-consequence domain, but it limits the claims that can be made today. Useful next evidence would include prospective time savings, error rates by agent, clinician override patterns and performance across languages and document quality. Privacy and medical-device obligations will depend on how and where the system is used.

For Australian health services, the architecture illustrates how agent specialisation can organise a difficult pathway without pretending that one model contains all medical expertise. The opportunity is to return time to geneticists and make evidence easier to inspect. The standard for success must remain a safer, more consistent clinical process, not simply a faster generated report.

Coordination creates its own failure modes

A fleet of specialists can reduce context overload, but it introduces hand-off risk. One agent may summarise a symptom imprecisely, and later agents can treat the summary as a confirmed fact. RareDx therefore needs typed outputs, source links and confidence carried through the coordinator. Critical observations should be rechecked against the original record rather than repeatedly paraphrased between agents.

The coordinator also needs rules for disagreement. A phenotype agent and literature agent may support different diagnoses, or a guideline may conflict with a recent paper. The safe response is to expose the conflict to the clinician, not to ask a model to create artificial consensus. Evaluation cases should include contradictory, incomplete and multilingual records because those are the situations where orchestration quality matters most.

Clinical teams will also need a clear correction workflow. When a physician changes an extracted symptom or rejects a code, the application should record the correction, propagate it to dependent steps and preserve the earlier state for audit. Silent edits can make a final report look coherent while hiding how much expert repair was required. That repair rate is itself an important evaluation metric.