Build an Intelligent Data Quality Monitoring System with Strands Agents

In data‑driven enterprises, poor data quality leads to costly decision errors, as illustrated by an e‑commerce case where a missing refund handling step caused duplicated sales figures and wasted ad spend.

Data engineers must monitor the accuracy, completeness and timeliness of increasingly complex pipelines, often overseeing hundreds of models, dealing with multi‑table anomalies, and spending excessive manual effort on root‑cause analysis.

To address these pain points, the article proposes an intelligent data‑quality monitoring system that combines the open‑source Strands Agents framework, Amazon Bedrock large‑language models, dbt (data build tool) and Amazon Redshift.

AI Agents

AI Agents pair an LLM with tool integrations to automatically scan data pipelines, detect metric deviations, and trigger analysis. For example, a data‑lineage Agent can draw flow graphs and, when a KPI drifts, correlate multi‑dimensional signals to pinpoint the anomaly.

Strands Agents Framework

Strands Agents is a lightweight, model‑driven AI‑Agent framework that offloads planning and execution to the LLM, eliminating hard‑coded workflow orchestration. Its key features include model‑agnostic LLM support, orchestration‑free design, built‑in OpenTelemetry observability, and integration with Anthropic’s MCP tool for seamless access to hundreds of external tools.

dbt and Redshift

dbt provides SQL‑based data modeling, testing and documentation, and integrates tightly with Amazon Redshift. The article uses the TICKIT sample dataset to demonstrate intentional modeling errors (e.g., extracting minutes instead of hours, wrong commission rate, missing space in full‑name concatenation) and shows how dbt’s MCP server exposes metadata and test results.

Agent Logic and Workflow

The data‑quality detection Agent uses a hybrid architecture: direct calls to dbt‑MCP tools (build, test, list, show) ensure data integrity, while a separate workflow orchestrates deep analysis and report generation. The workflow consists of three tasks:

deep_analysis : uses the think tool to analyze failed tests from technical and business perspectives.

english_report_generation : produces a bilingual report with exact file paths and line numbers for fixes.

chinese_translation : translates the report while preserving technical details.

Sample Python code shows how to instantiate the MCP client, load the Bedrock model ("us.amazon.nova‑premier‑v1:0"), and invoke the Agent with a system prompt that defines its role as a data‑quality assistant for the TICKIT system.

python -m venv .venv
source .venv/bin/activate  # For macOS/Linux
pip install strands-agents strands-agents-tools dbt-core dbt-redshift

Configuration files (e.g., profiles.yml) are adjusted to point to the Redshift cluster and the TICKIT schema. The Agent then executes the dbt tools, captures results, and runs the workflow to generate precise, code‑level remediation suggestions.

Production Recommendations

Least‑privilege and input validation : restrict Agent and tool permissions, validate all inputs to prevent injection or over‑privilege.

Containerization and elastic deployment : package Strands Agents in Docker and run on Amazon Fargate or EKS for scalability and high availability.

Monitoring and audit logging : integrate OpenTelemetry and Amazon CloudWatch to track token usage, latency, error rates, and tool invocations; enable detailed audit logs for compliance.

Deep dbt integration : automate dbt build/test commands, leverage dbt metadata for lineage tracing, and combine with Agent analysis to achieve end‑to‑end automated detection and fix recommendations.

By following this guide, teams can move from manual, rule‑based data‑quality checks to an automated, AI‑enhanced workflow that quickly identifies, diagnoses and resolves data issues before they impact business decisions.