Production-oriented Microsoft Agent Framework examples in C# — HITL, MCP, multi-agent, durability, OpenTelemetry and A2A

[Raj (rajshukla09)]

I've been working through a set of production-oriented patterns with Microsoft Agent Framework in C# and have published the complete implementations as two companion repositories.

Part 1 focuses on the foundations: structured outputs, conversations, sessions, persistence, tools, memory, context providers, reliable tool execution, and workflows.

Part 2 moves into human-in-the-loop workflows, MCP, dynamic and Magentic orchestration, multi-agent patterns, durable execution, OpenTelemetry, A2A, and production context engineering.

Each chapter is self-contained with implementation, tests, and setup documentation.

Part 1 repository:
https://github.com/rajshukla09/building-ai-agents-with-dotnet-part-1

Part 2 repository:
https://github.com/rajshukla09/building-ai-agents-with-dotnet-part-2

These repositories accompany my Building AI Agents with .NET books, but I've made the complete chapter implementations publicly available so they can also be explored independently.

I'd especially appreciate feedback from the Agent Framework community on the architecture and production patterns, and suggestions for scenarios worth exploring next.

[Harlock (capitaharlock)]

The A2A chapter is where I'd push hardest, because conformance is probed, not asserted. A card at /.well-known/agent.json that declares a bunch of returns tells you almost nothing until an external caller hits POST <card.url>/v1/<skill-id> with a real input and sees operational. Your implementation is already structured for that, but if the tests only check what the card says, you'll miss failures where the endpoint exists but doesn't match its declared skills.

A scenario worth adding next is an external A2A prober that runs against the live card, counts non-null declared returns, and flags drift between the card and the actual call. That's a production pattern that catches more than any self-reported metadata.