Microsoft Agent Framework 1.0: .NET and Python 2026

Release Facts

Semantic Kernel + AutoGen Merger

The architectural story is cleaner than people expected. Semantic Kernel does not go away — it becomesthe foundation layer. AutoGen's orchestration concepts get re-implemented as a graph workflow engine on top of that foundation.

Architecture Overview

Five layers, bottom to top:

1. Connectors — provider-specific adapters for Azure OpenAI, OpenAI, Anthropic Claude, Bedrock, Gemini, Ollama. Swappable with a single registration line.
2. Kernel — the DI container and configuration surface. Inherited from Semantic Kernel.
3. Agents — the first-class agent primitive: instructions, tools, memory, state. Pluggable at the agent level.
4. Orchestration — the graph workflow engine. Multi-agent patterns (round-robin, supervisor, hierarchical) live here.
5. Interop layer — MCP and A2A protocol adapters for talking to external tools and other-framework agents.

.NET Setup and Hello World

dotnet new console -n HelloAgent
cd HelloAgent
dotnet add package Microsoft.Agents.AI
dotnet add package Microsoft.Agents.AI.OpenAI

// Program.cs
using Microsoft.Agents.AI;
using Microsoft.Agents.AI.OpenAI;

var agent = new ChatAgent(
    name: "Concierge",
    instructions: "You are a concise agency concierge. Answer in one paragraph.",
    model: new OpenAIChatClient("gpt-4.1", Environment.GetEnvironmentVariable("OPENAI_API_KEY")!)
);

await foreach (var token in agent.RunStreamingAsync("Explain MCP in 3 sentences."))
{
    Console.Write(token.Text);
}

That is the full shape of a production-ready .NET agent: dependency-inject the connector, construct a ChatAgent, stream tokens. IAsyncEnumerable is the canonical streaming type — works with await foreach and with async pipelines in ASP.NET endpoints.

Python Setup and Hello World

pip install microsoft-agents-ai microsoft-agents-ai-openai

# main.py
import os
import asyncio
from microsoft.agents.ai import ChatAgent
from microsoft.agents.ai.openai import OpenAIChatClient

async def main():
    agent = ChatAgent(
        name="Concierge",
        instructions="You are a concise agency concierge. Answer in one paragraph.",
        model=OpenAIChatClient("gpt-4.1", api_key=os.environ["OPENAI_API_KEY"]),
    )

    async for token in agent.run_streaming("Explain MCP in 3 sentences."):
        print(token.text, end="", flush=True)

asyncio.run(main())

Same concepts, same ergonomics, idiomatic in each language. The Python API uses snake_case; the .NET API uses PascalCase. Otherwise the shape is identical — intentionally, so cross-team work stays portable.

Provider Matrix

MCP + A2A Interoperability

MCP and A2A are the two interop protocols that make Agent Framework 1.0 a first-class 2026 citizen. MCP is tools; A2A is agents.

MCP: dynamic tool discovery

// Attach any MCP server — Google Ads MCP, GitHub MCP, your own custom server
var agent = new ChatAgent("TradingAgent", instructions: "...")
    .WithMcpServer("https://mcp.example.com/trading")
    .WithMcpServer(new StdioMcpServer("pipx", "run", "--spec", "...", "google-ads-mcp"));

// Tools appear automatically; no code changes as the server catalog evolves.

A2A: cross-framework coordination

A2A 1.0 lets an Agent Framework agent coordinate with agents running in other frameworks (Hermes Agent, LangChain, custom implementations) via structured, protocol-driven messaging. Same wire format, regardless of the runtime.

Multi-Agent Orchestration Patterns

The graph workflow engine surfaces four canonical patterns out of the box. Each maps to a node topology in the workflow DSL.

DevUI: The Local Debugger

DevUI is a browser-based local debugger shipped alongside 1.0. Launch it with agent-framework devui; it surfaces agent execution traces, message flows, tool calls, and orchestration decisions in real time. Think Chrome DevTools for multi-agent workflows.

Azure App Service Deployment

Microsoft published an explicit Azure App Service reference architecture alongside 1.0. The pattern is intentionally uninteresting — and that is the point. Agents deploy like any other ASP.NET Core or FastAPI app.

1. Provision Azure App Service (Linux or Windows), Azure OpenAI, and a managed identity.
2. Grantthe App Service's managed identity access to the Azure OpenAI resource — no API keys in configuration.
3. Publish the app through your CI/CD. Slot-based deployments for blue/green rollouts.
4. Monitorwith Application Insights — Agent Framework's OpenTelemetry traces flow in automatically.
5. Scale via App Service plan autoscale rules based on CPU or queue depth.

Conclusion

Agent Framework 1.0 is the first enterprise agent SDK that ships as a 1.0 with both Semantic Kernel lineage and AutoGen orchestration patterns. For Microsoft-centric stacks, this is the canonical choice through 2027. For cross-platform teams, the Python SDK plus A2A interop keeps it a viable option even when the rest of your agent stack runs elsewhere.

Frequently Asked Questions