Building agentic AI applications with a modern data mesh strategy on AWS

This article demonstrates how to build governed agentic AI applications using a modern data mesh architecture on AWS that enforces fine-grained access control across multiple data layers.

• Replace vector databases with Amazon S3 Vectors for 90% cost reduction in moderate query-frequency workloads
• Use Amazon S3 Tables with Apache Iceberg and Lake Formation for row, column, and cell-level security on transactional data
• Expose data mesh as Model Context Protocol tools through AgentCore Gateway with Lambda-backed interceptors for deterministic access control
• Implement five overlapping governance layers: Athena workgroup cost controls, read-only IAM policies, Lake Formation fine-grained access, Gateway interceptors, and Bedrock Guardrails
• Enforce JWT scope-based tool authorization, dynamic tool filtering, and act-on-behalf identity propagation at the Gateway layer
• Apply guardrails at Gateway level rather than model inference to prevent prompt injection and harmful content in multi-step agent interactions

The architecture enables production-grade agentic AI deployment for regulated industries by implementing defense-in-depth governance at every data access decision point.