Google's CoDaS Turns 9,279 Wearable Users' Data Into Clinical Research in 8 Hours

Wearables spit out millions of heart rate, step, and sleep readings per user, but turning that firehose into something a clinician can actually use is grinding, manual work. A team from Google Research, Google DeepMind, and MIT just released CoDaS, a multi-agent system that automates the tedious middle of that pipeline: proposing hypotheses, running the stats, arguing against itself, and drafting a report. What it found is worth reading, and what it refused to report may be more instructive.

Why handcrafted biomarkers keep stalling

Current wearable biomarker research tends to look like a grad student staring at a Jupyter notebook for six months. Existing approaches rely on handcrafted features within narrowly defined disease settings, limiting generalizability across populations, devices, and clinical contexts. The combinatorial feature space is huge, spurious correlations are everywhere, and it is very easy to accidentally include a transformation of the label you are trying to predict.

LLM agents that write and execute code have started to encroach on this workflow, though most operate in symbolic, textual, or algorithmic domains rather than on real-world, high-dimensional physiological time series with clinical or validated survey endpoints. CoDaS is built specifically for that messier setting.

Six phases, two model tiers, one Fact Sheet

CoDaS runs as a set of AI agents coordinated through a shared workflow, using Gemini-3.1 Pro Preview for research-intensive reasoning and code generation and Gemini-3 Flash Preview for repeated lower-latency tasks. An Orchestrator agent dispatches specialized sub-agents through six phases:

1. Data profiling and literature grounding. A Scout agent maps the dataset schema, while a Researcher ensemble pulls in known biomarkers and mechanistic pathways from the literature.
2. Parallel statistical and ML exploration. Deterministic runners execute correlation tests and cross-validated Ridge or tree models while LLM interpreters propose composite features like ratios and variability indices.