Fabric Data Agents turn questions into queries

Microsoft Fabric Data Agents provide conversational question-and-answer over supported Fabric data sources. The standard runtime is generally available. Some newer sources and authoring tools remain in preview, so production readiness depends on the exact configuration.

Separate the runtime from preview features

Microsoft documents a generally available standard runtime and a preview runtime. The standard runtime includes query generation and orchestration for supported SQL, KQL, and Power BI sources. Preview features can be added separately and keep their own terms and limitations.

The runtime documentation is the right place to check that boundary. Do not treat the word “GA” as a blanket status for every source, model, or authoring experience.

Know what the agent can query

The current supported-source table lists lakehouses, warehouses, SQL databases, mirrored databases, Eventhouse KQL databases, and Power BI semantic models for the established query paths. Graph, ontology, and Azure AI Search paths are marked as preview.

The agent generates a query that the connected source executes. It does not repair weak definitions. Clear table names, relationships, measures, source instructions, and example queries reduce ambiguity.

Access controls still need review

Microsoft states that Data Agents respect supported Purview controls on underlying sources, but policy coverage varies by source and feature status. The Data Agent concepts page lists prerequisites, governance behaviour, and limitations.

Before release, test with identities that represent actual users. Confirm both allowed and denied questions. Do not rely on an administrator’s successful test as evidence for a viewer’s access boundary.

Build an evaluation set before launch

Use a small, named domain and record:

1. questions the agent should answer
2. expected sources, measures, and filters
3. questions it should refuse or escalate
4. generated-query accuracy and cost
5. the owner who corrects definitions or examples

Retest after model, source, schema, measure, instruction, or runtime changes. A conversational interface makes data easier to ask about; it does not remove the need to validate the answer.