The Two-Surface Approach
Why Two Surfaces?
Building an AI agent is iterative. You need fast feedback loops during development, but enterprise distribution for business users. This accelerator demonstrates both:
- GitHub Copilot CLI — Zero infrastructure. Connect MCP servers, write skills, iterate in your terminal. Perfect for prototyping and SE demos.
- Azure AI Foundry → M365/Teams — Same backends registered as Foundry platform tools. Published as an Agent Application with Entra identity, RBAC, and enterprise distribution.
Translation Mapping
| Capability | CLI (MCP Server) | Foundry (Platform Tool) | Same Backend? |
|---|---|---|---|
| Sales data queries | wwi-sales-data HTTP MCP | FabricIQPreviewTool | ✅ Same Data Agent |
| M365 activity | @microsoft/workiq MCP | WorkIQPreviewTool | ✅ Same WorkIQ service |
| Report generation | CLI skill (markdown) | Custom function + OneDrive | ✅ Same generator.py |
| Orchestration | Copilot CLI (built-in) | Foundry Responses API | Different — CLI is zero-code |
From Prototype to Production
Step 1: Build & test MCP servers locally
└─ Copilot CLI discovers tools automatically
└─ Fast iteration, no deployment needed
Step 2: Register same backends as Foundry tools
└─ FabricIQPreviewTool wraps the Data Agent MCP URL
└─ WorkIQPreviewTool wraps WorkIQ via A2A protocol
└─ Custom functions wrap report generator logic
Step 3: Publish Foundry Agent → M365/Teams
└─ Agent Application with stable endpoint
└─ Entra identity + RBAC
└─ Users @mention the agent in M365 Copilot Chat