Agentic AI Is Coming to Medicine. Cage It Where It Counts.

Eric Topol had a piece recently, Agentic AI Comes to Medicine, that is worth reading in full. He walks through two agentic systems published in Nature: MIRA for the emergency department and AMIE for outpatient care. The headline numbers are striking. MIRA reached 87.8 percent diagnostic accuracy against 78.1 percent for board-certified physicians, and on appendicitis it hit 100 percent against 88 percent. If you build in this space, results like that get your attention, and they should.

But Topol is careful, and so am I. Those systems ran on text-only inputs, clean and complete data, and patient actors or capped interactions. Real medicine is none of those things. It is incomplete, conflicting, and messy, which is where confident systems get into trouble. So my stance is not to dismiss agentic AI, and not to hand it the clinic either. It is to welcome the capability and cage the autonomy, giving the agent room to work exactly where the stakes and the mess are both low.

The results are real

I want to be honest that agentic AI just cleared a bar narrow AI never did: end-to-end reasoning that beats physicians on a controlled test. For years, medical AI meant a narrow model doing one step, usually flagging a diagnosis for a human to act on. What Topol describes is different: systems that reason across the whole encounter and make decisions, and do it well on the test they were given. The chart below is the headline. On a controlled diagnostic evaluation, the agentic system beat board-certified physicians. That is a real milestone, and pretending otherwise would be dishonest.

Clean study, messy clinic

The distance between that number and a safe deployment is the same distance between a clean dataset and a real patient. Here is what the number does not include. Real patients arrive with incomplete and conflicting information, not the tidy vignettes these systems were tested on. They speak, they hedge, they leave things out. The evaluations used text only, and often patient actors, with the interaction capped. None of that is a knock on the research, it is the research being honest about its limits. But it means the gap between 87.8 percent on a controlled test and safe autonomous care in a real emergency department is enormous. I have watched too many tools ace the demo and stumble on the mess to confuse the two.

Where agentic AI fits, and where it does not

So I do not ask whether to use agentic AI. I ask where, and the answer falls out of two questions: how high are the stakes, and how messy is the input. Map any task on those two axes. Low stakes and clean input is where I let an agent run with light oversight. High stakes or messy input is where I cage it: a bounded step, a human on the decision, a deterministic backbone around it. The dangerous corner is high stakes and messy input, which unfortunately describes a lot of real medicine, and which is exactly where a confident agent will hurt someone. Autonomy is a dial, and I set it by the corner the task lives in.

I am optimistic about agentic AI in medicine, and I am in no hurry to hand it the keys. Both can be true. The research is real progress, and real medicine is still messier than any benchmark. So I welcome each new result, I read the caveats as carefully as the headline, and I keep the agent caged where the stakes and the mess are high. When the evidence moves from clean studies to messy clinics, I will move the dial. Not before.