Build a semantic layer for agentic AI on AWS with Strategy Mosaic and Amazon Bedrock AgentCore
Amazon Web Services Zhili Cheng ● Covered by 3 sources
AWS and Strategy showed a way to let an AI agent query customer data through a semantic layer. The twist: the security and joins stay in Mosaic, not in the prompt.
Based on reporting by Amazon Web Services, Zhili Cheng — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS and Strategy are using the same customer-360 pattern again, but this time the semantic layer is Strategy Mosaic. The setup sits on Amazon Bedrock AgentCore, with a Strands agent reaching Mosaic through its MCP endpoint instead of stitching facts together in the prompt. That matters because the joins, metric definitions, and per-user filters live inside the model, where they can be governed.
The walkthrough keeps the data split simple and realistic. Amazon Aurora PostgreSQL-compatible edition holds customer-side records such as addresses, credit cards, and rewards enrollment. Amazon Redshift holds the analytical side: purchases, products, categories, and vendors. Mosaic links two models, one for Customer and one for Purchase, on shared attributes like Customer and Time, then pushes source-native SQL to each system and joins the intermediate results inside Mosaic.
A representative query asks which of the top 20 customers by lifetime spend still have not enrolled in rewards. That pulls profile data from Aurora, spend from Redshift, and enrollment status from Aurora, all through definitions that sit in the semantic layer. The agent does not invent its own join logic. It traverses the links that exist in the model, and nothing else.
Mosaic’s modeling is dimensional rather than ontology-based or view-based. Attributes have ID and display forms, hierarchies roll up things like Category to Department and Address to City to State, and metrics can be base, advanced, or compound. The example base metric is Purchase Amount, defined as price times quantity, while named filter objects and level qualifiers let the modeler pin down business logic like “large orders” or a customer-level grain.
Security is part of the design, not an afterthought. The user’s identity starts at AgentCore Gateway, passes through the Strands agent, and reaches Mosaic as an OAuth token tied to the human user, not a service account. Mosaic then applies the Use MCP Server privilege, checks certified-model access, and enforces row-level and object-level security before planning the query. The post also says Mosaic can publish the same governed definitions to Power BI, Tableau, notebooks, custom apps, and export them in Apache Ossie format.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous part of agentic AI that actually matters: put the facts in the model, not in a chat prompt and a prayer. The industry keeps acting like the hard problem is making agents clever; most of the time it’s making them stop freelancing with your data.
Read more about this at: Amazon Web Services