Towards Autonomous Environmental Health: AERIS, a Research Agent for Air-Pollution Exposure Assessment

Air pollution harms human health and contributes to millions of deaths each year, yet many public-health studies omit it even when they record the locations and times needed to estimate exposure. AERIS (Airborne Exposure Research & Investigation System) equips command-line agents with the tools and instructions to compute exposure from public air-quality measurements and study location data. To test the harness on a real-world, end-to-end, long-horizon research task, we introduce _source-substitution reanalysis_: AERIS rebuilds air-pollution datasets from public measurements, links them to published health outcomes, and reruns the statistical analyses. A structured literature search identified nine published studies as candidates for reanalysis based on open health-outcome data, sufficient methodological detail, and adequate air-quality coverage across the required locations and periods. Across these studies, AERIS processed 1.85 million air-quality monitor records into 20,568 area-time exposure estimates and 14 pollution–health comparisons across four pollutants. It matched the direction of all seven published positive associations; the median study-level difference between AERIS and published log effect ratios was 0.040. We then ran autonomous reanalyses as a robustness control and repeated them with the same model stripped of AERIS tools, instructions, and exposure engine as a harness ablation. Self-driving AERIS completed 23 of 27 attempts, compared with 7 of 27 for the raw agent, although its median log-effect difference rose to 0.169 without researcher review. AERIS turns air-pollution exposure assessment into a reviewable agent workflow and opens a path toward autonomous environmental-health research.