Introduction
Michel Schep
🤖 Microsoft Agent Framework

Getting Started
with MAF Agents
1.5

Build production-grade AI agents in .NET — from zero to running in minutes.

.NET 8 / 9 / 10 Open Source v1.5 — May 2026
Michel Schep 📅 May 9, 2026
1.5
Latest version
🚀
10K+
GitHub Stars
⭐
.NET + Py
Supported runtimes
🔷
01 / 10
Agenda
Michel Schep
Today's session

What we'll cover

01

What is an AI Agent?

The concept: LLM, Tools, Reasoning and Context — how they work together to create autonomous behaviour.

02

Introducing MAF

What Microsoft Agent Framework is, where it comes from, and why it matters for .NET developers.

03

The 3 Layers of a MAF Agent

OpenAI Client → IChatClient → AIAgent. How the abstraction stack fits together.

04

Live Demo + Aspire Observability

Build your first agent from scratch and watch it run — with full GenAI traces in .NET Aspire.

02 / 10
Concepts
Michel Schep
The concept

What is an AI Agent?

🧠 LLM — The Brain

A large language model (GPT-4o, Claude, Gemini…) that understands language, reasons over input, and decides what to do next.

🔧 Tools — The Hands

Functions the agent can call: search the web, query a database, call an API. The LLM decides when and how to use them.

📚 Context — The Memory

Conversation history, session state, user profile. Gives the agent awareness of what happened before this turn.

🔄 Autonomous loop 🎯 Goal-driven ⚡ Multi-step
The Agent Reasoning Loop
👤 User Input
↓
🗄️ RAG: Fetch Context
↓
🧠 LLM Reasoning + Context
↓
🔧 Tool Call?
↓
📥 Tool Result
↓
🧠 LLM Reasoning
↓
✅ Final Answer

An agent is not just a chatbot —
it thinks, decides, and acts.

03 / 10
Framework
Michel Schep
Microsoft Agent Framework

What is MAF?

The next generation of Semantic Kernel & AutoGen

Built by the same Microsoft teams. Combines AutoGen's simple agent abstractions with Semantic Kernel's enterprise features — session state, middleware, telemetry, and type safety.

🌐 Multi-provider support

Azure OpenAI · OpenAI · Anthropic · Gemini · GitHub Copilot · Ollama · Bedrock — switch with one line of code.

MIT License Open Source GA April 2026

🤝 Agents

Individual agents that use LLMs, call tools and MCP servers, and generate responses. Simple API, powerful runtime.

🕸️ Workflows

Graph-based multi-agent orchestration: sequential, concurrent, group chat, handoff, and Magentic patterns.

📡 Interoperability

A2A (Agent-to-Agent), AG-UI protocol, MCP servers, Azure Foundry hosting — works with the whole ecosystem.

04 / 10
Architecture
Michel Schep
Under the hood

The 3 Layers of a MAF Agent

🔌

Layer 1 — OpenAI Client

The raw connection to your LLM provider. Handles authentication, endpoint routing, and HTTP. Swap providers without changing anything above.

Azure.AI.OpenAI
→
🔁

Layer 2 — IChatClient

The standard AI abstraction from Microsoft.Extensions.AI. Provider-agnostic interface — swap models without changing agent code above.

M.Extensions.AI
→
🤖

Layer 3 — AIAgent

The MAF agent: adds tools, context, session state, instructions, and multi-turn conversation on top of IChatClient.

Microsoft.Agents.AI

💡 Each layer adds capabilities — but you only need the ones you use. Start with just OpenAI Client + AIAgent and add layers as you grow.

05 / 10
Observability
Michel Schep
Logging & Telemetry

Visibility into your agent — two lines

📋 .UseLogging(loggerFactory)

Plugs into the standard Microsoft.Extensions.Logging pipeline. Every agent action — LLM calls, tool invocations, errors — appears as structured log entries. Works with any log sink: console, Seq, Application Insights.

M.Extensions.Logging

📡 .UseTelemetry()

Emits OpenTelemetry spans following the GenAI semantic conventions. Captures model name, token counts, prompt content, tool calls, and latency — all as structured trace data.

OpenTelemetry GenAI conventions

💡 Add both to your AIAgentBuilder chain — they're composable middleware. Zero config required beyond a LoggerFactory.

What you get out of the box
🔍 Every LLM call logged

Model, tokens used, latency, prompt + completion content

🔧 Every tool call traced

Function name, input args, result, duration — as child spans

💰 Token usage tracked

Prompt tokens, completion tokens, per call and aggregated

🚨 Errors surfaced

Failed tool calls and LLM errors appear as error spans with full context

06 / 10
Code
Michel Schep
Build your first agent

From zero to running in 15 lines

Program.cs 
// Layer 1 — Connect to your LLM provider
var openAiClient = new OpenAIClient(
    new ApiKeyCredential(Environment.GetEnvironmentVariable("OPENAI_API_KEY")!));

// Layer 2 — Wrap as IChatClient (Microsoft.Extensions.AI)
IChatClient chatClient = openAiClient
    .AsChatClient(modelId: "gpt-4o-mini");

// Layer 3 — Build your AIAgent with observability
AIAgent agent = new AIAgentBuilder()
    .WithInstructions("You are a helpful assistant.")
    .UseChatClient(chatClient)
    .UseLogging(loggerFactory)    // → structured logs in Aspire
    .UseTelemetry()              // → OpenTelemetry GenAI traces
    .Build();

// Run it!
string answer = await agent.RunAsync("What is MAF 1.5?");
Console.WriteLine(answer);

📦 NuGet packages

Microsoft.Agents.AI
Microsoft.Agents.AI.OpenAI
Microsoft.Extensions.AI

🔌 Layer 1

Raw provider connection. Swap to AzureOpenAIClient, Anthropic or Gemini without touching anything else.

🔁 Layer 2

IChatClient is the standard .NET abstraction — works with any AI extension in the ecosystem.

📊 Observability

Two lines give you full GenAI trace data in .NET Aspire — token counts, latency, tool calls.

07 / 10
Observability
Michel Schep
.NET Aspire + GenAI

See exactly what your agent does

💬 Prompt & response inzichtelijk

Aspire toont de exacte user input en assistant output per LLM call — geen zwarte doos meer. Ideaal voor debugging en prompt tuning.

⏱️ Duration & token count

Elke LLM span toont duration (bijv. 1.79s) en token gebruik (bijv. 26 tokens) — direct zichtbaar zonder extra logging code.

🌳 Volledige trace hiërarchie

Van HTTP request → AgentRunner → invoke_agent → chat gpt-5.1: zie precies waar in de call stack de LLM wordt aangeroepen.

⚡ Gewoon .UseTelemetry()

Één regel in je AIAgentBuilder. Aspire pikt het automatisch op via OpenTelemetry — zero extra configuratie.

08 / 10
Demo
Michel Schep
💻
Live Demo

Let's build it live

Building a MAF 1.5 agent from scratch — in the IDE, step by step.

🔌 OpenAI Client
→
🔁 IChatClient
→
🤖 AIAgent
→
📊 Aspire
09 / 10
Get Started
Michel Schep
🤖

Start building today

MAF 1.5 is production-ready, open source, and designed for .NET developers. Three layers, fifteen lines of code, and you have a running AI agent with full observability.

📦 NuGet: Microsoft.Agents.AI → 📖 learn.microsoft.com/agent-framework → 🐙 github.com/microsoft/agent-framework → 💬 Discord: aka.ms/maf-discord → 🐙 github.com/michelschep → 💼 linkedin.com/in/michelschep →
Michel Schep Questions? 🙋
10 / 10