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

NeurIPS 2026 Workshop on GenAI for Health (GenAI4Health)
Romain Lacombe1· Ziv Lautman1· Nadav Efrat2· Yougeng Lu1· Michael P. Snyder1
1Stanford University   ·   2University of Illinois Urbana-Champaign
{rlacombe, lautman, yougengl, mpsnyder}@stanford.edu, nefrat2@illinois.edu
Read the Paper (PDF) Code & MCP Server Dataset & Agent Traces BibTeX
18,919 → 9
Studies Screened & Reanalyzed
Two-route PubMed & OpenAlex screen down to 21 candidates, 14 eligible attempts, and 9 fitted same-pollutant reanalyses.
1.04×
Median Effect Calibration
Median study-level |Δ log R| = 0.040 (Pearson r = 0.834); 7/7 published positive directions retained across 14 contrasts.
23 / 27
Zero-Intervention Autonomy
23/27 autonomous reanalyses completed across 3 campaigns (up to 9/9) vs. 7/27 for the same model without AERIS.
1.85M
Monitor Records Harmonized
1,847,519 station observations across 705 monitor-study associations aggregated into 20,568 area-time exposure estimates.
AERIS Architecture and Operating Modes
Figure 1 — AERIS Architecture & Operating Modes. In interactive or self-driving mode, a command-line agent (Codex, Claude Code, Antigravity, or any MCP host) loads study data, queries ambient sources, and computes exposure through the AERIS skill and local MCP tool server while preserving causal-design and provenance records in a durable .aeris workspace.

Overview

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 dates needed to estimate exposure. AERIS (Airborne Exposure Research & Investigation System) is a provider-neutral research harness and Model Context Protocol (MCP) server that gives coding agents the domain tools to link cohort locations and time windows to public air-quality records across fixed-location monitors, administrative polygons, and GPS trajectories.

To evaluate AERIS on a complete scientific workflow rather than isolated tool calls, we introduce source-substitution reanalysis: reconstructing a published study's pollution exposure from independent public measurements, linking it to the study's health outcomes, and re-estimating the health effects without access to the original authors' pollution tables.

Source-Substitution Reanalysis & Autonomy Controls

Screening 18,919 papers yielded nine published studies with open health data, reproducible statistical models, and sufficient air-quality data coverage. With researcher review, AERIS matched the direction of all seven published positive associations across 14 contrasts, with a median study-level log-effect difference of 0.040 (≈1.04× ratio; Pearson r = 0.834).

Study-Selection Funnel
Figure 2 — Study-Selection Funnel. Two parallel discovery routes across 18,919 PubMed and OpenAlex records yielded 21 unique candidates, 14 eligible attempts, and 9 completed same-pollutant reanalyses.
Effect Calibration across 14 Contrasts
Figure 3 — Effect-Size Calibration (14 Contrasts). Published vs. AERIS odds and rate ratios on log2 axes (median |Δ log R| = 0.040, geometric mean ratio 0.998, Pearson r = 0.834).

In three independent zero-intervention campaigns (27 study attempts), self-driving AERIS completed 23 of 27 reanalyses (8, 9, and 6 per campaign) compared with 7 of 27 (3, 3, and 1) for the same model without the AERIS harness—a 3.3× completion lift—though removing researcher review raised the median log-effect difference from 0.040 to 0.169.

Autonomy and Harness Ablation Comparison Across Campaigns
Figure 4 — Autonomy & Harness Ablation Comparison. (a) Fitted studies completed per nine-study campaign across AERIS with researcher review (9/9), self-driving AERIS (7.7 ± 1.5; campaigns C1–C3: 8, 9, 6), and the raw-agent ablation (2.3 ± 1.2; campaigns C1–C3: 3, 3, 1). (b) Pooled median study-level absolute log-effect difference |Δ log R| (left) and study-level Pearson correlation r by campaign (right).
Evaluation Metric AERIS + Researcher Review Self-Driving AERIS (3 Campaigns) Raw Agent Without AERIS (3 Campaigns)
Fitted Studies per Campaign 9 / 9 (100%) 7.7 ± 1.5 (23 / 27; C1–C3: 8, 9, 6) 2.3 ± 1.2 (7 / 27; C1–C3: 3, 3, 1)
AERIS Exposure & Provenance Record 9 / 9 26 / 27 0 / 27
Published Positive Direction Retained 7 / 7 11 / 16 3 / 4
Published Null / No-Evidence Retained 4 / 7 (all OECD) 14 / 21 3 / 3
Median Study-Level |Δ log R| 0.040 (~1.04×) 0.169 (~1.18×) 0.073 (7 fits only; 20 stopped)
Study-Level Pearson r by Campaign 0.834 0.507 · 0.091 · −0.170 0.016 · −0.675 · —

Quickstart

git clone https://github.com/rlacombe/aeris.git && cd aeris
pip install -e ".[all]"
export OPENAQ_API_KEY="your_openaq_key"

# Launch with Codex, Claude Code, or Antigravity
./aeris.sh --agent codex

Citation

@inproceedings{lacombe2026aeris,
  title     = {Towards Autonomous Environmental Health: AERIS, a Research Agent for Air-Pollution Exposure Assessment},
  author    = {Lacombe, Romain and Lautman, Ziv and Efrat, Nadav and Lu, Yougeng and Snyder, Michael P.},
  booktitle = {NeurIPS 2026 Workshop on GenAI for Health (GenAI4Health)},
  year      = {2026},
  url       = {https://rlacombe.github.io/aeris/}
}