Updated November 2026

Building AI Agents in .NET with Microsoft Agent Framework

Class Duration

21 hours of live training delivered over 3 days.

Student Prerequisites

Professional C# development experience on modern .NET, including async/await, LINQ, and dependency injection (see Modern C# Development with .NET)
Familiarity with LLM fundamentals: prompts, tool calling, and structured outputs
Working knowledge of ASP.NET Core and the .NET generic host is helpful for the hosting modules
Basic Azure experience (subscriptions, resource groups, Microsoft Entra ID)
An Azure subscription with access to a Microsoft Foundry project and a model deployment

Target Audience

C# and .NET developers, architects, and platform teams building AI agents on Microsoft's stack, including teams with Semantic Kernel or AutoGen code to migrate. Copilot Studio and the wider Foundry platform are covered in Building AI Agents with Microsoft Foundry and Copilot Studio. For other stacks, see Building AI Agents with Agent SDKs.

Description

Microsoft Agent Framework is Microsoft's open-source SDK for AI agents and multi-agent workflows. It is the successor to Semantic Kernel and AutoGen, built by the same teams. This course works in C# from a first ChatClientAgent through deployment. Students add tools, MCP servers, memory, and middleware, then build checkpointed graph workflows and multi-agent orchestrations. Final modules cover A2A, observability, evaluation, hosting on Microsoft Foundry or Azure Container Apps, and migration.

Learning Outcomes

Place Agent Framework alongside Foundry, Copilot Studio, Semantic Kernel, and AutoGen.
Choose between an agent, a workflow, or plain code for a task.
Build ChatClientAgent agents on Foundry, Azure OpenAI, OpenAI, and other providers.
Extend agents with function tools, hosted tools, approvals, and MCP servers.
Manage multi-turn state with sessions, history providers, and context providers.
Apply middleware for guardrails, logging, and compliance.
Use the Agent Harness for long-running, multi-step tasks.
Build graph workflows that checkpoint, resume, and pause for human input.
Orchestrate agents with sequential, concurrent, handoff, group chat, and Magentic patterns.
Call and host remote agents over A2A.
Instrument, debug, and evaluate agents with OpenTelemetry, DevUI, and evaluators.
Deploy agents to Foundry or Azure Container Apps, secured with Microsoft Entra ID.
Plan a migration from Semantic Kernel and AutoGen.

Training Materials

Comprehensive courseware is distributed online at the start of class. All students receive a downloadable MP4 recording of the training.

Software Requirements

The .NET 10 SDK, Visual Studio or VS Code with the C# Dev Kit, an Azure subscription with a Microsoft Foundry project and model deployment, the Azure CLI, the Azure Developer CLI (azd) with the azure.ai.agents extension, Docker or Podman, and Git.

Training Topics

Agent Framework in the Microsoft AI Stack

Successor to Semantic Kernel and AutoGen
Agents, the harness, workflows, and integrations
IChatClient and AIFunction foundations
Shared message types
Foundry, Copilot Studio, and Microsoft 365 Copilot
Choosing an agent, a workflow, or plain code
Stable and prerelease packages
The Python and Go SDKs

Agents and Model Providers

AIAgent and ChatClientAgent
Creating agents with AsAIAgent
Foundry, Azure OpenAI, and OpenAI clients
Anthropic, Bedrock, Gemini, and Ollama
RunAsync and RunStreamingAsync
Structured output and multimodal input
Background responses
Remote agent services

Tools, Function Calling, and MCP

Function tools with AIFunctionFactory
[Description] metadata
Code interpreter, file search, and web search
Approval-required tools
Handling tool approval requests
MCP servers with the MCP C# SDK
Hosted MCP tools
Microsoft Foundry Toolbox
Agents as tools and MCP servers

Conversation State and Memory

Creating and resuming an AgentSession
Serializing sessions
Service-managed versus local chat history
ChatHistoryProvider
Context providers for memory and RAG
Built-in providers such as Mem0 and Redis
Compaction for long conversations
Scoping conversation IDs to users and tenants

Middleware and the Agent Pipeline

How a run flows through the pipeline
Agent run middleware
Function calling middleware
IChatClient middleware
Registering middleware with AsBuilder()
Guardrails and early termination
Result overrides and exception handling
Microsoft Purview middleware

The Agent Harness

HarnessAgent and AsHarnessAgent
Planning and todo tracking
File memory and Agent Skills
Standing tool approvals
Built-in web search and telemetry
File access, shell tools, and background agents
Bounded looping
Customizing defaults with HarnessAgentOptions

Graph-Based Workflows

WorkflowBuilder, executors, and edges
Conditional routing
Fan-out and fan-in
Supersteps and workflow events
Streaming runs with InProcessExecution
Shared state and sub-workflows
Agents as executors
Workflows as agents
Declarative workflows in YAML

Checkpointing and Human-in-the-Loop

Checkpoints with CheckpointManager
Restoring and rehydrating workflows
Stable identities for resumable workflows
RequestPort and RequestInfoEvent
Tool approvals inside workflows
Checkpoint storage as a trust boundary

Multi-Agent Orchestration Patterns

Sequential orchestration
Concurrent orchestration
Handoff orchestration with AgentWorkflowBuilder
Group chat managers
Magentic orchestration and progress ledgers
Stall detection and plan review
Agent-as-tool versus handoff

Agent-to-Agent (A2A) Interoperability

Agent cards and discovery
Calling remote agents with A2AAgent
Resolving cards with A2ACardResolver
Hosting agents with AddA2AServer
JSON-RPC and HTTP+JSON endpoints
Continuation tokens and stream reconnection
Durable session and task stores

Observability, Debugging, and Evaluation

OpenTelemetry with UseOpenTelemetry
GenAI semantic conventions
Application Insights and the Aspire dashboard
Sensitive data in telemetry
Local debugging with DevUI
LocalEvaluator and FoundryEvals
Microsoft.Extensions.AI.Evaluation
Repetitions and expected outputs
Evaluating workflows per agent

Hosting on Microsoft Foundry and Azure Container Apps

Foundry hosted agents versus self-hosting
The Responses and Invocations protocols
azd ai agent init and azd ai agent run
Deploying with azd deploy
Self-hosting with AddAIAgent and MapOpenAIResponses
Containers on Azure Container Apps
Durable agents on Azure Functions
The Durable Task Scheduler

Identity and Security

ManagedIdentityCredential in production
Developer credentials locally
Entra agent identities for hosted agents
Per-user session isolation
Authenticating MCP and A2A connections
On-Behalf-Of passthrough
Foundry guardrails and prompt-injection defenses
Secrets in Azure Key Vault

Migrating from Semantic Kernel and AutoGen

Semantic Kernel plugins to tools
ChatCompletionAgent to ChatClientAgent
AgentThread to AgentSession
InvokeAsync to RunAsync
Filters to middleware
AutoGen teams and GraphFlow to workflows
Planning an incremental migration