Microsoft is selling four ways to build an AI agent and a fifth product to govern them. Copilot Studio for business users, the Microsoft 365 Agents SDK for in-Microsoft 365 experiences, Foundry Agent Service for managed agents, and Microsoft Agent Framework for pro-code Python and .NET, with Microsoft Agent 365 sitting above all four as the tenant-wide control plane. The names overlap, the messaging blurs, and most teams I have advised end up paying for the wrong choice in wasted weeks. This book is the long answer to "which one, and how."
The book is built for the developer who has decided, or is about to decide, that pro-code is the right answer. After two short orientation chapters that frame the landscape and pin down where Microsoft Foundry fits underneath everything, the rest of the book is hands-on Microsoft Agent Framework, end to end. Every code listing is complete and runnable. Every API is verified against `agent-framework` 1.2.0 by a verification harness that runs on every commit to the companion repository at `github.com/rchaganti/maf-book`. Every chapter advances a single customer-support agent, from "answer a billing question" in chapter 3 to "approve refunds via the order database, escalate to a human, surface in Microsoft Teams, and register with the tenant's Agent 365 control plane" by the closing chapter.
Part I is the orientation. Two chapters that map the four-option landscape, introduce Microsoft Foundry as the platform underneath, and make the case for MAF, with a one-page comparison table and a decision tree the reader can pin somewhere visible.
Part II is the foundations. You install the framework, point it at an Azure OpenAI deployment in a Foundry project, and grow the fifteen-line hello-world from chapter 2 into the customer-support agent with a function tool and a streaming response. You learn the MAF client landscape, `OpenAIChatClient`, `OpenAIResponsesClient`, `FoundryChatClient`, and `FoundryAgent`, and the decision-tree question of which one to use when.
Part III extends the agent. Function tools in depth (with the `@tool` decorator, approval modes, validation), multi-turn conversation through `AgentSession`, structured outputs through `response_format`, and the persistent-agent path that connects directly back to Foundry's managed runtime.
Part IV is tools and context. The Model Context Protocol covered three ways: local MCP servers, Foundry-hosted MCP servers, and exposing your own agent as an MCP server. Agent Skills, the capability layer that complements MCP. Context providers and pluggable memory.
Part V is multi-agent orchestration. Sequential, concurrent, handoff, group-chat, and Magentic patterns, then the workflow primitives underneath them, executors, edges, events, and BSP supersteps, and how to visualize and checkpoint long-running workflows.
Part VI is production. Middleware for redaction, retries, and approvals. Human-in-the-loop function approvals. OpenTelemetry observability that flows into Foundry. The DevUI inspector for local development. Three production hosting paths compared: Azure App Service, Azure Functions, and Foundry-hosted agents.
Part VII is interoperability. Cross-framework communication via the Agent-to-Agent (A2A) protocol. AG-UI for production chat surfaces. The bridge back to Copilot Studio and Microsoft Teams via the M365 Agents SDK and the Microsoft Agent 365 SDK. A migration chapter for teams coming from AutoGen or Semantic Kernel.
By the end of the book, you have a working multi-agent system you can run, a mental model for extending it, and the operational story for taking it to production. Whether you are a Python developer building a customer-support backend, an engineer migrating from AutoGen, or an architect deciding between the four Microsoft authoring paths, the book gives you the pro-code path through the Microsoft enterprise AI landscape, with no rewrites, no fabricated APIs, and no hand-waving over the parts that matter.
