Build a semantic layer for agentic AI on AWS with Denodo and Amazon Bedrock AgentCore
Amazon Web Services Suntosh Murthy ● Covered by 3 sources
AWS shows how to let an agent query Aurora, Redshift, and on‑prem Postgres through Denodo instead of moving data around. The twist: the joins and access rules stay in the semantic layer, not in the prompt.
Based on reporting by Amazon Web Services, Suntosh Murthy — 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 is keeping the same customer 360 pattern from its earlier semantic-layer post, but this time the setup reaches beyond the comfort zone of an AWS-only stack. The new version uses Denodo Platform as the semantic layer and adds an on-premises PostgreSQL rewards database to the mix alongside Amazon Aurora and Amazon Redshift. That matters because the federation problem is no longer neatly contained inside an Amazon VPC. It now has to span places where enterprise data actually lives.
The example is simple enough to be useful: a retailer wants to know which of its top 20 customers by lifetime spend still haven’t joined the rewards program. Aurora holds customer profile data, Redshift holds order history, and the on-premises PostgreSQL cluster holds rewards enrollment. The shared key is cid, and there’s no foreign key bridging the systems. No ETL job is building a joined table behind the scenes either. Denodo resolves the joins at query time, deciding what to push down to each source and how to combine the results.
That semantic layer is built out of base views, a derived view, and a metric view. Base views map physical tables into Denodo Virtual DataPort. The customer_360 derived view then joins Aurora customer and address data with Aurora credit card data and the on-premises rewards table. The metric view defines reusable business logic once, so spend calculations don’t get rewritten every time another view needs them. And because the model is meant for agents, the security and metadata ride along with the data model, including tags for sensitive columns like ssn and card_num.
The runtime flow is equally opinionated. Amazon Bedrock AgentCore Gateway handles the inbound request, validates the user’s JWT against the enterprise identity provider, then routes the call to the hosted agent. On the way out, AgentCore Identity does an on-behalf-of token exchange so Denodo sees a scoped token for the same user. The Denodo MCP Server receives the VQL request, VDP validates the token, applies access policies, and sends rewritten SQL to Aurora, Redshift, and PostgreSQL. The model never owns the joins. That’s the point.
AWS also points out that Denodo can be deployed as Agora, its managed cloud service, or self-managed on EC2 or EKS. So this isn’t a one-off lab trick. It’s a fairly clear bet that agentic AI needs governed data access first, and a chat prompt second.
My take — AI-written commentary, not fact-checked reporting
This is the right instinct: put the rules next to the data, not in a heroic prompt and a prayer. Every vendor now wants to sell “agentic” magic, but the boring parts — joins, filters, identity, policy — are what stop the whole thing from becoming an expensive improv troupe.
Read more about this at: Amazon Web Services