Fabric can now draft a data agent—but its configuration still needs a review gate

Microsoft has put another agent inside the data-agent development loop. The new Build Agent with AI preview generates and refines Microsoft Fabric Data Agent configurations through an interactive workflow, using schema and conversation context to recommend agent instructions, data-source guidance and example queries. Microsoft says the preview currently focuses on SQL and Eventhouse scenarios. (Microsoft Fabric What's New)

That is useful because these fields are not cosmetic prompts. Fabric’s own architecture says instructions influence which source an agent selects, while example question-query pairs teach it how to handle recurring requests. Microsoft’s documentation gives routing financial metrics to a Power BI semantic model, raw exploration to a lakehouse and log analysis to a KQL database as examples of developer-supplied guidance. (Fabric data agent creation)

Review the generated contract

The preview therefore automates part of the agent’s operating contract, not just a setup form. A sensible release process should review generated configuration in three passes.

First, inspect scope. Fabric agents can combine as many as five sources, including warehouses, lakehouses, semantic models, KQL databases, ontologies and Microsoft Graph. Microsoft recommends choosing only relevant tables. An AI-generated recommendation that widens the source set may change query routing even when it does not change permissions. (Fabric data agent creation)

Second, test definitions and examples. Example pairs can anchor common SQL or KQL questions, but Microsoft says they are not currently supported for Power BI semantic-model sources. Teams should not assume the builder creates equivalent calibration coverage across every connector. (Fabric data agent creation)

Third, verify policy precedence. Fabric places organizational policy above role-based settings, developer instructions and end-user prompts, in that order. It also says the agent uses the requesting user’s credentials, remains read-only and applies supported Purview controls to the underlying data. Generated instructions cannot legitimately override those boundaries—but they can still route an allowed question toward the wrong allowed source or encode an incorrect business definition. (Fabric data agent creation)

Treat configuration like code

The practical gain is faster authoring: the preview can propose the repetitive guidance that previously required manual setup. The operational risk is quieter. A plausible instruction can look finished before anyone tests whether it selects the correct source, respects local terminology or produces the expected query.

For production teams, the right unit of approval is the complete configuration change: selected sources and tables, generated instructions, example pairs, and a regression set of representative questions. Keep the builder’s output as a candidate, compare it with the previous configuration, and publish only after answer and routing behavior pass review.

Build Agent with AI lowers the cost of creating the contract. It does not remove the need to own that contract.