Clinical Agent Governance Engine
NANDA makes clinical AI agents discoverable and interoperable.
CAGe makes them governable.
The Problem
Healthcare agents will increasingly retrieve clinical information, coordinate with other agents, access EHR data, draft recommendations, and initiate actions. The key question is no longer just whether a model is accurate — it's whether a networked agent is authorized, operating within its validated scope, using permitted data and tools, escalating appropriately, and leaving a trustworthy record of what it did.
Project NANDA provides agent identity, discovery, and interoperability at internet scale. CAGe is a proposed open-source layer that sits between clinical agents and the systems they use — supplying the domain-specific controls healthcare needs before those capabilities can touch real patients.
A Governed Clinical Action
Agent requests an action
e.g., draft a medication recommendation
Gateway verifies identity & scope
NANDA identity · Clinical AgentFacts · validated population
Local policy is enforced
ALLOW · ALLOW WITH REVIEW · DENY · ESCALATE · DEGRADE
Clinician reviews & decides
Accept, edit, reject, or escalate — with evidence shown
Signed receipt generated
Privacy-preserving audit record — no PHI duplicated
The Central Idea
Govern the whole clinical action — agent, user, patient context, purpose, data, tools, evidence, and outcome — not just the underlying model.
What CAGe Provides
A healthcare extension to NANDA AgentFacts: machine-readable declarations of intended use, exclusions, evidence, validated populations, data requirements, permitted actions, and required oversight.
Middleware that intercepts agent-to-agent and agent-to-tool calls (MCP, A2A, NANDA, FHIR) and returns ALLOW, DENY, or ALLOW WITH REVIEW decisions — zero trust, applied to agents.
Reusable organizational policies for medication advice, patient communication, documentation, pediatrics, behavioral health, and sensitive data — authored in simple YAML, enforced rigorously.
Signed, privacy-preserving records of each meaningful agentic event: models, data categories, tools, policies, evidence, approvals, and outcomes — auditable without duplicating PHI.
Explicit delegation patterns — from Inform to Prohibited — specifying what an agent may prepare, what a person must inspect, and who carries responsibility.
Interfaces for clinicians, governance committees, developers — and eventually patients — showing agent inventories, decisions, receipts, violations, and review dates.
Human Oversight as Infrastructure
The governance engine selects the required mode based on action risk, patient population, evidence quality, and local approval.
Built on Project NANDA
NANDA answers "who is this agent and how do I reach it?" — it deliberately doesn't answer "may it act on this patient?" CAGe fills that gap as a layer on top of the NANDA Index: Clinical AgentFacts is an extension schema the Index can resolve like any AgentFacts record, and the gateway sits on the protocols agents already speak.
Through NANDA Town, CAGe would ship as a callable governance skill, a governed clinical-agent starter template, conformance tests, and machine-verifiable governance badges — badges that assert governance properties, never medical approval. Authority stays local, with each healthcare organization.
Governance skill callable by any NANDA-compatible agent before a clinical action.
Clinical AgentFacts extension indexed and displayed with the base agent record.
Starter repository & policy packs published through NANDA Town.
Adapters for NANDA messaging, MCP tool calls, A2A communication, and FHIR access.
Conformance tests & governance badges attached to NANDA Town listings.
First Demonstration
Not an autonomous diagnostician — a deliberately narrow agent that retrieves approved local guidance, checks the patient is within the validated population, requests only the minimum FHIR data, drafts an evidence-linked recommendation, requires clinician confirmation, cannot place orders, and generates a signed receipt. Clinically meaningful, while keeping the safety and regulatory burden manageable.
Open
Source · Apache 2.0
FHIR
MCP · A2A · NANDA
Local
Institutional Authority
Human
Centered Governance
Status
A perspective paper describing the CAGe architecture, its integration with the NANDA Index and NANDA Town, and a technical, clinical, and human-factors evaluation agenda is in preparation (arXiv preprint, 2026). Interested in collaborating, contributing a policy pack, or piloting the governed guideline agent at your institution?