The Microsoft Agent Stack Decision Map

Someone asked you to ship an agent. Your company runs on Microsoft 365, so the answer is somewhere in Microsoft's stack, and that is exactly where it stops being obvious. Foundry, Copilot Studio, a Copilot extension, plain Azure OpenAI. Four ways to do it, all documented separately, none of them telling you which one is yours. So you pick the one you read about most recently, and you find out whether that was right in week three.

This is the page that settles it before you build. One side is a five-question router that dead-ends on a surface. The other is every tradeoff that matters, laid out side by side, so you can defend the choice to an architect who was not in the room when you made it.

Inside:

The five-question router. Start from who owns it after launch, not from the technology. Five questions, each resolving to a surface, and a last one that resolves to governance regardless of which you picked.

The comparison matrix. Foundry, Copilot Studio, Copilot extension and Azure OpenAI across seven rows: build surface, best for, grounding, identity and governance, cost model, control ceiling and time to first ship. The cost and control rows are the ones people forward to their architect.

Layers, not rivals. Why most production agents are a chain rather than a pick, and what the common shape actually is: a Foundry agent behind a Copilot Studio front door, published to Teams, registered in Agent 365.

The four expensive ways to get this wrong. Foundry for a FAQ bot, Copilot Studio for work that must run the same way every time, raw Azure OpenAI inside a Microsoft shop, and no Agent 365 registration. What each one costs you, and how late you find out.