Microsoft Foundry Hosted Agent – Fabric Data Agent Tool fails with 400 tool_user_error despite successful Foundry authentication

Hi @Allan Solomon Mejia

Thank you for the clarification.

We have already provided all the required permissions to the end user, including access to the Fabric workspace, Fabric Data Agent, and underlying semantic model/data.

Also, H_Analysis_Agent is a prompt agent, and when we invoke it directly from the Foundry Playground, the Fabric Data Agent works successfully.

The same nested-agent architecture also works when we test it locally using az login credentials.

The issue occurs specifically when the same architecture is deployed/hosted and invoked with the delegated end-user identity.

Hello @Manudeep Yadav Vanneboina

Your Router Agent working confirms that DefaultAzureCredential() and the service principal can authenticate the hosted agent to the Foundry project, but it doesn't establish an identity that the Fabric Data Agent tool can use.

The key limitation here is that Microsoft currently documents the Fabric Data Agent integration as using end-user identity passthrough (On-Behalf-Of/OBO). It specifically states that user identity authentication is required and service principal authentication isn't supported for the Fabric Data Agent tool.

So in your architecture:

Granting the service principal additional permissions to the Fabric workspace or semantic model therefore won't necessarily resolve the 400 tool_user_error. The issue isn't simply whether the SP has Fabric permissions; it's that the Fabric tool expects a supported user identity context when it executes the query.

Microsoft also states that the Fabric tool runs queries using the signed-in user's identity and that each end user needs access to both the Fabric Data Agent and its underlying data sources.

This makes the error you're seeing during:

Create assistant failed

code: tool_user_error

consistent with the Fabric tool being instantiated without the user identity context it requires.

I would therefore test H_AnalysisAgent separately using an interactive/delegated user identity that has access to:

1. The Foundry project.
2. The published Fabric Data Agent.
3. The underlying Fabric semantic model/data source.

If that succeeds, you've isolated the problem to identity propagation from the MAF Hosted Agent, rather than the Fabric connection itself.

Also verify that the Fabric Data Agent is published and that the Fabric Data Agent and Foundry project are in the same tenant. Microsoft lists both as prerequisites.

For the hosted-agent architecture specifically, I would not try to solve this by placing AZURE_CLIENT_ID, AZURE_CLIENT_SECRET, and AZURE_TENANT_ID in the container and granting the application progressively broader Fabric permissions. That still provides an application identity, while the Fabric integration currently expects the user's delegated identity.

If you need fully unattended service-to-service execution with no signed-in user, the documented Fabric Data Agent tool integration doesn't currently support that authentication model. You would need either a different integration pattern that supports application identity, or confirmation from Microsoft that Hosted Agents provide a supported mechanism for propagating delegated user identity through this particular MAF orchestration path.

Given that this is a Hosted Agent → Prompt Agent → Fabric Data Agent scenario and the Fabric integration is still in preview, if you are already supplying a delegated end-user token and still receive this error, I would recommend escalating it with the request_id: 33935f862d81c6d9b4bc16220e8ecb21

That would require Microsoft to determine whether OBO identity propagation through this particular nested-agent architecture is currently supported.

Sharing these references with you:

Use the Microsoft Fabric data agent with Foundry agents

Consume a data agent in Microsoft Foundry

Help make this community better for everyone: if this answer resolved your issue, please accept it or leave an upvote. If not, share more details in a comment so we can continue the discussion and find the right solution.