Abstract
Public-interest focus. This paper examines agentic infrastructure as a means of advancing democratization of care: wider practical access, affordability, informed choice and equitable participation. Claims of benefit require evidence about who gains and who remains excluded.
Question. How can programmable agents coordinate heterogeneous care systems without confusing the ability to invoke a tool with the authority or competence to perform a consequential action? This paper develops a reference architecture for programmable care-coordination systems in which agent reasoning, clinical semantics, authorization and workflow execution remain distinct but connected layers.
Evidence base. The technical analysis draws on HL7 FHIR, SMART App Launch, MCP authorization and security documentation, NIST's generative AI risk profile and primary vendor materials.[1][2][3][4][7] These establish interface concepts and risk-management guidance. They do not establish that an implementation exists or that an agent has demonstrated clinical safety.
Finding. MCP can expose discoverable tools to an agent; FHIR can structure health information; SMART can support scoped application authorization. None of these alone resolves patient identity, clinical appropriateness, cross-organization accountability or the economics of implementation. A useful care hub must compose these mechanisms around a defined workflow with explicit control of writes, disclosure and escalation.
Proposed contribution. The paper specifies a bounded execution architecture, an interoperability contract, a threat model, a synthetic evaluation suite and a staged developer program. Neural networks and transformer models are treated as probabilistic components. Clusters provide computational capacity and isolation strategies; they are not evidence of general intelligence or reliability.
Design thesis: Let an agent propose and interpret; let policy determine permission; let a controlled executor perform authorized actions; let independently observable evidence determine success.
The term “agentic” is used here for a system that can select and sequence tools within an assigned task. It does not imply AGI, autonomous medical practice or unlimited delegation. The proposed research and analytics function monitors evidence and performance, but cannot silently broaden its own authority.
Key findings and implications
- FHIR structures health information; MCP exposes tool interfaces; neither alone establishes authority for a consequential action.
- A controlled executor and observable workflow state can separate model proposals from authorized actions.
- Inclusive agent evaluation should measure language access, human handoff and integration burdens across organizational sizes.
Why this matters for democratization of care
Agentic systems as accessible coordination infrastructure
Programmable agents can be studied as an accessibility mechanism: they may help people interpret administrative language, move between services and express preferences through different communication modes. These possibilities do not establish that an agent is safe, accurate or beneficial. The democratization question is whether the entire service expands effective choice and successful care completion across groups.
An agent should not become a compulsory gatekeeper. People need understandable routes to human assistance, correction and withdrawal from an automated interaction. The care organization must retain responsibility for consequential decisions and ensure that failures do not strand a patient between systems. A service that performs well for fluent, digitally confident users may still exclude people with limited connectivity, disability or uncommon communication needs.
Developer participation also has distributional consequences. Documentation, affordable testing environments, accessible conformance tools and transparent connector terms can enable community organizations and smaller developers to contribute. A nominally open protocol can coexist with inaccessible commercial conditions. Research should measure practical participation rather than count published tool definitions.
| Dimension | Proposed measure | Interpretation safeguard |
|---|---|---|
| Language access | Task completion and error correction by preferred language | Evaluate intended use with community input |
| Human assistance | Time to an effective human handoff | Do not count a displayed phone number as a successful handoff |
| Provider participation | Integration burden by organizational size | Include maintenance and incident response |
| Accountability | Ability to inspect, challenge and reverse appropriate actions | Evaluate real disputes and recovery pathways |
These measures are a proposed evaluation framework, not established findings about an existing service. Report baseline conditions, uncertainty, excluded populations and adverse results. A credible study can conclude that an intervention is useful, ineffective or inequitable; democratization is the question being tested, not a benefit assumed in advance.
Methods and evidence boundaries
The review selected primary standards and vendor documentation relevant to agent-facing tools, health-data exchange, authorization and risk management. Specifications were read as normative interface descriptions; vendor materials were treated as claims about an offering. The review did not execute conformance tests, inspect implementation source code or access a production care environment.
The architectural synthesis is a proposed design argument. It decomposes a consequential task into interpretation, authorization, execution and verification, then identifies failures at each boundary. The design is assessed for testability and accountability, not claimed empirical superiority. A comparison with a simpler non-agent workflow remains necessary.
This is an AI-assisted, targeted research working paper, not an independently peer-reviewed study. Proposed models and interventions are not evidence of deployed capabilities or measured outcomes. The full PDF contains the detailed analysis, assumptions and limitations.
References and source notes
Reference numbers match the PDF. Public sources are linked below; preliminary supplied planning materials are identified as such and do not constitute independent verification.
- HL7 International. FHIR Release 4: Overview. FHIR R4, version 4.0.1. Primary technical specification. Version is explicit; this paper does not assert it is the newest FHIR release.Read the source
- HL7 International. SMART App Launch Implementation Guide. Version 2.2.0 consulted. Primary technical specification. Authorization and application-launch context; deployed versions vary.Read the source
- Model Context Protocol. Authorization. Specification revision 2025-11-25. Primary protocol specification. Describes authorization for HTTP-based transports; not healthcare law or a clinical semantic standard.Read the source
- Model Context Protocol. Security Best Practices. 2025-11-25 documentation path. Primary security documentation. Protocol-related threats and mitigations; implementation must be tested locally.Read the source
- Cognizant. Cognizant Brings Agentic AI and MCP tool library to Core Claims Operations with Workflow Agentic Processing for TriZetto. September 28, 2026. Primary vendor announcement, not an independent outcome evaluation. Announces 100+ MCP tools and availability for Facets and QNXT; performance results were described as forthcoming.Read the source
- Stedi. MCP server documentation. Product documentation accessed for 2026 review. Primary vendor documentation. Evidence of an interface offering, not proof of Careverse™ integration or operational effectiveness.Read the source
- National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. July 26, 2024. NIST AI 600-1. DOI: 10.6028/NIST.AI.600-1. Voluntary risk-management guidance; not a certification.Read the source