Fabric Data Agent Upgrade and Foundry Migration Checklist (November 2026)

This November, Microsoft Fabric Data Agent work sits in two separate windows. They are not one toggle. One window upgrades how the standard runtime writes SQL and DAX. The other retires an older way to wire a Fabric data agent into Microsoft Foundry. We haven't tested whether each tenant has already switched; the project's Fabric runtime selector and Foundry Tools list are the live signal. For reference only.

Audience stays narrow: teams that build or operate Fabric data agents, and teams that attach those agents inside Foundry. Ordinary Copilot chat users who never touch a data agent or a Foundry tool catalog can skip this page. Related Foundry agent-runtime context: Microsoft Copilot Harness Lands in Foundry Agent Service Preview.

Why two checklists

The Fabric Community upgrade post covers Advanced NL2SQL, advanced DAX generation, and the move of those core improvements—plus GPT-5.6 Terra—onto the standard runtime on a rolling calendar of about November 2 through November 10. The Fabric Community retirement post covers a different product surface: the earlier custom tool connection from Foundry to a Fabric data agent stops being supported starting November 6, 2026, with Fabric IQ (OneLake Catalog) as the replacement path (still labeled Preview).

Mixing the two into one “Fabric just changed” story creates the wrong action list. A tenant can pick up the runtime upgrade and still need a Foundry connection migration—or the reverse. Run the checklists below as independent passes.

---

Checklist A — Standard runtime upgrade (about November 2–10)

What is rolling out

Per the Fabric Community upgrade post titled *Advanced NL2SQL and advanced DAX generation in Fabric data agents (Generally Available)*:

• Core Advanced NL2SQL and advanced DAX generation improvements begin rolling to the standard runtime on November 2, with the rollout expected to finish by November 10.
• As part of that rollout, the standard runtime also moves to GPT-5.6 Terra for orchestration and supported data agent capabilities.
• Data agents on the standard runtime receive these improvements by default. No opt-in is required when the rollout reaches that environment.
• Schema descriptions and topics for SQL sources remain on the preview runtime. They are not part of what moves to standard in this wave.

Advanced NL2SQL, in the same post, is framed as explore-then-write behavior: schema exploration, join tests, and real filter-value lookup before SQL is emitted, with modes such as example-query following, clarifying questions, and data exploration. Advanced DAX generation is framed as an agentic tool-calling loop over semantic models, including instance value indexing so filters match stored values more consistently across Data Agent, Power BI, and Microsoft 365 Copilot experiences described there.

What does not change in this wave

• Preview-only configurations for SQL—schema descriptions and topics—stay preview.
• The upgrade post does not publish a named timezone or a single UTC hour for each tenant flip. Reading November 2–November 10 as a rolling calendar window is safer than inventing a midnight cutover.
• We haven't checked each tenant's switch timing. When the standard runtime in a given environment already shows the new behavior, that environment has the upgrade; when it does not, retesting continues as the window progresses.

Internal benchmarks are not a business guarantee

The upgrade post shares Microsoft internal evaluation numbers (for example, higher smart-answer accuracy and lower P50/P95 latency on internal sets for Advanced NL2SQL and advanced DAX generation versus prior baselines). Those figures are useful context for why Microsoft is promoting the change. They are not a promise that every production warehouse, lakehouse, or semantic model will jump by the same margins. Workload, data model, instructions, examples, and configuration still dominate outcomes. Internal benchmarks are Microsoft's lab signal—not a substitute for the team's own question set.

Checklist A — actions

1. Confirm the surface. The agent under test is a Fabric data agent on the standard or preview runtime, not a free-form Copilot chat that never queries through a data agent.
2. Runtime selector. As the November 2–10 window progresses, the data agent experience shows whether the standard runtime has the new stack. When it does, re-run the team's existing question set there. The same class of changes is available earlier on the preview runtime, as the upgrade post describes.
3. Re-run evaluations with the Fabric data agent SDK and a ground-truth set when that workflow already exists. Compare against the previous baseline. Visible reasoning / run steps, where the product exposes them, show where assumptions shifted after explore-then-write behavior.
4. Retune examples and instructions. Guidance written only to work around older one-shot generation may be unnecessary—or actively misleading—once the agent explores schema and substitutes filter values more aggressively.
5. Preview configs stay in the preview bucket. If the team relies on schema descriptions or topics for SQL, those controls stay on preview in this wave; they did not silently become standard-runtime features on November 2.
6. Record tenant reality, not calendar folklore. We haven't tested each tenant. What the project's own runtime UI and evaluation runs show between November 2 and November 10 is what counts.

Short form for Checklist A: standard runtime gets Advanced NL2SQL + advanced DAX + GPT-5.6 Terra on a November 2–10 roll, no opt-in; schema descriptions / topics for SQL stay preview; retest with the old question set; internal scores ≠ business accuracy.

---

Checklist B — Foundry custom tool → Fabric IQ (deadline November 6, 2026)

What is changing

Per the Fabric Community post *Retiring the Fabric data agent custom tool in Microsoft Foundry: move to Fabric IQ*:

• The earlier integration let Foundry agents consume a Fabric data agent by adding it as a custom tool.
• That custom-connection integration is being deprecated. Starting November 6, 2026, it is no longer supported.
• Foundry agents that still use a Fabric data agent as a custom tool need to migrate before that date.
• The replacement path (labeled Preview) uses the Fabric IQ (OneLake Catalog) tool so published Fabric data agents can be selected from the catalog inside the Foundry portal.
• “No longer supported” is the official support cutoff wording. It is not proof that every leftover custom connection drops for every tenant at the same instant. Plan for the published support end date; verify connection health in the Foundry project after migration.
• The replacement is Preview, not a claim that Fabric IQ for this path is generally available.

Why Microsoft is pushing the new path (as stated)

The retirement post calls out catalog discovery by name (without pasting workspace ID / artifact ID), the ability to attach more than one Fabric data agent to a Foundry agent (the older custom path supported only one), and presentation of each selected data agent through a Model Context Protocol (MCP) endpoint. Coming-soon wording in that post also mentions returning the query used to produce an answer—that wording is forward-looking, not a production feature already in verified use.

Checklist B — portal migration (official sequence)

Complete these steps before November 6, 2026 if any Foundry agent still uses a Fabric data agent as a custom tool:

1. In Microsoft Fabric, publish the Fabric data agent with a rich, detailed description (helps catalog discovery and instruction writing).
2. Open the agent in the Microsoft Foundry portal.
3. In Tools, add Fabric IQ (OneLake Catalog), then select the published data agent from the catalog (filter by data agent type as shown in the post).
4. Update the Foundry agent's instructions: name each connected data agent and describe which question types it should handle.
5. Remove the earlier custom tool, then save.
6. Test with representative questions for each connected data agent.

For builders who wire connections in code, the same post points to updating the implementation to the Fabric IQ tool, including MCP server URL details under the data agent's Model Context Protocol settings, and sample flows in Python / C# / JavaScript on Microsoft's “Connect agents to Microsoft Fabric with Fabric IQ” guidance. The official sample path is the one to follow; a third connection style is out of scope.

Checklist B — verification notes

• After cutover, confirm the Foundry agent lists Fabric IQ tools and no leftover custom-tool entry for the Fabric data agent.
• Confirm multi-agent setups explicitly: the new path allows more than one data agent; instructions must disambiguate which agent handles which questions.
• We haven't tested each tenant's migration state. If a project still shows only the old custom tool after November 6, that is an urgent support / migration issue against the published cutoff—not proof the cutoff was soft.

Short form for Checklist B: by November 6, 2026, replace Foundry custom tool links to Fabric data agents with Fabric IQ (OneLake Catalog) (Preview); publish + catalog select + rewrite instructions + remove old tool + retest; support end ≠ proof every leftover connection drops instantly.

---

Scope and reading order

Checklist A applies to every Fabric data agent the team ships answers from. Checklist B applies only where Foundry agents still consume Fabric data agents through the old custom tool. Many teams need both; some need only one.

For broader Foundry agent-hosting context (Copilot harness preview on Foundry Agent Service—a different surface from Fabric IQ catalog wiring), see Microsoft Copilot Harness Lands in Foundry Agent Service Preview. Harness product shape and billing trade-offs in the abstract are covered in Same Model, Different Coding Harness: Pass Rate Barely Moves, the Bill Does.

Closing

November 2026 is an execution week for Fabric Data Agent operators: retest standard-runtime query quality as Advanced NL2SQL / advanced DAX / GPT-5.6 Terra arrive without opt-in; SQL schema descriptions / topics stay in the preview mental model; Foundry migrations off the custom tool onto Fabric IQ finish before the November 6 support cutoff. We haven't tested every tenant. Official Fabric Community upgrade and retirement posts remain the wording to trust when console UI and this page disagree. For reference only.