AWS Builder Center

Builder Center

Building AeroGraph: Autonomous Multi-Agent Air Quality Intelligence with AWS Strands Agents SDK & Amazon Bedrock

How we built a production-grade, local-first multi-agent system that shifts air quality monitoring from reactive metrics to predictive early warnings and dynamic, role-tailored operational playbooks.

The Problem: Why Traditional Air Quality Dashboards Fail

Most environmental intelligence dashboards suffer from three fundamental architectural flaws:
  1. They are strictly reactive: They report a dangerous AQI number hours after pollutants have already settled into urban lungs.
  1. They lack causal attribution: An AQI of 240 tells you the air is unhealthy, but doesn't tell you why—is it localized vehicular emissions, stagnant secondary aerosols, or transboundary stubble burning?
  1. They offer generic advice: A blanket public health notification like "Limit outdoor activity" is practically useless for school administrators, HVAC facility operators, or municipal anti-smog squads who need operational directives.
To solve this, we engineered AeroGraph: a predictive, multi-agent environmental intelligence platform powered by the AWS Strands Agents SDK and Amazon Bedrock.


High-Level Architecture

Instead of relying on a monolithic prompt that hallucinates generic recommendations, AeroGraph utilizes an agentic pipeline where specialized agents collaborate across deterministic and generative stages:
  1. Telemetry & Forecasting Layer: Real-time environmental metrics (PM2.5, PM10, NO2, Wind Vector, Humidity) are ingested alongside a multi-step Scikit-Learn time-series forecast predicting pollutant surge deltas at $t+1, t+2,$ and $t+3$ hours.
  1. Sentinel Agent (AWS Strands SDK): Continuously evaluates incoming telemetry against predictive thresholds. If a delta exceeds $\ge 15$ AQI points within 3 hours, it triggers an AI Early-Warning surge alert before ambient air degrades.
  1. Diagnostic Agent: Performs chemical fingerprinting:
  • Computes $\text{PM}{2.5} / \text{PM}{10}$ ratios to distinguish fine combustion aerosols from coarse fugitive dust.
  • Calculates upwind trajectory vectors to attribute emissions to localized stagnation or regional transport corridors.
  1. Playbook Planner Agent (Amazon Bedrock): Translates diagnostic telemetry and forecast vectors into 4 concurrent, persona-tailored operational playbooks using structured output schemas.
  1. Deterministic Safety Clamps: Post-LLM code clamps enforce immutable public health rules (e.g., if AQI > 250, N95 respirators and indoor HVAC loops are strictly required, regardless of generative variations).


Orchestrating with the AWS Strands Agents SDK

We adopted the open-source AWS Strands Agents SDK ( strands-agents ) to manage the state machine and agent lifecycle locally. The framework allows each agent to operate with clear boundaries:
Python
from strands_agents import Agent, Pipeline



# Sentinel Agent evaluates forecast deltas

sentinel = Agent(

    name="Sentinel",

    role="Detect imminent AQI surge deltas from ML telemetry stream"

)



# Diagnostic Agent attributes pollution fingerprints

diagnostic = Agent(

    name="Diagnostic",

    role="Determine chemical source profile and trajectory vectors"

)



# Planner Agent queries Amazon Bedrock for playbook synthesis

planner = Agent(

    name="PlaybookPlanner",

    role="Synthesize MultiPersonaPlaybook using strict Pydantic schemas"

)



pipeline = Pipeline(stages=[sentinel, diagnostic, planner])

# ai-agents# ai-assisted-workflows# amazon-bedrock-agents# amazon-nova# ai-ml
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article