Building Production-Grade AI Agents with Azure AI Foundry

This full-day advanced training focuses on building production-grade AI agents using Azure AI Foundry, starting from Model Context Protocol (MCP) and progressing toward a complete, intent-driven agent workflow. The session is practical and engineering-oriented, showing how real systems are composed rather than how isolated features work.

The day starts with Model Context Protocol as the foundation. Participants build an MCP server using Azure Functions, exposing tools and resources backed by a database. The MCP server is tested locally using the MCP Inspector, allowing participants to validate behavior, understand the contract, and debug issues in isolation.

Once the MCP layer is in place, an agent is created in Azure AI Foundry and linked to the MCP server. This section focuses on how agents invoke MCP tools and resources, how execution flows work, and how the same MCP endpoint can be reused by multiple agents. The same MCP server is then connected to Copilot Studio, demonstrating how Foundry agents and Copilot Studio copilots can share the same backend logic while providing different user experiences, including a small Copilot Studio demonstration.

The training then moves to knowledge bases, focusing on their real implementation through Azure AI Search. Knowledge bases are presented not as simple vector stores, but as hybrid retrieval systems combining vector search, keyword search, and semantic search. The session explains how a model controls the retrieval process, evaluates results, refines queries, and repeats searches when necessary to improve accuracy and grounding.

Knowledge bases are exposed through MCP endpoints, treating them as reusable services. Participants see how the same knowledge base can be consumed by both Azure AI Foundry agents and Copilot Studio through a consistent MCP URL, enabling predictable and repeatable retrieval behavior across tools.

The day concludes with a Foundry agent workflow that integrates both retrieval and structured data access in a single agent. The workflow identifies the user’s intent and routes the request accordingly, deciding whether to answer from the knowledge base (RAG backed by Azure AI Search), from the database exposed by the MCP server, or from a combination of both.

📚 Topics covered (table of contents)

🧩 MCP foundations (comes before agents)
➡ What MCP is and why it is the first building block
➡ MCP servers, tools, and resources
➡ Designing MCP endpoints for reuse and safety

🛠 Building an MCP server with Azure Functions
➡ Creating an MCP server
➡ Exposing database-backed tools and resources
➡ Local execution and debugging
➡ Testing endpoints using MCP Inspector

🧪 Testing MCP with MCP Inspector
➡ Installing the MCP Inspector
➡ Validating inputs and outputs
➡ Troubleshooting common MCP issues

🤖 Creating an agent in Azure AI Foundry
➡ Defining a Foundry agent
➡ Linking the agent to the MCP server
➡ Invoking MCP tools and resources
➡ Understanding execution flow

🔁 Reusing the same MCP backend in Copilot Studio
➡ Connecting Copilot Studio to the MCP server
➡ Small Copilot Studio demonstration
➡ Same backend, different user experience

📖 Knowledge bases backed by Azure AI Search
➡ Where knowledge bases are located
➡ Combining vector search, keyword search, and semantic search
➡ Model-controlled retrieval
➡ Query refinement and repeated search for accuracy

🔌 Exposing knowledge bases through MCP
➡ Knowledge bases accessed via MCP URLs
➡ Connecting knowledge bases to Foundry agents
➡ Reusing the same knowledge bases in Copilot Studio

🧠 Final build: intent-driven agent workflow
➡ Identifying user intent
➡ Routing between RAG (knowledge base) and structured data (database via MCP)
➡ Combining unstructured and structured responses
➡ Designing predictable and explainable agent behavior

- Understand the creation of MCP Servers - Understand knowledge bases - Create agents and workflows in Foundry