Google's personal health agent: experts liked it 17.1% more than users

Google researchers published a personal health agent paper in Nature that puts numbers on a familiar gap: clinicians and patients do not rate the same system the same way.

On complex queries, rating differences against baselines averaged 17.1% for health experts and 4.6% for end users. The paper, “The anatomy of a personal health agent,” went online Oct. 21, 2026, and appears in the Nov. 5, 2026 issue (Vol. 659, doi:10.1038/s41586-026-11161-8). Lead authors include A. Ali Heydari, Ken Gu, Vidya Srinivas, Hong Yu and Akshay Paruchuri of Google Research.

The ranges behind those averages were 5.7% to 39.0% for experts and 1.6% to 15.4% for end users.

Three specialists and an orchestrator

The system splits work across three specialist sub-agents: a data science agent for wearable and medical time series, a domain expert agent for health knowledge, and a health coach agent for goals and behavior change. An orchestrator decides who leads and who supports on each query.

The design grew out of four user journeys the authors mapped: general health knowledge, personal data interpretation, wellness advice and symptom assessment. The evaluation covered 11 benchmark tasks with more than 7,000 annotations and about 1,350 hours of effort from experts and end users.

The authors call the system a research prototype for consumer wellness, not a clinical diagnostic tool.

46.1% diagnostic accuracy, and a 2% harm flag

On a retrospective set of 2,000 real-world consumer health cases, the domain expert agent reached 46.1% top-one diagnostic accuracy, 4.7 points above a prior Gemini-based state of the art. The authors say the system was not built to diagnose.

An expert safety review of domain expert summaries found no or minimal potential harm in 83% of cases, mild potential harm in 15%, and moderate to severe potential harm in 2%.

Why the expert-user gap matters

Experts gave the multi-agent system higher overall quality scores than single-agent or parallel multi-agent baselines. End users also preferred it, but by a smaller margin. The paper treats that gap as a finding, not a footnote: conversation quality is what users see, and conversation quality is not the same as clinical soundness.

The authors point future work toward helping people log symptoms, spot patterns and prepare questions for clinicians, with prospective tests still needed.

What it means for you

If you use a health chatbot, treat it as a way to organize questions for a clinician, not as a diagnosis. This paper’s own accuracy and harm numbers argue for that limit. Bring the printout. Do not let the fluent answer stand in for a clinical visit.

Sources: Nature article; Google Research blog.