AWS Details AtScale Semantic Layer for Agents

AWS describes an AtScale-based semantic layer for agentic AI that connects a Strands agent hosted on Amazon Bedrock AgentCore to customer data in Amazon Redshift. The walkthrough follows earlier posts that used Stardog and Denodo for the same Customer 360 pattern. It uses AtScale C2026.7 and Anthropic Claude Sonnet 4.6 as the foundation model that plans and writes queries.

AtScale represents the data as a dimensional model of dimensions, hierarchies, facts and measures, rather than Stardog’s graph of triples or Denodo’s views. Modelers define it in Semantic Modeling Language (SML), an open YAML-based specification whose files are versioned in Git. A single deployed model exposes the same definitions through SQL, XML for Analysis (XMLA) and the Model Context Protocol (MCP), so the Strands agent and existing business intelligence tools can receive the same computed answers.

The example is a retail company whose customer, order, product and rewards information spans systems. Unlike the earlier examples, AtScale plans each dimensional query as native warehouse SQL and sends it to one engine; the walkthrough consolidates all analytical data in Amazon Redshift Serverless. Amazon Aurora PostgreSQL remains the operational source of record, with AWS Database Migration Service or Aurora zero-ETL replication keeping the Redshift copy current. The model maps to eight Redshift tables: customer, address, credit_card, rewards_account, purchase, product, category and vendor. The article identifies nine columns of interest; customer records link to purchases through cid, purchases link to products through pid, and products link to categories through dept.

A sample request asks which of the top 20 customers by lifetime spend have not enrolled in rewards. The model has seven dimensions, including customer, address, rewards, product, date and vendor dimensions, with a product-to-category hierarchy. Its fact_purchase dataset supports measures such as total_revenue, calculated as SUM(price quantity), order_count as COUNT(id), and units_sold as SUM(quantity). Calculated measures such as customer_lifetime_value, big_spender and spend_tier apply business logic at query time. The post gives an example threshold change from $10,000 to $15,000: the definition is updated in a YAML file, reviewed as a pull request and deployed through an AtScale CI hook.

Amazon Bedrock AgentCore handles the agent runtime and credential flow. On the inbound path, AgentCore Gateway validates the human user’s JSON Web Token against the enterprise identity provider and routes the request to the agent. On the outbound path, the agent calls AtScale’s MCP server through Gateway, while AgentCore Identity injects the credential. AtScale accepts a bearer token for development and single-identity use, or an OAuth 2.0 access token from its bundled Keycloak instance for production. AtScale resolves the caller to a role and applies row security and other model access rules; the agent itself does not see the credentials.

AtScale’s AI Computation Engine parses the agent’s dimensional SQL, resolves its references to model datasets, applies the caller’s row security and plans execution against Redshift. Joins run as native SQL in Redshift rather than being pre-materialized in an ETL pipeline. The post also describes AtScale’s autonomous aggregate management, which monitors query patterns and creates warehouse summary tables when a pattern is expensive enough to justify one; matching later queries can use those aggregates transparently. The deployed model can also serve SQL clients, XMLA tools such as Excel and Power BI, and MCP clients. The article’s walkthrough includes modeling, query planning, identity and access controls, and deployment on Amazon EKS using AtScale’s Helm chart and Kubernetes blueprints.

AWS describes an AtScale-based semantic layer for agentic AI that connects a Strands agent hosted on Amazon Bedrock AgentCore to customer data in Amazon Redshift. The walkthrough follows earlier posts that used Stardog and Denodo for the same Customer 360 pattern. It uses AtScale C2026.7 and Anthropic Claude Sonnet 4.6 as the foundation model that plans and writes queries.

AtScale represents the data as a dimensional model of dimensions, hierarchies, facts and measures, rather than Stardog’s graph of triples or Denodo’s views. Modelers define it in Semantic Modeling Language (SML), an open YAML-based specification whose files are versioned in Git. A single deployed model exposes the same definitions through SQL, XML for Analysis (XMLA) and the Model Context Protocol (MCP), so the Strands agent and existing business intelligence tools can receive the same computed answers.

The example is a retail company whose customer, order, product and rewards information spans systems. Unlike the earlier examples, AtScale plans each dimensional query as native warehouse SQL and sends it to one engine; the walkthrough consolidates all analytical data in Amazon Redshift Serverless. Amazon Aurora PostgreSQL remains the operational source of record, with AWS Database Migration Service or Aurora zero-ETL replication keeping the Redshift copy current. The model maps to eight Redshift tables: customer, address, credit_card, rewards_account, purchase, product, category and vendor. The article identifies nine columns of interest; customer records link to purchases through cid, purchases link to products through pid, and products link to categories through dept.

A sample request asks which of the top 20 customers by lifetime spend have not enrolled in rewards. The model has seven dimensions, including customer, address, rewards, product, date and vendor dimensions, with a product-to-category hierarchy. Its fact_purchase dataset supports measures such as total_revenue, calculated as SUM(price quantity), order_count as COUNT(id), and units_sold as SUM(quantity). Calculated measures such as customer_lifetime_value, big_spender and spend_tier apply business logic at query time. The post gives an example threshold change from $10,000 to $15,000: the definition is updated in a YAML file, reviewed as a pull request and deployed through an AtScale CI hook.

Amazon Bedrock AgentCore handles the agent runtime and credential flow. On the inbound path, AgentCore Gateway validates the human user’s JSON Web Token against the enterprise identity provider and routes the request to the agent. On the outbound path, the agent calls AtScale’s MCP server through Gateway, while AgentCore Identity injects the credential. AtScale accepts a bearer token for development and single-identity use, or an OAuth 2.0 access token from its bundled Keycloak instance for production. AtScale resolves the caller to a role and applies row security and other model access rules; the agent itself does not see the credentials.

AtScale’s AI Computation Engine parses the agent’s dimensional SQL, resolves its references to model datasets, applies the caller’s row security and plans execution against Redshift. Joins run as native SQL in Redshift rather than being pre-materialized in an ETL pipeline. The post also describes AtScale’s autonomous aggregate management, which monitors query patterns and creates warehouse summary tables when a pattern is expensive enough to justify one; matching later queries can use those aggregates transparently. The deployed model can also serve SQL clients, XMLA tools such as Excel and Power BI, and MCP clients. The article’s walkthrough includes modeling, query planning, identity and access controls, and deployment on Amazon EKS using AtScale’s Helm chart and Kubernetes blueprints.