The Plumbing for AI Agents: MCP, A2A, and Why Healthcare Should Care

Let me sell you on the least glamorous topic in AI, because it is quietly the most important one for healthcare. Everyone is excited about agents, AI that does not just answer but acts. Here is the boring truth nobody puts on a keynote slide: an agent is useless until it can safely reach the systems it needs, the chart, the drug database, the guidelines. Right now, wiring each agent to each system is a bespoke, brittle mess. As Wolters Kluwer puts it, without a standardized way to connect, every new agent adds friction, risk, and cost. The fix is a pair of protocols with deeply unsexy names, and they are about to matter enormously.

Two protocols, one goal: let agents actually do things

There are two acronyms worth learning, and the difference between them is simple once you see it. One connects an agent to tools. The other connects agents to each other.

If you read my piece on the multi-agent turf war, A2A is the layer that is supposed to keep that turf war from happening, a shared language so agents negotiate like colleagues instead of strangers. But A2A only matters once each agent can actually do something on its own, and that is MCP's job. So let us start there.

MCP is the universal translator

The cleanest way to picture MCP is a universal translator sitting between the agent and the messy world of clinical systems. The agent does not learn a hundred different APIs. It asks the MCP server for what it needs, and the server fetches structured data from trusted sources through one consistent interface. Tap through the shape.

The quietly brilliant node is the audit log. Because every request flows through one interface, you can record exactly what an agent asked for and got. In most software that is nice. In healthcare, where you may one day have to explain to a regulator or a family exactly why an agent did what it did, a complete, reviewable trail is not a feature, it is the price of being allowed to play.

The healthcare catch: do not trust this plumbing yet

Now the cold water, because I would be doing you a disservice to sell only the dream. A recent Springer perspective on agentic systems and MCP in medicine is refreshingly blunt: the ecosystem is immature, current implementations lack unified authentication, standardized logging, and secure-by-design protocols, and full autonomy for high-stakes decisions is ethically contentious and operationally unproven. Translated: the pipes are exciting, and they are also leaky in ways that matter when the fluid is patient data and clinical action. The right response is not to wait, and not to go all in. It is to match autonomy to stakes.

That table is the whole adoption strategy on one page. Point your agents at the low-stakes, high-annoyance work first, the summarizing and fetching and drafting, where a mistake is cheap and a human is still in the chair. Earn trust there while the security and liability standards catch up. Do not let a slick MCP demo talk you into wiring an autonomous agent to a dose calculation, because the protocol that makes that easy to build does nothing to make it safe to do.

Here is my actual advice, from someone who has watched integration debt sink more healthcare projects than bad models ever did. Learn MCP and A2A now, while they are still emerging, the same way the smart teams learned FHIR before it was mandatory. Build your first agents on the standard, keep them on the low-stakes end of that table, and insist on the audit log from day one. Do that, and when the plumbing finally hardens into something you can trust for the big stuff, you will already be living in the right house. Everyone else will be ripping out custom pipe.