For more than a decade we have been collecting data from wearables at an incredible rate: steps, heart rate, heart rate variability, sleep, GPS traces, training sessions. Yet the step that matters most has always been the hardest one: going from numbers to a decision. What should this person do next week, given their data, their health history and their life? In high-performance sport we solve that problem with teams of people. For the general population, that team simply does not exist.
A paper just published in Nature by Heydari and colleagues at Google Research, “The anatomy of a personal health agent” (21 October 2026), is the most serious attempt I have seen so far to build that team in software. It is worth reading carefully, both for what it shows and for what it does not show, because it gives us a realistic preview of AI-generated, personalised activity prescriptions based on our own data.
What they built
The authors started from users rather than from technology. They analysed 1,370 real consumer health queries from search, Gemini and the Fitbit forum, surveyed 555 people using an experimental Fitbit Labs chatbot and ran a workshop with 14 experts. Four types of need emerged: general health knowledge, interpretation of personal data, wellness advice, and symptom assessment. These needs require very different skills — numerical precision and code for data questions, factual grounding for knowledge and symptoms, empathetic open-ended dialogue for coaching — and the team found that a single large language model struggled to balance them.
Their solution is a modular, multi-agent system — the Personal Health Agent (PHA) — built on Gemini 2.0, made of three specialists coordinated by an orchestrator:
- A data science agent that turns a vague question (“am I getting fitter?”) into an explicit analysis plan, then writes and runs Python code on the user’s wearable time series, including comparisons with population percentiles by age and sex.
- A domain expert agent that reasons over wearable data, blood tests and health history, using tools such as NCBI literature search, web search and population statistics to ground its answers.
- A health coach agent built around motivational interviewing: it identifies goals first, explores constraints, asks for feedback and only then offers a SMART recommendation.
The orchestrator decides which agent leads and which ones support, rewrites a targeted sub-question for each, checks the combined answer for coherence (and to avoid asking the user for data it already has), and keeps a memory of goals, barriers and preferences across turns.
What the evaluation shows
The evaluation is large: 11 benchmark tasks, more than 7,000 annotations and about 1,350 hours of work from end users and experts, grounded in the WEAR-ME study (1,165 Fitbit users with blood tests and questionnaires). A few results stood out for me.
Data analysis. The data science agent’s analysis plans scored 75.6% against 53.7% for the base model. The most striking gain was in checking whether there was enough data to answer the question in the first place: 66.5% of plans did so, against just 5.3% for the base model. Anyone who has worked with wearable data knows that missing days, non-wear periods and artefacts are where most bad conclusions come from. Its code passed 75.5% of expert-written unit tests at the first attempt (58.4% for the base model), rising to 79% with self-correction.
Health knowledge and context. On more than 2,000 board-style questions in endocrinology, cardiology, sleep medicine and strength and conditioning, the gain over the base model was small (83.6% vs 81.8%). Much bigger differences appeared when personal context mattered: end users rated 96.9% of the domain expert’s personalised answers as trustworthy against 38.7% for the base model, and five clinicians judged 65% of its multimodal health summaries useful against 30%.
Coaching. Users preferred the coach agent on five of six coaching themes, and expert coaches preferred it on all six. Interestingly, both groups gave the plain baseline a slight edge on progress measurement — a weakness the authors then addressed in the integrated system.
The integrated agent. In live multi-turn conversations, 45 users role-playing real WEAR-ME profiles ranked the orchestrated PHA first 46.7% of the time, against 28.9% for a parallel multi-agent design and 24.4% for a single agent. The experts were far more decisive: the PHA was ranked best in 66.7% of conversations, roughly five times as often as the single agent, with an overall quality score of 72.8% against 64.3% and 54.4%.
Why this matters for activity prescription
What struck me most is that the architecture mirrors how a good performance support team works. Someone looks at the data properly, someone interprets it in the context of the person’s health and history, and someone turns it into a plan the person will actually follow. Exercise prescription needs all three. Frequency, intensity, time and type are not chosen from a guideline table alone; they depend on where the person is starting from, what they can safely tolerate, and what they are willing and able to do.
Seen through that lens, the near-future pipeline is easy to imagine:
- Baseline from your own data: real habitual steps, active minutes, resting heart rate, HRV and sleep — computed, not guessed, and checked for missing data.
- Comparison and context: where you sit relative to people of your age and sex, and what your blood markers and conditions mean for what is safe.
- A negotiated dose: a first step that fits your constraints and motivation rather than a generic “150 minutes a week”.
- Progression: memory of what was agreed and what actually happened, so the next prescription builds on the last one.
The example conversation in the paper’s extended data shows this already. A user with type 2 diabetes, a BMI of about 36, around 5,500 steps per day and a recent hospital admission asks how to get back to 10,000 steps. Instead of prescribing the target, the agent explores why it matters to her, picks up symptoms that suggest possible sleep apnoea, recommends speaking to her doctor, and agrees on several short 10-minute walks per day as a starting point. That is, in essence, a graded prescription with a red-flag screen built in. Another example I liked: asked whether more activity leads to more deep sleep, the data science agent reported a correlation of about 0.14 and said plainly that the relationship was weak — exactly the kind of honest answer a good analyst gives and a lot of consumer apps do not.
The real opportunity is scale. One practitioner cannot review thousands of people’s data every day and adjust their plan. A system like this could make the first layer of individualised prescription and follow-up available to anyone wearing a device, and free human experts to focus on the people and decisions that need them.
The risks we should not gloss over
The authors are refreshingly upfront about limitations, and they describe the PHA as a research prototype rather than a clinical tool. Still, for anyone tempted to treat this as a ready-made prescription engine, these are the issues I would keep in mind.
- Safety at scale. In the expert safety review of the clinical summaries, 83% carried no or minimal potential for harm, but 15% were judged mildly harmful and 2% moderately to severely harmful. Two per cent sounds small until it is multiplied by millions of users.
- No outcome data. Everything measured is conversation quality and short-term task performance. There is no evidence yet that people change behaviour, get fitter or become healthier. For an activity prescription, the only test that ultimately matters is the dose–response over weeks and months.
- Data quality and generalisability. The data come from one device ecosystem (Fitbit), and consumer estimates of sleep stages or HRV have their own validity limits. Even the improved agent checked data sufficiency in only two-thirds of its plans, so one-third did not.
- Simulated users. In the live studies, participants role-played personas matched to their own profiles rather than seeking help for themselves, and parts of the evaluation relied on LLM-based autoraters. Both are reasonable research choices, but neither is the real world.
- Training load is not in scope. The system was not evaluated on periodisation, load management, injury risk, environmental stress such as heat, or athletic populations. Extending it to performance settings needs its own validation.
- Evidence quality. Open-web retrieval can surface weak or conflicting evidence. The authors themselves suggest restricting retrieval to curated sources and showing provenance next to every recommendation.
- Privacy, equity and cost. Real deployment needs proper privacy and security controls, fairness checks across populations, languages and literacy levels, and a cheaper architecture: the PHA averaged 6.5 model calls and about 205 seconds per response, against 3 calls and 36 seconds for a single agent.
- Who built and judged it. The work was funded by Alphabet, most authors are Google employees, and the system runs on Google models and Fitbit data. That does not invalidate the findings — and releasing the code and benchmark is a genuinely good step — but independent replication will matter.
- Coaching versus prescribing. The coach agent sometimes declines to give a concrete number and steers the user to decide for themselves. Good motivational practice, but people asking for an exercise prescription often need a clear, specific dose — and someone accountable for it.
What I would want to see next
Before trusting an AI-generated activity prescription, I would want to know four things: whether it checked the quality of the data before drawing conclusions, whether its evidence comes from curated sources it shows me, whether it recognises red flags and hands off to a clinician rather than reassuring, and whether its prescriptions are reviewed against what actually changed over time. The next generation of studies should be prospective, longitudinal and outcome-based, with people using the system for themselves, across devices and populations, and with clinicians and coaches in the loop.
Bottom line
This paper does not show that AI can prescribe exercise better than a qualified professional. It shows something more modest and, I think, more important: that splitting the job into data analysis, domain expertise and coaching, and orchestrating them carefully, produces answers that experts judge to be substantially better than a single general model. Personalised activity prescriptions built from our own data are no longer a distant idea. The challenge now is to prove that they are safe and that they actually change behaviour and health — and to make sure the humans who are accountable for those decisions stay firmly in the loop.
Reference: Heydari, A. A. et al. The anatomy of a personal health agent. Nature (2026). https://doi.org/10.1038/s41586-026-11161-8 (open access). Code: Google-Health/consumer-health-research.
You must be logged in to post a comment.