Forum Discussion
Fabric Data Agent vs MicrosoftFoundry Agent+MCP—Is dedicated Fabric capacity worth it for AI agents?
Hi Corujo,
Thank you for reaching out to the Microsoft fabric community forum.
Based on Microsoft's documentation, Azure AI Foundry supports Microsoft Entra-based agent identities, allowing agents to securely access downstream resources using Entra authentication and RBAC.For your testing, it may be useful to validate how identities are propagated across the full request path and whether authorization is consistently enforced for both internal and B2B guest users. On the Fabric side, One Lake Security provides centralized access controls that can be assigned to users, groups, and service principals to help govern data access.
Consider following points:- Identity propagation across each layer of the architecture.
- RBAC enforcement for agent identities and end users.
- Authentication and authorization behavior for internal and guest users.
- Consistency of OneLake Security permissions when data is accessed through the agent workflow.
These tests should help determine whether the Foundry + MCP architecture aligns with your security and governance requirements alongside your ongoing performance and cost evaluations.
Please go through the Microsoft Documentations:
https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/agent-identityhttps://learn.microsoft.com/en-us/fabric/onelake/security/get-started-security
Thank you.
Super UserHey Corujo, we ran basically this same eval. Here's some quick thoughts:
Your 66% is real but partly a smoothing artifact. Data Agent is a background job, so 62 queries crammed into one eval window spikes way harder than the same 62 spread across a normal day. The bigger thing people miss: the LLM tokens are only half the bill. The SQL the agent generates against your Lakehouse gets billed separately through the query engine. So it's not 62 calls, it's 62 calls plus 62 queries hitting the same pool your pipelines and reports run on. And there's no per-workload reservation, it all shares one capacity, and the agent gets rejected first under pressure.
On cost, your read is right.
At around 100 queries/day, Foundry + SQL MCP on pay-per-token wins easily, and you get real per-agent cost tracing on top. Plus the moment you want write ops or multi-tool agents, Data Agent isn't even in the running.
It's read-only Q&A.
Where it does make sense, honestly, comes down to one question: will you need F64+ for other reasons? If yes, the agent overhead vanishes into headroom you've already paid for, and the semantic-model integration is genuinely nice. On F8 you're basically paying capacity just to run the agent, which is the worst spot to be in.
Our call at your scale: not worth it. We went Foundry + MCP for the agent and kept Fabric for what it's good at.
One catch before you commit, though. If your semantic models carry a lot of business logic, budget time to rebuild that in the MCP layer or your answer quality drops. That's the cost nobody puts in the spreadsheet.
If this works for you, kindly mark it as the solution and give a thumbs up.
Best,
Shai Karmani
Hi Shai, thanks again for the insights — really useful thread.
One more thing worth adding: we've also been digging into the security side of the Foundry + MCP approach, which was one of our main concerns before committing to anything. The model looks promising — Entra-based agent identities, RBAC at the Agent Application level, and OAuth Identity Passthrough for per-user context. Part of our upcoming testing will focus specifically on security validation: understanding how identity propagates through the stack, what access control options are available at each layer, and how different authentication modes behave in practice — including for our mixed user base of internal and B2B guest users. That's something we need to validate properly before drawing any conclusions.
That said, we're not committing to Foundry yet. Performance, query quality, cost at scale, and security validation are still being tested over the coming weeks. But so far results are very promising — and if testing continues in this direction, Foundry + MCP will very likely be our choice.
If you have any recommendations or things you've tested and found useful along the way, they're very welcome.