Enterprise AI has spent years treating the data catalog like a library card: useful for finding information, but separate from the system doing the work. Amazon Quick's new Agentic Catalog Experience collapses that distance. A curator can describe an analytics use case, let an agent recommend catalog assets, create datasets in bulk, and assemble a multi-dataset topic with inherited semantics.
The important shift is not conversational search. It is that metadata can now become executable context.
AWS launched the feature in preview for AWS Glue Data Catalog and Databricks Unity Catalog. Its documentation describes a four-stage flow: discover relevant tables from natural language; create DirectQuery datasets; inherit or infer relationships; then carry table and column descriptions into a reusable topic. The upstream catalog remains the source of truth.
That design can compress setup work. It also turns weak metadata into a production risk.
The catalog-to-agent chain
The feature exposes a new enterprise AI supply chain:
1. Discovery: Quick uses descriptions, data-quality scores, and lineage to recommend tables.
2. Materialization: Approved assets become DirectQuery datasets without replacing the upstream catalog.
3. Composition: Existing relationships are inherited, additional ones may be inferred, and selected datasets become a multi-dataset topic.
4. Grounding: Catalog descriptions become business context for downstream questions.
Each step removes manual translation. Each step can also transmit an error farther.
A stale column description is no longer merely bad documentation. An incorrect relationship is no longer confined to a modeling notebook. If those semantics are inherited by an agent-facing topic, they can shape which data is selected and how a business question is interpreted.
AWS acknowledges the boundary directly: authors should review discovered tables, inferred relationships, and inherited descriptions before proceeding. The product automates assembly, not accountability.
A semantic control plane
Independent reporting from VentureBeat shows that this is bigger than one AWS feature. Google, Databricks, Snowflake, and Microsoft are also positioning semantic catalogs and federation as infrastructure for agents. Vendors differ on who creates and maintains the semantics, but the market is converging on the same premise: data access without business meaning is not enough.
Operators should treat the catalog as a semantic control plane with four controls:
- Source truth: Keep ownership, lineage, permissions, and refresh behavior anchored upstream.
- Context boundary: Approve only the assets needed for a specific use case. More tables create more ambiguity, not automatically better answers.
- Review gate: Require a named owner to approve recommendations and every inferred relationship.
- Regression receipt: Save the selected assets, descriptions, relationships, and test questions as a versioned deployment record.
This reframes readiness. A connector being online is not success. Success means the agent receives the right slice of data, with semantics that survive representative questions and permission checks.
The operator playbook
Before enabling an agentic catalog workflow, teams should run four tests.
First, test descriptions as instructions. Ask whether a new analyst could distinguish similarly named measures, time windows, and entities from the metadata alone. If not, an agent probably cannot do so reliably either.
Second, test relationship direction and grain. An inferred join can look plausible while duplicating measures or mixing incompatible levels of detail. Use known-answer queries that expose fanout and aggregation errors.
Third, test the context boundary. Build one narrow topic for one decision workflow. Compare it with a broader version and measure whether extra assets improve or degrade retrieval and interpretation.
Fourth, test change. When a schema, description, or relationship changes, rerun the same questions before promoting the new semantic version.
For founders, this creates a practical tooling opportunity. Enterprise teams will need semantic linters, catalog-quality scores, relationship test suites, approval workflows, and drift monitoring that work across vendors. The winning products may look less like another chatbot and more like CI/CD for business meaning.
Amazon Quick's preview makes the direction legible: the enterprise AI bottleneck is moving upstream. Models still matter, but the operating advantage will increasingly come from making context bounded, reviewable, and testable before an agent acts on it.
Sources
- AWS: https://aws.amazon.com/blogs/machine-learning/announcing-the-agentic-catalog-experience-in-amazon-quick/
- Amazon Quick documentation: https://docs.aws.amazon.com/quick/latest/userguide/catalog-integration-agentic-flow.html
- VentureBeat: https://venturebeat.com/data/the-modern-data-stack-was-built-for-humans-asking-questions-google-just-rebuilt-its-for-agents-taking-action/