AWS and Strategy outlined a customer-data workflow that connects a Strands agent to Strategy Mosaic while preserving caller-specific access controls.
AWS published a technical walkthrough on October 20, 2026, showing how Strategy Mosaic can serve as a governed data layer for an AI agent running on Amazon Bedrock AgentCore. The example builds a customer 360 view from Amazon Redshift and Amazon Aurora, linking two Mosaic models without an ETL step in the demonstrated setup.
A Strands agent hosted on AgentCore reaches Mosaic through its Model Context Protocol (MCP) server. The agent first discovers accessible projects and certified models, then retrieves the model’s business definitions before issuing a read-only query. Rather than asking the model to invent joins or business calculations, Mosaic supplies the relationships, metrics and permitted data paths. Strategy says Mosaic can also serve definitions to other tools through JDBC/ODBC over Trino, XMLA for Power BI and Excel, and REST for custom applications.
The walkthrough separates operational customer records and enrollment information in Aurora from purchase analytics in Redshift. Mosaic links the sources on shared customer attributes and plans queries across them. In one example, the agent ranks customers by lifetime spending and identifies those without rewards accounts. The post also describes using the same model for questions about customers in a particular state or top-selling products.
Identity and permissions are applied to the human caller rather than replaced with a shared service identity. The article describes row-level and object-level security, including a test in which a fraud analyst can access Social Security number data while a marketing analyst cannot. In a separate check, totals from the agent matched results obtained from the same certified model through an Excel PivotTable.
The post is the third entry in an AWS series comparing semantic-layer approaches for agentic analytics; earlier examples used Stardog and Denodo. For this Strategy Mosaic implementation, the authors also discuss choosing live-source or in-memory execution and exporting definitions in Apache Ossie format. The article documents an architecture and example workflow, not a general availability announcement for a new product.