Microsoft Agent Framework

The framework has two main layers. An agent combines a chat client for a model provider with instructions, tools (function tools, local or hosted MCP tools), middleware and a session that carries conversation state. Chat history is kept in memory or in the model service, and context providers add longer-term memory. An opinionated harness agent adds planning, context compaction, file access and tool-approval defaults for long tasks.

A workflow is a graph of executors joined by edges. It runs in supersteps, and a checkpoint is written after each superstep with executor state, queued messages, pending requests and shared state, so runs can be resumed or restored. Agents can be executors, and a workflow can be wrapped back into the agent interface.

Multi-agent patterns are packaged as orchestrations built on workflows: sequential, concurrent, handoff, group chat and Magentic, where a manager agent plans and assigns work. Hosting options include Microsoft Foundry hosted agents, Azure Functions with a durable extension, and self-hosted A2A, MCP and OpenAI-compatible endpoints.

Agents: a tool marked as approval-required makes the run end with an approval request instead of a final answer; the caller collects the decision and passes it back in the next run. Workflows: an executor sends a request (ctx.request_info() with a @response_handler in Python, a RequestPort in .NET), the workflow emits a request event, and the response is routed back to that executor. Since Python 1.13.0 checkpoints record delivered responses, so a human-in-the-loop continuation can be replayed from a checkpoint. The built-in orchestrations accept tool approvals and request-info pauses, and the AG-UI integration carries approvals to a web frontend.

pip install agent-framework azure-identity

import asyncio, os
from agent_framework.foundry import FoundryChatClient
from agent_framework.orchestrations import SequentialBuilder
from azure.identity import AzureCliCredential

async def main() -> None:
    client = FoundryChatClient(  # run `az login` first
        project_endpoint=os.environ["FOUNDRY_PROJECT_ENDPOINT"],
        model=os.environ["FOUNDRY_MODEL"],
        credential=AzureCliCredential(),
    )
    writer = client.as_agent(name="writer", instructions="Write one punchy product sentence.")
    reviewer = client.as_agent(name="reviewer", instructions="Give brief feedback on the sentence.")
    workflow = SequentialBuilder(participants=[writer, reviewer]).build()
    events = await workflow.run("Tagline for a budget-friendly eBike.")
    for msg in events.get_outputs()[0].messages:
        print(f"[{msg.author_name}] {msg.text}")

asyncio.run(main())

• The official quickstarts use Microsoft Foundry with Azure CLI credentials; run az login first or switch to another provider's chat client.
• pip install agent-framework installs the standard package set; experimental modules come from agent-framework-lab, and several integrations (A2A, AG-UI, Claude, MCP hosting) need --pre.
• On minimal Python installs, MCP tools may need pip install mcp --pre.
• Python 1.13.0 changed workflow checkpoint behavior; code that depends on superstep counts or checkpoint ordering should follow the upgrade guide.
• The docs warn that DefaultAzureCredential is meant for development; use a specific credential in production.

Yes. Microsoft's AutoGen migration guide calls it the direct successor, built by the same teams, and AutoGen's README puts AutoGen in maintenance mode and points new users here.

Yes. Agents consume MCP tools in Python and .NET, Python can serve an agent over MCP, and A2A works as client (A2AAgent) and server through hosting packages.

Python and .NET in this repository. A Go SDK lives in microsoft/agent-framework-go and is in public preview.

The framework is MIT-licensed. Hosting on Microsoft Foundry, including Foundry hosted agents, is a paid Azure service.

Yes. The agent-framework-claude package provides ClaudeAgent, which wraps the Claude Agent SDK and requires the Claude Code CLI to be installed.