Microsoft Agent Framework (.NET) — 1.0.0-rc4

Verified against dotnet-1.0.0-rc4 tag of microsoft/agent-framework.

Quick Reference

Item	Value
Core package	Microsoft.Agents.AI
Provider package	Microsoft.Agents.AI.OpenAI (Azure OpenAI / OpenAI)
Namespace	Microsoft.Agents.AI
Base class	AIAgent — all agents derive from this
Requires	.NET 10+ recommended, .NET 8+ supported
Repo	github.com/microsoft/agent-framework
Docs	learn.microsoft.com/en-us/agent-framework/

Architecture

┌─────────────────────────────────────────────────┐
│                  AIAgent (base)                  │
├──────────┬──────────┬───────────┬───────────────┤
│ Providers│  Tools   │  Memory   │  Middleware    │
│ AzureOAI │ Function │ Context   │ Agent Run     │
│ OpenAI   │ MCP      │ Providers │ Function Call │
│ Anthropic│ Code Int.│ Session   │ IChatClient   │
│ Ollama   │ File Srch│ StateBag  │               │
└──────────┴──────────┴───────────┴───────────────┘
         │                    │
    ┌────▼────┐         ┌────▼────┐
    │Workflows│         │ Hosting │
    │Executors│         │ A2A     │
    │Edges    │         │ AG-UI   │
    │Checkpoint         │ Azure Fn│
    └─────────┘         └─────────┘

Core Rules

• All agents derive from AIAgent — consistent interface for multi-agent scenarios.
• IChatClient from Microsoft.Extensions.AI is the core inference abstraction.
• Agent run middleware uses innerAgent.RunAsync() pattern — not a next delegate.
• BuildAIAgent() on IChatClient builder creates agent with chat client middleware pre-configured.
• Production: replace DefaultAzureCredential with ManagedIdentityCredential to avoid latency.
• Sessions are agent/service-specific — do not reuse across different agent configurations.
• [MessageHandler] on workflow executors uses compile-time source generation (Native AOT compatible).
• OpenTelemetry integration is built-in via Microsoft.Agents.AI package.

Cheat Sheet

Task	Pattern
Create agent	chatClient.AsAIAgent(instructions: ..., tools: [...])
Run (non-streaming)	await agent.RunAsync("prompt")
Run (streaming)	await foreach (var u in agent.RunStreamingAsync("prompt"))
Multi-turn	var s = await agent.CreateSessionAsync(); agent.RunAsync("msg", s)
Function tool	AIFunctionFactory.Create(MyMethod) → pass to tools:
Local MCP tools	McpClient.CreateAsync(transport) → ListToolsAsync() → cast to AITool
Hosted MCP	new HostedMcpServerTool(serverName, serverAddress)
MCP approval	Check AgentResponse for McpServerToolApprovalRequestContent
Memory/context	Implement AIContextProvider → pass to AIContextProviders in options
Agent middleware	agent.AsBuilder().Use(runMiddleware, null).Build()
Function middleware	agent.AsBuilder().Use(funcMiddleware).Build()
Chat middleware	chatClient.AsBuilder().Use(getResponseFunc: ...).BuildAIAgent(...)
Context middleware	agent.AsBuilder().UseAIContextProviders(provider).Build()
Serialize session	agent.SerializeSessionAsync(session)
Deserialize session	agent.DeserializeSessionAsync(json)
Workflow	WorkflowBuilder → AddEdge() → Build() → InProcessExecution.RunAsync()
Custom executor	partial class MyExec : Executor + [MessageHandler]
Azure Fn hosting	FunctionsApplication.CreateBuilder(args).ConfigureDurableAgents(opts => opts.AddAIAgent(...))

Additional Resources

Example Files

Complete, runnable .cs examples in examples/:

• examples/hello-agent.cs — Minimal setup, Azure OpenAI, streaming, Chat vs Responses provider, function tools
• examples/sessions-memory.cs — AgentSession multi-turn, session serialization, AIContextProvider implementation
• examples/middleware.cs — Agent run, function calling, IChatClient middleware; AIContextProvider as middleware
• examples/mcp-tools.cs — Local MCP (stdio), hosted MCP (Azure AI Foundry), tool approval (human-in-the-loop)
• examples/workflows.cs — WorkflowBuilder, lambda executor, custom Executor class, [MessageHandler], Azure Functions hosting

Reference Files

Detailed tables in references/:

• references/packages-protocols.md — All NuGet packages, provider client types, tool support matrix, protocols, workflow concepts