Microsoft Agent Framework: build a travel assistant

Rakesh Lakshminarayana introduces Microsoft Agent Framework through a travel assistant and focused code examples. He describes it as an open source SDK for .NET and Python, with replaceable model and memory providers. The session explains the AI agent loop, then shows how context providers add information before a model call and how tools fetch data during execution.

The examples use C# and require the .NET 10 SDK. A React interface connects to an ASP.NET Core backend, while Docker and Aspire support development and trace inspection. The demonstrated application uses Azure OpenAI models, Azure Container Apps and Cosmos DB. Although the speaker covers running the application locally and mentions Ollama as another provider, the demo uses cloud services and does not establish an offline setup.

Skills load instructions and supporting files on demand to limit unnecessary context. The speaker notes that the skill provider is experimental and its API may change. An MCP server supplies flight search and booking tools; booking waits for user approval. The full application stores travel preferences in Cosmos DB and uses AGUI with CopilotKit for result cards and approval prompts. The final walkthrough splits responsibilities between specialist agents and a triage agent that routes requests through a handoff workflow.