Skip to content
CSA Loom — the Microsoft Fabric experience for Azure tenants where Fabric isn't yet available: lakehouses, warehouses, notebooks, semantic models, Activator rules, Data Agents, across Commercial, GCC, GCC-High, and DoD IL5

Data Agents parity

What Fabric does

Data Agents (renamed from "AI Skills" in 2026) are NL-to-query agents over Fabric items. GA at FabCon March 2026. Built on Azure OpenAI Assistant APIs. Multi-tool NL→query over up to 5 sources per agent:

  • NL2SQL for Lakehouse / Warehouse / SQL DB / Mirrored DB
  • NL2DAX for Power BI semantic models
  • NL2KQL for KQL DBs / Eventhouse (incl. UDFs)
  • Microsoft Graph queries
  • Custom AI Search index for unstructured grounding (Ignite 2025)

Read-only by design. Identity-passthrough: agent invocations execute under the calling user's Entra identity, so RLS / CLS / object-level security applies naturally.

Per-agent configuration: instructions, example queries (Q→SQL/DAX/KQL pairs — heavily weighted by the LLM for shape-matching), verified answers, field descriptions, sensitivity policy. Surfaceable as tools in Foundry Agent Service agents.

CSA Loom parity design — extends apps/copilot/ + azure-functions/copilot-chat/

Per temp/fiab-research/07-existing-repo-scope.md, the csa-inabox Copilot scaffold is ~70% there. Reuse: - PydanticAI agent (Loom Data Agents runtime) - Azure Functions backend (security: rate limit, PII redaction, content safety, telemetry — all keep) - AI Search vector index becomes per-agent grounding store

New tools added

class LoomDataAgent:
    @tool
    async def nl2sql(self, question: str, data_source_id: str, user_token: str) -> QueryResult:
        """Generate and execute SQL against a registered lakehouse / warehouse."""
        # Load schema from UC (Commercial) or Purview (Gov)
        # Load few-shot example Q/SQL pairs from Cosmos DB
        # Call AOAI with system prompt + grounding
        # Execute under user_token (OBO) against Databricks SQL Warehouse
        # or Synapse Serverless
        # Return rows + generated SQL as citation

    @tool
    async def nl2dax(self, question: str, semantic_model_id: str, user_token: str) -> QueryResult:
        """Generate and execute DAX against a Power BI semantic model via XMLA."""

    @tool
    async def nl2kql(self, question: str, adx_cluster: str, database: str, user_token: str) -> QueryResult:
        """Generate and execute KQL against an ADX database."""

    @tool
    async def graph_search(self, query: str, user_token: str) -> List[Entity]:
        """Microsoft Graph search (people, files, mails) — OBO."""

    @tool
    async def custom_search(self, query: str, search_index: str, user_token: str) -> List[Doc]:
        """Search a custom Azure AI Search index (PDFs, docs, etc.)."""

Per-agent configuration (Cosmos DB schema)

Per PRP-09:

{
  "id": "agent-finance-analyst",
  "workspaceId": "ws-001",
  "name": "Finance Analyst Agent",
  "description": "Answers questions about sales, expenses, and AR/AP",
  "instructions": "You are an expert at answering finance questions...",
  "dataSources": [
    {"type": "lakehouse", "id": "finance-lakehouse", "instructions": "..."},
    {"type": "semantic-model", "id": "finance-semantic-model", "instructions": "..."},
    {"type": "adx", "cluster": "adx-loom-eastus2", "database": "finance-events"}
  ],
  "exampleQueries": [
    {"question": "What were sales last quarter by region?", "language": "DAX", "query": "..."}
  ],
  "verifiedAnswers": [],
  "fieldDescriptions": {},
  "sensitivityPolicy": "Block agent on tables tagged 'PII-restricted'"
}

Identity passthrough (OBO)

Per ADR fiab-0009 and AMENDMENTS A15: every tool call carries the caller's Entra token. RLS / CLS / object-level security apply naturally because the underlying engine (Databricks SQL, Synapse Serverless, ADX, Power BI XMLA) honors caller identity.

Foundry integration (Commercial only)

Published Loom Data Agents are surfaceable as tools in Foundry Agent Service agents via the standard MCP-compatible endpoint registration. One Foundry agent attaches one Loom Data Agent (mirroring Fabric's current limitation). In Gov where Foundry isn't GA, agents are callable via REST.

Per-boundary behavior

Boundary NL2SQL NL2DAX NL2KQL Foundry integration
Commercial ✅ Databricks SQL Warehouse ✅ Power BI XMLA ✅ ADX ✅ Foundry Agent Service
GCC ✅ Databricks SQL Warehouse ✅ Power BI XMLA ✅ ADX ✅ Foundry Agent Service
GCC-High / IL4 ✅ Synapse Serverless ✅ Power BI XMLA ✅ ADX ❌ MAF + AOAI direct only
IL5 (v1.1) ✅ Synapse Serverless ✅ Power BI XMLA ✅ ADX ❌ MAF + AOAI direct only

Honest gaps

  • NL2DAX maturity — production NL2DAX is materially less mature than NL2SQL. Loom Data Agents ships NL2DAX but flags accuracy expectations in the agent's system prompt; v1.1 hardens with more training examples
  • "Prep for AI" semantic-model annotations — Loom's equivalent (Cosmos DB store edited via Console) is functional but Fabric's UX is more polished; v1.1 invests
  • Foundry integration in Gov — unavailable until Foundry Agent Service Gov-GAs

Forward migration

Agent definitions export to JSON + Foundry Agent JSON formats; import into Fabric Data Agents via the public REST API (preview at time of writing). Example queries port 1:1.