How we are building the personal health coach

Google Research just pulled back the curtain on something it's been quietly building for a while: an AI-powered personal health coach living inside the Fitbit app, powered by Gemini. Starting tomorrow, eligible US-based Fitbit Premium users on Android can opt into a public preview, with iOS support promised soon after. The pitch is straightforward — instead of a doctor telling you to lose weight or see a specialist and then leaving you to figure out the rest, the coach is meant to connect those dots itself, using your actual sleep, activity and physiological data.

What's actually interesting here is the engineering underneath a question as simple-sounding as 'do I sleep better after I exercise?' Google says answering that well requires real numerical reasoning over time-series health data, not just a chatbot bolted onto a fitness app. The system checks whether recent data exists, picks the right metrics, compares days against your personal baseline and against population data, and folds in what you've asked before. That's a lot of scaffolding for one sentence of advice.

To pull it off, Google built a multi-agent setup rather than relying on one big model to do everything. A conversational agent handles the back-and-forth and figures out what you actually want. A data science agent fetches and crunches the numbers, generating code on the fly when needed. And domain-expert agents — a fitness specialist, for instance — build and adjust actual coaching plans as your progress changes. Gemini's raw capability gets steered hard with custom evaluations built specifically around consumer health needs, because a general-purpose model answering health questions carelessly is a liability, not a feature.

Google leans heavily on the claim that this isn't just a model demo — it's been checked against something they call the SHARP framework, covering safety, helpfulness, accuracy, relevance and personalization. That involved over a million human annotations and more than 100,000 hours of evaluation from specialists in sleep medicine, cardiology, endocrinology, sports science and behavioral psychology. There's also a standing Consumer Health Advisory Panel and ongoing input from professional fitness coaches, plus feedback loops from thousands of users already testing features like Fitbit's Sleep and Symptom Checker Labs.

None of that guarantees the coach will actually be good — evaluation frameworks are easy to describe and hard to verify from outside. But the emphasis on expert oversight and iterative testing suggests Google knows this is a different risk category than a coding assistant or an image generator. Health advice that's wrong or overconfident has real consequences, and Google seems, at least on paper, to be treating that seriously rather than just shipping a slick Gemini wrapper and calling it a coach.