Nature: MIRA and AMIE as medical AI agents for clinical workflows
Nature has published two papers on medical AI agents: MIRA manages a patient case through a virtual emergency department, while AMIE plans treatment across three outpatient visits
Neither system was tested as a reference tool for physicians, but as a participant in the clinical process. MIRA operated in an electronic health record sandbox, asked questions, ordered tests, and proposed an admission plan. AMIE conducted a dialogue with a standardized patient and built a treatment plan across multiple visits.
On 17 June, Nature published two papers on medical AI agents. Medical AI has usually been tested on questions: make a diagnosis from a description, choose an answer on a test, write a note for a doctor. That is useful, but it resembles an exam taken at a desk. In a real hospital, a physician does more: questions the patient, reviews the medical history, orders tests, waits for results, revises the hypothesis, prescribes medications, and decides whether to send the person home or admit them to the hospital.
A few days earlier, Nature Medicine compared medical AI systems on real physician queries: 100 de-identified questions from NYU Langone, six systems, and blinded evaluation by 12 clinicians. The new Nature papers take on the next segment: the sequence of clinical actions.
The first article in Nature describes MIRA, Medical Intelligence for Reasoning and Action. It is an agent embedded in a virtual EHR, that is, an electronic health record sandbox. It has a set of permitted actions: take a history, order laboratory tests, microbiology, and imaging, read the results, produce a differential diagnosis, prescribe medications, a procedure, or surgery, and propose hospitalization.
The authors ran MIRA on 574 cases from MIMIC-IV, a large de-identified database of real intensive care patients and hospital records. In a head-to-head comparison on 311 cases, MIRA made the correct diagnosis in 87.8% of cases, while four board-certified physicians did so in 78.1%. The system also looked careful on medications: out of 468 prescriptions, 467 were rated as a "clinically useful and correct dosing instruction." Most often, the system made mistakes in the route of administration: intravenous, oral, subcutaneous, and so on.
The second article approaches the problem from the other side. Google DeepMind and Google Research expanded AMIE, Articulate Medical Intelligence Explorer, for longitudinal disease management. The system did not operate in the emergency department, but in a virtual outpatient scenario: 100 clinical cases, three visits, a standardized patient, comparison with 21 primary care physicians, and blinded specialist evaluation.
AMIE drew on UK NICE clinical guidelines, BMJ Best Practice, and drug references so that its treatment plan rested on external sources. In specialists' evaluation, it was no worse than physicians in reasoning about patient management and performed better on the accuracy of prescriptions, tests, and agreement with guidelines. Separately, the authors created RxQA, a test based on US and UK drug formularies that checks safe drug selection.
The phrase "AI outperformed doctors" erases the main limitation. Both papers tested simulations. MIRA lived in a controlled sandbox. AMIE spoke with standardized patients, and the three visits were spaced about two days apart. Ordinary clinical practice adds queues, rescheduling, forgotten tests, conflicting complaints, and economic pressure on decisions.
That boundary is where the real result lies: these papers test medical AI as a working system. Such a system maintains context, calls tools, and leaves structured actions inside a medical interface.
Medicine depends both on physician knowledge and on hospital throughput: who will gather the data, not miss contraindications, cross-check guidelines, notice changes between visits, and do it equally well across a thousand clinics. So far, MIRA and AMIE have been validated inside simulations. The next stage is prospective studies in real clinical pathways.