The following is a guest article by Adam Farren, CEO at Canvas Medical
The Model Context Protocol (MCP) has already started to reshape how AI systems access data across every software market. For any organization looking for better outputs from LLMs and agents, MCP holds real promise. It gives agents a standardized way to reach external context and invoke tools in a structured interface. Originally proposed by Anthropic and now maintained as an open standard with contributions from OpenAI, Google, and Microsoft, MCP has significant industry momentum behind it as a foundational layer for AI applications.
But in healthcare, MCP by itself is not enough. Not because the protocol is flawed; it is well-designed for what it does. The problem is that healthcare’s hardest challenges exist at a layer MCP was never designed to address.
MCP Solves Communication, Not Coordination
MCP extends a large language model’s capabilities by giving it access to new context and new tools through a structured client-server interface. That is a meaningful unlock.
However, in healthcare, the challenges are not about communication between agents and data sources. They are about coordination, governance, and state management across systems where the patient record can span multiple context sources and systems.
Consider a multi-agent system handling prior authorizations. One agent gathers clinical documentation, another checks formulary requirements, and a third submits to the payer. What central system coordinates their work? Who ensures they are not operating on stale data? Who resolves conflicts when two agents attempt to update the same record? Who guarantees observability of the agent’s autonomous action?
MCP does not answer these questions. It is a protocol for connections, not a framework for coordination, and healthcare organizations need both.
The Infrastructure Problems MCP Leaves Open
Identity and access are out of scope for MCP. MCP intentionally keeps organizational identity governance out of scope. Healthcare organizations still need role-based access controls, audit logging, and the ability to prove to regulators exactly which agent accessed which patient data and when. This is a table-stakes capability that any production-ready healthcare technology must offer.
Shared state is genuinely hard, and MCP doesn’t solve it. When one agent updates a patient’s chart, every other agent working on that patient’s care needs to know immediately. Without a centralized coordination layer, you end up with multiple agents operating on outdated or conflicting information. One may recommend a medication another has already flagged, or document a visit a third agent is still processing. In healthcare, that kind of lag is a major patient safety risk.
Lifecycle management isn’t provided by MCP. Sampling agent outputs for quality, shutting down an agent that’s making errors, managing schema versioning, and resolving conflicts when agents work at cross purposes are all examples of the need for centralized oversight that MCP does not provide.
Healthcare AI Needs a Hub-and-Spoke Model
When every agent communicates directly with every other without a central authority, complexity scales exponentially. In healthcare, this is not an engineering problem; it is a patient safety issue. The answer is not to abandon MCP, but to recognize that MCP needs orchestration around a centralized coordination authority. Healthcare requires a hub-and-spoke architecture where:
- The EMR is the hub; it is the source of truth for patient data and the center of clinical workflows, and it is already established for deterministic workflows and in the right system, extensible for agentic ones
- Agents are spokes; they operate in a coordinated runtime, hosted centrally within the EMR runtime, which means with shared access to structured, real-time context centered on the patient record
- MCP enables connections at the periphery; external data sources and tools connect through MCP, and that access is managed through the governed orchestrator, not freelancing across a mesh
This is where a well-designed software development kit becomes essential infrastructure. The right SDK gives developers the ability to build and deploy custom workflows natively within the EMR environment, across scheduling, charting, and billing, without standing up separate systems. The SDK provides a consistent, low-complexity toolset for building multi-agent systems without incurring the combinatorial cost of stitching together context, tools, and state externally.
That foundation delivers:
- Native EMR deployment across scheduling, charting, billing, and beyond — no separate systems to stand up or maintain
- Real-time event streams emitting clinical, operational, and financial data, so agents always operate on the current patient context
- Built-in governance with conflict resolution, safety rails, and end-to-end auditability from a single point of control
- Pre-built integrations with AI providers, cloud services, and communication tools that reduce the surface area developers have to manage
The entire model stays auditable from a single point of control.
Implications for Tech Leaders
For leaders building or buying healthcare AI infrastructure, the priorities are clear.
Demand an orchestration layer, not just MCP connectivity. The ability to govern, monitor, and coordinate agents in real time is what separates a compelling pilot from a production-ready system. If a vendor shows you MCP connections but cannot explain the coordination authority, keep looking.
Treat the EMR as the coordination hub. Any architecture that routes agents around the system of record will eventually break under the weight of state synchronization failures and compliance gaps. Build from there.
Build auditability from day one. In regulated environments, the ability to trace every agent action back to a responsible system is not optional. It is the baseline for earning clinical trust.
Where MCP Goes From Here
MCP will succeed in healthcare, but only as one layer of a multi-layer stack, not as the entire solution.
The multi-agent future in healthcare is coordinated intelligence, centered on the system of record. MCP handles external connections, while a robust platform handles coordination, governance, task orientation, and safety. This is not a limitation of MCP, but rather an acknowledgment that healthcare’s problems require solutions at multiple layers, and that agent coordination with shared context is foundational to patient safety.
Adam leads Canvas Medical with over 15 years of experience in startups and healthcare. He combines his deep technical expertise with a passion for transforming care delivery. Before joining Canvas, he served as Chief Growth Officer at Elation Health and Osmind, where he played a pivotal role in expanding both companies into national markets, driving rapid customer acquisition, and creating new revenue streams alongside their EMR solutions. Adam holds a BA from Princeton University and an MBA from the Haas School of Business at UC Berkeley.
No comments:
Post a Comment