Microsoft and Azure agent framework
Microsoft Agent Framework
Microsoft's open-source framework for building, orchestrating, and deploying AI agents and multi-agent workflows in Python, .NET, and Go, built as the direct successor to AutoGen and Semantic Kernel.
Quick decision
What it is
Microsoft Learn describes Agent Framework as the direct successor to AutoGen and Semantic Kernel, created by the same teams. It combines AutoGen's simple agent abstractions with Semantic Kernel's enterprise features, including session-based state management, type safety, middleware, and telemetry, and adds graph-based workflows for explicit multi-agent orchestration.
The building blocks are agents that use LLMs to process inputs, call tools and MCP servers, and generate responses across providers such as Microsoft Foundry, Anthropic, Azure OpenAI, OpenAI, and Ollama; a Harness agent with planning, context compaction, file access, and tool approval for long multi-step tasks; workflows that connect agents and functions through explicit execution paths; and integrations for model providers, tools, context providers, and middleware, with MCP clients for tool integration.
The framework ships as language SDKs rather than a hosted service. Python installs with pip install agent-framework, .NET uses the Microsoft.Agents.AI NuGet packages, and a Go module is published from the same repository. Microsoft publishes migration guides from both Semantic Kernel and AutoGen, because the APIs are new rather than a drop-in rename.
What it does
- Agents that call tools and MCP servers, with model clients for Foundry, Anthropic, Azure OpenAI, OpenAI, Ollama, and more
- Graph-based and functional workflows that connect agents and functions through explicit execution paths, with sequential, concurrent, handoff, and group chat orchestration patterns
- Agent sessions for state management and context providers for memory across multi-turn work
- A Harness agent for long multi-step tasks, with planning and todo tracking, context compaction, and do-not-ask-again tool approval
- Middleware that intercepts agent actions, plus telemetry carried over from the Semantic Kernel lineage
What it does not do
- A hosted agent service. Agent Framework is an SDK and runtime you run in your own process; hosting and model access are yours to arrange
- A drop-in upgrade for AutoGen or Semantic Kernel code. Microsoft ships migration guides from both because the programming model changed
- Responsibility for third-party systems. The Microsoft Learn overview states that building on third-party servers, agents, code, or non-Azure models is at your own risk and under their own license terms
Setup
Create a first Agent Framework agent (Python)
Install the agent-framework package from PyPI, then build an Agent on the OpenAI chat client, following the package README's simple-agent example. Set OPENAI_API_KEY first, or swap in another provider client such as Foundry.
pip install agent-framework
import asyncio
from agent_framework import Agent
from agent_framework.openai import OpenAIChatClient
async def main():
agent = Agent(
client=OpenAIChatClient(),
instructions="You are a friendly assistant. Keep your answers brief.",
)
result = await agent.run("What is the largest city in France?")
print(result)
asyncio.run(main())Environment variables (names only; use your own values)
OPENAI_API_KEY
Source: https://pypi.org/pypi/agent-framework/json. Examples use placeholders only.
MCP support: MCP clients are a foundational building block: agents connect to MCP servers for tool integration and can call tools and MCP servers directly, per the Microsoft Learn overview.
Health and maintenance
14,374
Checked 2026-10-29
2,535
Checked 2026-10-29
Stars, forks, MIT license, and active, not-archived status read from the GitHub API for microsoft/agent-framework on 2026-10-15, with the latest push the same day. PyPI showed agent-framework 1.23.0 as the current release in the 2026-10-29 check.
Verified compatible tools
Via Agent Framework MCP client connected to the official Tavily MCP server (tavily-mcp). The Microsoft Learn overview lists MCP clients among the framework's foundational building blocks and states agents call tools and MCP servers. Tavily publishes an official MCP server, catalogued on AgentsUse at /mcp/tavily, so the pairing runs over MCP rather than a Tavily-specific adapter. Evidence. Setup detail
Via Agent Framework MCP client connected to the official Exa MCP server (exa-mcp-server). The Microsoft Learn overview lists MCP clients among the framework's foundational building blocks and states agents call tools and MCP servers. Exa publishes an official MCP server, catalogued on AgentsUse at /mcp/exa, so the pairing runs over MCP rather than an Exa-specific adapter. Evidence. Setup detail
Via Agent Framework MCP client connected to the official Firecrawl MCP server (firecrawl-mcp). The Microsoft Learn overview lists MCP clients among the framework's foundational building blocks and states agents call tools and MCP servers. Firecrawl publishes an official MCP server, catalogued on AgentsUse at /mcp/firecrawl, so the pairing runs over MCP rather than a Firecrawl-specific adapter. Evidence. Setup detail
Grouped view: tools for Microsoft Agent Framework.
Sources and freshness
- Microsoft Learn: Agent Framework overview: https://learn.microsoft.com/en-us/agent-framework/overview/
- GitHub repository: https://github.com/microsoft/agent-framework
- PyPI agent-framework: https://pypi.org/pypi/agent-framework/json
Checked 2026-10-29. Source-verified means the claims above trace to the listed sources; it does not mean AgentsUse ran the framework. Send corrections through the contact page. Back to the framework index.