The anatomy of a personal health agent

Google's research arm just published one of the most detailed looks yet at what it takes to build an AI that can actually help you with your health, not just recite WebMD. The project, called the Personal Health Agent, skips the usual approach of throwing one giant model at every question. Instead it splits the job into three specialists: a data science agent that crunches your Fitbit numbers, a domain expert agent that pulls from sources like the NCBI database, and a health coach agent trained on techniques like motivational interviewing.

The reasoning is pretty intuitive once you see it laid out. Asking "how many hours did I sleep last month" is a statistics problem. Asking "how do I sleep better" is a behavior-change problem. Asking about symptoms is a clinical-knowledge problem. Google found this out the hard way, apparently, after mining more than 1,300 real health queries from forums and surveying over 500 users, which pointed to four recurring needs: general health info, personal data interpretation, wellness advice, and symptom checks. That's a lot of different skill sets for one model to fake convincingly.

What makes this more than a thought experiment is the scale of the evaluation. Google ran the system against real wearables data, blood tests, and questionnaires from roughly 1,200 consenting participants in an IRB-reviewed study, then had health experts and end-users grade it across 10 benchmark tasks, racking up over 7,000 annotations and 1,100 hours of human review. The data science agent scored 75.6% on generating sound analysis plans versus 53.7% for a plain base model. The domain expert agent beat baseline on clinician-graded relevance and consumer-rated trust. The coaching agent won out with both users and professional coaches, mainly because it didn't just talk — it gave usable advice.

The most interesting result, though, is what happened when Google combined all three under an orchestrator that decides which agent leads and which support, then merges their output into one answer. That setup beat not only a single all-in-one agent but also a simpler version where all three specialists just answered in parallel without coordination. In other words, the orchestration itself, the part that mimics how a real care team divides labor and checks each other's work, is where the actual gains showed up.

Google is careful to frame this as research, not a product announcement, and there's no Fitbit app update coming next week. But the direction is clear: they're betting that health AI improves not by making one model smarter, but by making several narrower models work together like a clinical team would.