2.3.3. Foundry and Fabric Data Agents
💡 First Principle: Connected agents are Copilot Studio's embassy system — pro-code and data-platform specialists keep their citizenship (their home platform, tooling, and owners) while serving inside your conversational solution. The parent gains capabilities no low-code rebuild could match, because nothing is rebuilt.
A Foundry agent brings the pro-code stack of 1.4.2 into a conversation: an agent built by developers in Microsoft Foundry — possibly wielding fine-tuned models, custom code paths, and deep Azure integration — added to your Copilot Studio parent as a connected agent. The division of labor is clean: Copilot Studio owns the channel presence, governance, and conversational front door; Foundry owns the specialized reasoning. This is the standard exam answer when a requirement says "the data science team built an agent in Foundry; make it available in the existing Teams agent."
A Fabric data agent answers questions over enterprise data in Microsoft Fabric — lakehouses, warehouses, KQL databases — using Fabric's own understanding of the data estate. Connect one and your parent can field "which region's sales dipped last quarter?" without you modeling any of that data in Copilot Studio. Fabric data agents carry sharper constraints, which exams love: they cannot be invoked via topic redirect or referenced in instructions (orchestrator-routed only), and they are incompatible with deployment to Microsoft 365 Copilot.
Both integrations are the narrowest-change pattern from 1.4.2 in multi-agent form: keep the front door, connect the specialist. (Both are newer, preview-era surfaces — expect the exam to test the architecture and constraints, not pixel-level steps.)
⚠️ Common Misconception: Foundry and Fabric agents must be rebuilt in Copilot Studio to be used there. They are connected in place — configuration, not migration. The rebuild answer is nearly always the distractor.
Reflection Question: A scenario requires natural-language analytics over a Fabric lakehouse inside an agent deployed to Microsoft 365 Copilot. What conflict do you spot, and what would you push back on?
