Build a semantic layer for agentic AI on AWS with AtScale and Amazon Bedrock AgentCore
Amazon Web Services Frank Lapolito ● Covered by 3 sources
AWS shows AtScale as the semantic layer for agentic AI on Redshift. Same model now feeds the agent and BI tools through SQL, XMLA, and MCP.
Based on reporting by Amazon Web Services, Frank Lapolito — 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’s latest C360 walkthrough keeps the same basic idea running through the whole series: put business meaning in a semantic layer, then let the agent ask questions without guessing at joins. This time the layer is AtScale, not Stardog or Denodo. And the shape changes with it. Instead of triples or views, AtScale uses a dimensional model built in Semantic Modeling Language, an open YAML spec that lives in Git and can be exposed through SQL, XMLA, and the Model Context Protocol from one deployed model.
The reference stack is Amazon Bedrock AgentCore on the front end and Amazon Redshift Serverless on the data side. Claude Sonnet 4.6 plans the query. AgentCore hosts the Strands agent and handles the outbound credential. The semantic layer does the hard part: joins, metric definitions, filters, and row security live there, not inside the prompt. AWS is careful about identity too. The user’s JWT comes in through AgentCore Gateway, then the identity is carried through AgentCore Identity to AtScale, where the engine resolves the caller to a role and applies RBAC, row security, and perspectives at query time.
The sample domain is the familiar retail customer 360 setup: customer records, orders, products, rewards enrollments, and a question about which of the top 20 customers by lifetime spend still haven’t joined the rewards program. In this version, all of the analytical data sits in Amazon Redshift, while Amazon Aurora PostgreSQL stays the operational system of record. Replication keeps the warehouse copy current, either with AWS Database Migration Service or Amazon Aurora zero-ETL to Amazon Redshift. That matters because AtScale pushes each dimensional query down as native warehouse SQL to one engine per query. There’s no cross-engine federation at query time, and no ETL job trying to prebuild every answer in advance.
The model itself is laid out like a star schema, but every object is a YAML file. Datasets map the Redshift tables. Dimensions and hierarchies define customer, product, date, vendor, and rewards relationships. Facts and measures define revenue, order count, and units sold. Calculated measures like customer_lifetime_value and big_spender are evaluated at query time, so a threshold change goes into one file and rolls out through Git and AtScale’s CI hook. AtScale also watches query patterns and builds aggregates when it sees repeated expensive requests, which is a practical touch for agent traffic that tends to ask the same broad questions in slightly different words. The deployment piece is equally concrete: the post says AtScale can be run on Amazon EKS with the AtScale Helm chart and the AtScale K8S Blueprints repository, using the AtScale C2026.7 release as the source of truth.
The nice part here is that AWS isn’t pretending the agent is magical. It’s a query planner with manners. The semantic layer does the actual governance, and that’s exactly where it belongs.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of boring. Put the meaning in the model, not in a prompt, and suddenly “agentic AI” looks a lot less like a demo and a lot more like software. The open YAML bit is the best part: if the model can’t be reviewed in Git, it’s just vibes in a trench coat.
Read more about this at: Amazon Web Services