Blog Post

Fabric Updates Blog
4 MIN READ

Prepare your Fabric Data Agent integrations for Assistants API retirement

mayurjain's avatar
mayurjain
Icon for Microsoft Employee rankMicrosoft Employee
2 months ago

The OpenAI Assistants API that powers the orchestration layer for the Microsoft Fabric data agent is currently scheduled to be shut down by OpenAI on August 26, 2026. After that date, direct calls to the Assistants API will stop working. If you connect to a Fabric data agent programmatically through the Assistants API, you need to migrate to the data agent Model Context Protocol (MCP) endpoint before then. This post explains what is changing, who is affected, and the steps to move over smoothly.


What’s changing

Fabric data agent answers questions about your data in natural language. Behind the scenes, it has used the OpenAI Assistants API as its orchestration layer — the part that keeps track of a conversation, decides which tools to run (such as generating a query or running code), and assembles the final answer. When you ask a data agent a question through the Fabric portal, the SDK, or directly through the API, the request flows through that orchestration layer.


The public programmatic interface exposed this model directly. Developers could create assistants, manage conversation threads, and submit runs against their data agents in their own code. That direct dependency on the Assistants API is what the OpenAI deprecation affects.


OpenAI is moving from the Assistants API to the Responses API. Because the Responses API is not a drop-in replacement for the Assistants API, Fabric recommends migrating integrations to the data agent MCP endpoint rather than rebuilding directly against OpenAI APIs.


The MCP endpoint exposes your data agent as a Model Context Protocol server and abstracts the underlying orchestration API. This provides a standards-based interface that can continue working even as the implementation evolves.


Who is affected

The impact depends entirely on how you connect to the data agent today.

  • If you programmatically use Data Agents by invoking the Assistants API or the Public REST API within or outside Fabric to create assistants, manage threads, and submit runs — you must migrate to the MCP endpoint.

Publishing tab in Data Agent settings exposing the Public REST API

  • If you use the data agent through the SDK, it will be migrated to use the Responses API internally. SDK document with revised code samples demonstrate the new implementation.
  • If you use the data agent through the Fabric portal, you do not need to take any action. Fabric is moving this experience to the Responses API internally and will complete that work before August 26, 2026.


For the experiences that Fabric will migrate internally, you should see no change in performance or answer quality compared to the Assistants API. The internal migration is designed to preserve the behavior you rely on today, but you can expect a one-time reset of conversation history.


Action required: Migrate to MCP before August 26, 2026

If your code calls the Assistants API directly, plan to migrate to the data agent as a Model Context Protocol server before August 26, 2026. The MCP endpoint is the supported programmatic interface going forward.


To get started, after publishing the data agent, go to the Settings of the agent and open the Model Context Protocol tab. The tab provides the following information:

  • Data agent MCP server name
  • MCP server URL (copy this URL to use in your application)
  • Data agent MCP tool name
  • MCP server tool description

Model Context Protocol tab in Data Agent settings

Refer to the example for setting up MCP server in VS Code: Data agent as Model Context Protocol server. Validate your integration end to end after you switch, so you can confirm that the performance matches your expectations before you retire the old path.


The MCP endpoint provides a few key benefits:

  • MCP gives you an open, model-agnostic interface instead of tying your integration to provider-specific assistant, thread, and run concepts.
  • By abstracting the orchestration layer, MCP helps your integration keep working as underlying APIs evolve.
  • Since Responses API is not a drop-in replacement, MCP helps you avoid rebuilding conversation-state management and lets you focus on your application.


What stays the same

Your data agent configuration does not change. The data sources you have connected, the instructions you have written, and the tools you have enabled all remain intact. This is a change to how your code connects to the data agent, not to what the data agent does.


It is also worth noting that conversation persistence behaves the same way it always has for data agents. For the authoritative details on conversation history behavior and other current considerations, refer to the data agent documentation — documentation is the maintained source of truth and is the right place to check current behavior over time.


Next steps

We recommend beginning your migration as soon as possible so you have time to validate and test before the deprecation deadline. Migrating well ahead of the deadline gives you time to validate your integration and avoid any last-minute disruption. Treat this migration like any other dependency update: schedule it, test it in a non-production environment first, and roll it out to production once you are confident in the results.


To get started, refer to Connect to a Fabric data agent with the MCP endpoint.

Updated 2 months ago
Version 1.0

1 Comment

  • I’m not sure whether this is related to what you're describing in your post, but since yesterday I've been experiencing major issues with a Fabric Agent that I have published to Microsoft 365.

    The agent works correctly inside Fabric, but it is no longer working properly in Microsoft 365. It struggles to answer even the simplest follow-up questions. Up until Friday, it was working fine, and I haven't changed anything in the configuration, data sources, or agent settings.

    For example, if I ask:

    "Give me the payments for July."

    and then follow up with:

    "Also give me the payments for June."

    the agent responds with something like:

    "What would you like me to provide?"

    It used to understand this type of conversational context perfectly, but since yesterday it appears to have lost that capability. Has anyone else experienced similar behavior?