Nature Medicine Studies Show Agentic AI Can Aid Patient Care — But Isn't Clinic-Ready

Two experimental AI agents can take a patient history, order tests, weigh the results and draft a treatment plan — and in simulated cases they matched or beat physicians on diagnosis and management. But the researchers behind both systems, and the editors who spotlighted their work, land on the same conclusion: neither is ready to touch a real patient.

That is the throughline of a Nature Medicine research highlight published on 30 June 2026, which examined a pair of agentic AI systems, MIRA and AMIE, whose primary studies appeared together in Nature on 17 June. "Two studies show that agentic artificial intelligence (AI) models could aid decision-making at various stages of patient management, including diagnosis, treatment and hospital admission, but neither model is ready yet for real-world clinical use," wrote Karen O'Leary, an associate research analysis editor at Nature Medicine, in the highlight titled "Agents AMIE and MIRA advance medical AI capabilities."

The framing matters. Both systems move past the narrow, single-task chatbots that have dominated clinical AI toward autonomous "agents" that string together multi-step workflows — exactly the capability that would make them useful in a real clinic, and exactly the capability that makes them hard to validate.

MIRA: an agent inside a sealed medical record

MIRA — Medical Intelligence for Reasoning and Action — was built by Jakob Kather's group at TUD Dresden University of Technology and Heidelberg University, among other institutions, and described in the paper "Towards autonomous medical artificial intelligence agents" (Ferber, Hilgers, Höper et al.). It operates as an autonomous agent inside a sealed, virtual electronic health record, choosing from more than 85,000 options across eleven tools to order labs, microbiology and imaging, interpret results, generate differential diagnoses and write treatment plans that include prescriptions, procedures and hospital admissions.

The team tested MIRA on more than 500 real emergency department cases drawn from the public MIMIC-IV dataset, with a second AI agent role-playing the patient and sharing only what the actual medical record documented. In a head-to-head subset of 311 cases, MIRA reached the correct diagnosis 87.8% of the time versus 78.1% for a panel of experienced specialists. Blinded specialist reviewers found no dangerous drug interactions, no incorrect dosing for patients with impaired kidney function and no risky painkiller prescriptions in its recommendations. MIRA was nearly perfect at capturing a patient's current medications and, notably, did not miss a single case that required hospitalization.

Kather compared the technology to aviation autopilot. "These systems can support and relieve medical professionals by taking over routine tasks, but ultimate responsibility will always remain with the physicians," he told the Financial Times, adding that the work offers "a preview of how AI could transform medicine." His own team flagged the caveats: MIRA recommended care that "deviated from best practices" for a "small but non-zero" share of patients, and the simulated patient's answers may have been "more structured than real speech of patients in emergency departments."

AMIE: reasoning across multiple visits

Google's AMIE — Articulate Medical Intelligence Explorer — was described by Mike Schaekermann and colleagues in "Towards Conversational AI for Disease Management" (Liévin, Palepu, Weng et al.). Built on Gemini, it pairs a conversational agent that talks to the patient with a second, slower agent that cross-references the case against clinical guidelines and drug formularies, and it can reason continuously across multiple patient visits.

In a virtual clinical examination, AMIE was compared with 21 primary care physicians across 100 multi-visit case scenarios spanning five specialties, benchmarked against UK NICE guidance and BMJ Best Practice. AMIE matched physicians on management reasoning and outperformed them on the precision of treatments and investigations and on alignment with guidelines; at the first visit, its overall plan was rated appropriate in 95% of cases versus 72% for physicians. On RxQA, a new medication-reasoning benchmark verified by licensed pharmacists, AMIE beat physicians on the hardest questions — though scores stayed below 75% on easier ones for both sides.

The AMIE team called the work a milestone but stressed that neither the curated cases nor the text-only conversations reflect a real clinic; the system shows "promising capabilities" but is "not ready for real-world translation," with more work needed to address "latent reasoning errors" in its hidden reasoning steps.

Why It Matters

Agentic AI is the frontier of clinical decision support because it promises to do more than answer a question — it can act across a workflow, which is where physician time is actually consumed and where staffing shortages bite. But autonomy multiplies the safety surface. Both studies ran in simulation. Catherine Pope, a professor of medical sociology at the University of Oxford, told the FT the results sit "some remove from the messy, complex, human world of everyday healthcare." Julie Jacko, a health informatics professor at the University of Edinburgh, cautioned that many reported advantages reflected "precision and completeness of plans" rather than "clear differences in clinical correctness." Both papers also carry a validation asterisk: the MIMIC-IV data are public and may have leaked into model training, and AMIE's own experiments showed its elaborate scaffolding added little once a newer base model was swapped in — a warning that today's benchmarks may not describe tomorrow's systems.

What to Watch

The gap now is not raw capability but evidence. Prospective clinical trials of agentic pipelines remain scarce, and regulators have yet to settle how to classify systems that reason and act autonomously. Watch for the first prospective, real-world studies — MIRA's team has signaled clinical trials are next — and for regulatory guidance that defines validation standards for agents rather than static software. Until an agent is tested on real patients speaking in real, unstructured ways, "at least as good as doctors" remains a simulation result, not a clinical one.