A personal health agent can pass a code test, produce an answer users prefer and hold a well-rated coaching conversation. Each result answers a different question about its usefulness. Moving between them requires care, especially when the subject is health.

In “The anatomy of a personal health agent”, published in Nature on 21 October 2026, A. Ali Heydari and colleagues describe a system with three specialist components: one for data analysis, one for health-domain interpretation and one for coaching. A coordinating component assigns work and combines their outputs. The researchers built the system around the Gemini model family.

The design combines language-model responses with tools and personal data. Its inputs include wearable time series and health questionnaires; medical-record inputs in this study were restricted to blood-test results because of data-access constraints.

Three evaluations show how the researchers assessed parts of that system. This analysis covers selected accessible main-text passages, checked on 6 November. The available source extraction was truncated before the complete integrated results, Discussion, Methods and supplements. Recruitment, full rating rubrics and follow-up remain unverified here. That access limit bounds this interpretation; it does not establish what the unseen sections contain.

Code tests check specified computational behavior

The data-science component writes an analysis plan, generates Python code and uses a correction loop to respond to execution errors. In the code-generation evaluation, described with Figure 2a, seven data-science experts created 173 unit tests across 25 challenging query–plan pairs.

A unit test checks whether code behaves as specified for a particular case. The authors report that the specialist component passed more tests on its first attempt than base Gemini, with further improvement after iterative correction.

The counts describe the benchmark’s construction. The 173 tests sit within 25 query–plan pairs; they are not 173 patients or independent health recommendations. The result concerns code performing the operations those tests check. Whether the analysis plan asks the right question, or the resulting advice helps a person, requires a different assessment.

Users judged some answer qualities more favorably

The health-domain component was evaluated using 50 curated personalized health questions and 17 end-user raters. In “Personalized answers to health queries” and Figure 3b, the authors describe a blinded comparison with baseline Gemini responses, using yes-or-no rubric judgments and side-by-side rankings.

Users preferred the specialist’s answers on personalization and contextualization, and on the credibility of citations. The authors also report higher rubric-rated trustworthiness. Relevance to the question and groundedness in the data were comparable between the systems.

That combination matters. An answer can feel more tailored while receiving similar judgments for relevance and use of the underlying data. Citation credibility is also a judgment made in this evaluation; it does not establish that every citation was independently checked. Likewise, rated trustworthiness describes the assessment, rather than medically verified safety.

The question and rater counts do not reveal the total number of ratings or how questions were allocated. They also do not establish that participants judged answers based on their own health records.

Coaching ratings include an advantage for the baseline

For the end-user coaching evaluation, reported with Figure 4b, 31 participants took part in blinded, counterbalanced sessions comparing the coaching component with a baseline built on the same base Gemini model. Counterbalancing varies the order in which people encounter the systems.

Questionnaires assessed usability, the working relationship with the coach, motivational interviewing and goal setting. The authors report advantages for the coaching component in conversation flow, motivational interviewing and handling feedback.

The baseline scored higher on progress measurement. Here, progress measurement is a rated dimension of the coaching evaluation. It does not mean baseline users made greater health progress.

These observations concern participants’ experience and judgments of the conversations. Sustained behavior change and health benefit are different endpoints. The selected coaching passages alone cannot establish that connection.

Reviewers also identified potential harm in summaries

The accessible text also contains a caution beyond these three examples. In “Multimodal clinical health summaries”, the authors report that expert safety reviewers identified potential harm in generated summaries. This is a reviewer-assessed concern about the content; it should not be read as a count of observed patient injuries.