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

Tutorial: Agent flow editor

CSA Loom agent-flow editor — a standalone visual multi-agent workflow. Chain grounded data tools (lakehouse / warehouse / KQL / AI Search), capability tools (MCP servers, OpenAPI, functions), and connected sub-agents on a canvas, then run the flow against your real Azure backends and publish it as an MCP tool. Azure-native — no Microsoft Fabric or Foundry dependency on the default path.

What it is

A single chat agent answers from whatever you paste into its prompt. An agent flow is a design: an orchestrator with explicit instructions, a set of tools that are bound to your real Loom items, and sub-agents it can delegate to. At run time, Azure OpenAI grounds over the bound items and delegates to each sub-agent, and the whole run is recorded.

Three tabs: Design, Runs, Publish.

When to use it

  • One question needs data from more than one place (an order table and live telemetry) with a citation trail.
  • You want specialized sub-agents (pricing, compliance, logistics) that an orchestrator can call.
  • You want the finished flow callable by other systems as an MCP tool.

Step-by-step in Loom

  1. Create the item. + New item → Agent flow, then Create agent flow.
  2. Write the orchestrator instructions. On Design, the first card is the orchestrator prompt — what the flow should do and how it should combine its tools and sub-agents. This grounds the Azure OpenAI orchestrator at run time. For example:

    You are a supply-chain analyst. Use the warehouse tool for order data and the KQL tool for live telemetry, delegate pricing questions to the pricing sub-agent, and always cite the source.

  3. Compose the flow on the canvas. The AgentFlowCanvas below lets you add:
  4. Grounded data tools — bound to real Loom items (lakehouse, warehouse, KQL database, AI Search index).
  5. Capability tools — MCP servers, OpenAPI operations, functions.
  6. Connected sub-agents — other agents the orchestrator can delegate to.
  7. Guardrails — the flow's guardrail settings, shown inline.

Node positions persist as part of the item's layout. 4. Run it. The canvas has an embedded run pane. Ask a question there and the flow executes for real through the item's own owner-scoped run route (POST /api/items/agent-flow/<id>/run) — real grounded orchestration over the bound items and sub-agents, not a simulation. 5. Save. The ribbon's Save and the bottom save bar both persist the instructions, tools, sub-agents, guardrails, and layout; the save bar shows the last saved time. 6. Review the history. Runs lists every execution: started, the question, how many grounded sources and capability tools were used, how many sub-agents (with a check mark when delegation actually happened), total tokens, and status. Runs are persisted with the item and survive reloads — newest first, up to 50 retained. 7. Publish as an MCP tool. Publish exposes the saved flow as an MCP tool other systems can call, showing the tool name and the publish time. Publishing requires a flow with content (instructions, tools, or sub-agents) and no unsaved changes — if the item is dirty, an info MessageBar tells you to save first so the MCP server serves the latest design.

The Azure backend it rides on

  • Orchestration: Azure OpenAI connected-agents runtime, grounded over your bound Loom items.
  • Grounded data tools: the real backends behind those items — ADLS Gen2 / Delta (lakehouse), Synapse SQL (warehouse), Azure Data Explorer (KQL), Azure AI Search.
  • Capability tools: your deployed MCP servers, OpenAPI endpoints, and functions.
  • Routes: POST /api/items/agent-flow/<id>/run, GET /api/items/agent-flow/<id>/runs — owner-scoped, with the run history persisted to the item's state.

Honest gates

Condition What you see Exact remediation
No runs yet "No runs yet. Switch to Design and ask a question in the run pane." Run the flow once
Unsaved changes on Publish Info MessageBar "Unsaved changes"; publish is blocked Save the flow, then publish
Empty flow Publish is unavailable until the flow has instructions, tools, or sub-agents Author the flow first
A bound tool's backend is unconfigured That tool's own honest gate surfaces at run time with its remediation Configure the named backend for that item type

No Fabric required

Azure OpenAI plus the Azure backends behind your bound items. No Fabric capacity, workspace, OneLake path, or Power BI workspace is used on the default path.

Learn more