Fabric data agent

Fabric Data Agent

Use this skill for Fabric Data Agent design and lifecycle tasks. Data Agents are read-only conversational analytics artifacts; they are not general-purpose automation agents.

Required References

• Data Agent reference: FABRIC-DATA-AGENT-CORE.md
• ALM reference: FABRIC-ALM-CORE.md
• Governance patterns: COMMON-CORE.md

Workflow

1. Clarify the audience, business domain, and questions the Data Agent should answer.
2. Select data sources and tables deliberately. Prefer certified semantic models for governed metrics, Warehouse/Lakehouse tables for relational exploration, and KQL databases for telemetry.
3. Confirm permissions, capacity, tenant AI settings, cross-region constraints, and Purview/DLP implications.
4. Draft the Data Agent instructions: glossary, source routing, metrics, fiscal calendar, ambiguity handling, and refusal rules.
5. Add validated SQL/KQL example query pairs where supported.
6. Build an evaluation set and test generated queries/intermediate steps before publishing.
7. Plan sharing, monitoring, periodic review, and ALM promotion through Git/deployment pipelines where supported.

Delegation

Need	Delegate
Querying or validating Lakehouse/Warehouse data with T-SQL	sqldw-consumption-cli
Querying or validating KQL database behavior	eventhouse-consumption-cli
Semantic model metadata or DAX validation	powerbi-consumption-cli
Source table design for Data Agent readiness	fabric-lakehouse, sqldw-authoring-cli
ALM and deployment of Data Agent configuration	fabric-alm-cicd
Governance/security review	FabricAdmin

Must

• Keep Data Agent behavior read-only.
• Respect Purview, RLS/CLS, workspace roles, model Read permissions, and selected data-source scope.
• Use clear English instructions and examples.
• Validate generated SQL/KQL/DAX against the actual source schema when possible.
• Surface limitations early, especially row/column caps, unsupported unstructured files, semantic-model example-query limitations, and cross-region constraints.

Prefer

• Narrow, trusted source selection over broad workspace exposure.
• Business glossary and metric definitions embedded in Data Agent instructions.
• Evaluation questions covering happy paths, ambiguous wording, access-denied scenarios, and sensitive data boundaries.
• Published Data Agent descriptions that explain purpose, scope, owner, data freshness, and escalation path.

Avoid

• Using a Data Agent for write operations, remediation automation, or full dataset export.
• Adding raw files directly as sources; expose them through tables first.
• Routing certified KPI questions to raw Lakehouse tables when a semantic model owns the business logic.
• Ignoring Purview audit/eDiscovery implications for sensitive workloads.

Output Format

For design requests, return:

• Use-case and audience summary
• Source selection and routing rules
• Governance and permission model
• Data Agent instructions draft
• Example query plan
• Evaluation set
• Publish/share/ALM plan