A Nature paper published this week is worth your attention.
MIRA (Medical Intelligence for Reasoning and Action) is an autonomous AI agent that operates directly within EHR environments. Tested against 574 real-world emergency department cases from MIMIC-IV, it achieved 88.9% diagnostic accuracy — and in a head-to-head comparison with experienced physicians, outperformed them 87.8% to 78.1%. Of nearly 500 medications it ordered, 99.8% were correct for indication and safety.
This isn't a chatbot. It takes patient histories, orders and interprets labs and imaging, generates differential diagnoses, and proposes treatment plans — all within a FHIR-compliant EHR environment.
The researchers are careful to note this is not a physician replacement. It wasn't tested in live clinical environments, and gaps remain — particularly around nuanced antibiotic selection. Human supervision remains essential.
But for those of us building healthcare products, the implications are hard to ignore.
We've spent years thinking about AI as a layer on top of clinical workflows — a tool that surfaces information, flags risks, supports decisions. MIRA suggests something more fundamental is possible: an agent that doesn't just support the workflow, but operates within it.
The questions this raises for product builders aren't just technical. They're about trust, governance, liability, and what it actually means to deploy something like this responsibly in a real health system. The gap between a sandboxed research environment and a live ED is enormous — and bridging it will require as much product and organizational thinking as it does engineering.
This is the most interesting space in healthcare right now. What are you watching?