It’s 3 AM in an extremely busy emergency department. A man in his mid-30s is lurching over with pain in his lower right abdomen, and he’s worried because his friend, who experienced similar pain, was recently diagnosed with cancer. A trustworthy doctor would ease his nerves and ask all of the right questions. He would then order blood tests and conduct an ultrasound, identify appendicitis, prescribe antibiotics, and schedule him for surgery. Now imagine an AI agent handling that entire process. That’s exactly what MIRA (Medical Intelligence for Reasoning and Action) was built to do, and a new study published by Dyke Ferber, Jakob Nikolas Kather, and colleagues in Germany and the US proves how close it comes.

LLMs (large language models) have been able to pass medical exams and ace multiple-choice tests, but diagnosing and conversing with a patient demands a whole different set of skills. Clinical work mainly relies on EHRs (electronic health records), serving as a digital record of a patient’s medical history and doctors’ orders. MIRA, powered by OpenAI’s GPT-4o, operates in a controlled, sandbox EHR environment and conducts full end-to-end clinical workflows. It has eleven medical tools covering around 85,000 possible actions: ordering labs, prescribing drugs, jotting patient history, admitting patients, and more. The plethora of options was also programmed to adhere to up-to-date medical codes, so that MIRA couldn’t order false tests or invent actions that don’t exist in real life.

To test how accurate MIRA’s decision-making was, researchers benchmarked the agent across 574 emergency cases from MIMIC-IV, a database containing health data from Beth Israel Deaconess Medical Center. Researchers selected data on eight prominent conditions, including pancreatitis, appendicitis, and pneumonia. In a trial against human doctors, MIRA achieved 87.8% diagnostic accuracy, which exceeded board-certified physicians (78.1%) and a team of residents and specialists (71.1%). MIRA’s performance was exceptional for conditions like pancreatitis, with 95.2% accuracy compared to doctors' 78.6%. Additionally, when recommending procedures or surgeries, MIRA requested 53.5% of the procedures patients actually received, compared to 38.3% for doctors.

MIRA consistently followed a structured clinical sequence: history, examination, labs, imaging, and lastly treatment. Its prescriptions matched treatment standards by +35 percentage points over those of board-certified physicians. MIRA recorded no severe drug-drug interactions or allergy mismatches when prescribing. Steady performance was also maintained when prompt-level patient biases were tested, like anxiety or stubbornness in accepting a diagnosis and treatment.

While MIRA matched or exceeded experienced doctors, specialists, and physicians across tasks, researchers emphasize that agents are meant to support professionals, not replace them. Much of MIRA’s value may lie in the work surrounding diagnosis. The tools it uses to reason through cases (writing prescriptions, admitting patients, placing imaging orders) are the administrative tasks that take up a lot of a physician’s day in EHRs. Under human oversight, MIRA could handle documentation, so the doctor in that 3 AM emergency department can spend less time on a screen and more time addressing patients face-to-face.