microsoft-agent-framework

Microsoft Agent Framework 技能

版本：0.1.0-preview | 最后更新：2025-11-15 | 框架版本：0.1.0-preview 语言：Python 3.10+、C#（.NET 8.0+）| 许可证：MIT

快速参考

Microsoft Agent Framework 是一个用于构建生产级 AI Agent 和工作流的开源平台，将 AutoGen 的简洁性与 Semantic Kernel 的企业级功能融为一体。

核心功能：AI Agent（有状态对话、工具集成）| 工作流（基于图的编排、并行处理）| 企业级功能（遥测、中间件、MCP 支持）

安装：

• Python：pip install agent-framework-core --pre
• C#：dotnet add package Microsoft.Agents.AI --prerelease

仓库：https://github.com/microsoft/agent-framework（5.1k 颗星）

---

何时使用此技能

在以下情况下使用 Microsoft Agent Framework：

1. 生产级 AI Agent，需要企业级功能（遥测、中间件、结构化输出）
2. 多 Agent 编排，通过基于图的工作流进行条件路由
3. 工具/函数集成，包含审批工作流和错误处理
4. 跨平台开发，需要同时使用 Python 和 C# 实现
5. 从研究到生产的流程，利用 AutoGen 与 Semantic Kernel 的融合

与 amplihack 集成：使用 Agent Framework 构建有状态的对话式 Agent和实现复杂编排。使用 amplihack 的原生 Agent 系统进行无状态任务委派和简单编排。详细指南请参阅 @integration/decision-framework.md。

---

核心概念

1. AI Agent

处理消息、调用工具并保留上下文的有状态对话实体。

Python 示例：

from agents_framework import Agent, ModelClient

# Create agent with model
agent = Agent(
    name="assistant",
    model=ModelClient(model="gpt-4"),
    instructions="You are a helpful assistant"
)

# Single-turn conversation
response = await agent.run(message="Hello!")
print(response.content)

# Multi-turn with thread
from agents_framework import Thread
thread = Thread()
response = await agent.run(thread=thread, message="What's 2+2?")
response = await agent.run(thread=thread, message="Double that")

C# 示例：

using Microsoft.Agents.AI;

var agent = new Agent(
    name: "assistant",
    model: new ModelClient(model: "gpt-4"),
    instructions: "You are a helpful assistant"
);

var response = await agent.RunAsync("Hello!");
Console.WriteLine(response.Content);

2. 工具与函数

通过提供可调用的函数扩展 Agent 的功能。

Python 示例：

from agents_framework import function_tool

@function_tool
def get_weather(location: str) -> str:
    """Get weather for a location."""
    return f"Weather in {location}: Sunny, 72°F"

agent = Agent(
    name="assistant",
    model=ModelClient(model="gpt-4"),
    tools=[get_weather]
)

response = await agent.run(message="What's the weather in Seattle?")
# Agent automatically calls get_weather() and responds with result

C# 示例：

[FunctionTool]
public static string GetWeather(string location)
{
    return $"Weather in {location}: Sunny, 72°F";
}

var agent = new Agent(
    name: "assistant",
    model: new ModelClient(model: "gpt-4"),
    tools: new[] { typeof(Tools).GetMethod("GetWeather") }
);

3. 工作流

用于多 Agent 系统的基于图的编排，支持条件路由和并行执行。

Python 示例：

from agents_framework import Workflow, GraphWorkflow

# Define workflow graph
workflow = GraphWorkflow()

# Add agents as nodes
workflow.add_node("researcher", research_agent)
workflow.add_node("writer", writer_agent)
workflow.add_node("reviewer", review_agent)

# Define edges (control flow)
workflow.add_edge("researcher", "writer")  # Sequential
workflow.add_edge("writer", "reviewer")

# Conditional routing
def should_revise(state):
    return state.get("needs_revision", False)

workflow.add_conditional_edge(
    "reviewer",
    should_revise,
    {"revise": "writer", "done": "END"}
)

# Execute workflow
result = await workflow.run(initial_message="Research AI trends")

C# 示例：

var workflow = new GraphWorkflow();

workflow.AddNode("researcher", researchAgent);
workflow.AddNode("writer", writerAgent);
workflow.AddNode("reviewer", reviewAgent);

workflow.AddEdge("researcher", "writer");
workflow.AddEdge("writer", "reviewer");

var result = await workflow.RunAsync("Research AI trends");

4. 上下文与状态管理

在多个 Agent 之间保留对话历史和共享状态。

Python：

from agents_framework import Thread, ContextProvider

# Thread maintains conversation history
thread = Thread()
await agent.run(thread=thread, message="Remember: My name is Alice")
await agent.run(thread=thread, message="What's my name?")  # "Alice"

# Custom context provider
class DatabaseContext(ContextProvider):
    async def get_context(self, thread_id: str):
        return await db.fetch_history(thread_id)

    async def save_context(self, thread_id: str, messages):
        await db.save_history(thread_id, messages)

agent = Agent(model=model, context_provider=DatabaseContext())

5. 中间件与遥测

添加日志记录、身份验证和监控等横切关注点。

Python：

from agents_framework import Middleware
from opentelemetry import trace

# Custom middleware
class LoggingMiddleware(Middleware):
    async def process(self, message, next_handler):
        print(f"Processing: {message.content}")
        response = await next_handler(message)
        print(f"Response: {response.content}")
        return response

# OpenTelemetry integration
tracer = trace.get_tracer(__name__)
with tracer.start_as_current_span("agent-run"):
    response = await agent.run(message="Hello")

C#：

public class LoggingMiddleware : IMiddleware
{
    public async Task<Message> ProcessAsync(Message message, Func<Message, Task<Message>> next)
    {
        Console.WriteLine($"Processing: {message.Content}");
        var response = await next(message);
        Console.WriteLine($"Response: {response.Content}");
        return response;
    }
}

---

常见模式

人在回路审批

from agents_framework import HumanInTheLoop

@function_tool
def delete_file(path: str) -> str:
    """Delete a file (requires approval)."""
    return f"Deleted {path}"

# Add approval wrapper
delete_file_with_approval = HumanInTheLoop(
    tool=delete_file,
    approval_prompt="Approve deletion of {path}?"
)

agent = Agent(tools=[delete_file_with_approval])

并行执行 Agent

workflow = GraphWorkflow()

# Add multiple agents
workflow.add_node("analyst1", analyst_agent)
workflow.add_node("analyst2", analyst_agent)
workflow.add_node("synthesizer", synthesis_agent)

# Parallel execution
workflow.add_edge("START", ["analyst1", "analyst2"])  # Both run in parallel
workflow.add_edge(["analyst1", "analyst2"], "synthesizer")  # Wait for both

result = await workflow.run(message="Analyze market trends")

结构化输出生成

from pydantic import BaseModel

class WeatherReport(BaseModel):
    location: str
    temperature: float
    conditions: str

agent = Agent(
    model=model,
    instructions="Generate weather reports",
    response_format=WeatherReport
)

response = await agent.run(message="Weather in Seattle")
report: WeatherReport = response.parsed
print(f"{report.location}: {report.temperature}°F, {report.conditions}")

错误处理与重试

from agents_framework import RetryPolicy

agent = Agent(
    model=model,
    retry_policy=RetryPolicy(
        max_retries=3,
        backoff_factor=2.0,
        exceptions=[TimeoutError, ConnectionError]
    )
)

try:
    response = await agent.run(message="Hello")
except Exception as e:
    print(f"Failed after retries: {e}")

---

与 amplihack 集成

决策框架

在以下情况下使用 Microsoft Agent Framework：

• 构建有状态的对话式 Agent（多轮对话）
• 需要企业级功能（遥测、中间件、身份验证）
• 需要通过条件路由实现复杂的多 Agent 编排
• 有跨平台需求（Python + C#）
• 需要与 Microsoft 生态系统集成（Azure、M365）

在以下情况下使用 amplihack 原生 Agent：

• 无状态任务委派（代码审查、分析）
• 简单的顺序/并行编排
• 基于文件的操作和本地工具
• 无需基础设施的快速原型开发
• 注重令牌效率的技能型架构

混合方案：

# Use amplihack for orchestration
from claude import Agent as ClaudeAgent

orchestrator = ClaudeAgent("orchestrator.md")

# Delegate to Agent Framework for stateful agents
from agents_framework import Agent, Thread

conversational_agent = Agent(
    model=ModelClient(model="gpt-4"),
    instructions="Maintain conversation context"
)

thread = Thread()
response1 = await conversational_agent.run(thread=thread, message="Start task")
response2 = await conversational_agent.run(thread=thread, message="Continue")

# Use amplihack for final synthesis
result = orchestrator.process({"responses": [response1, response2]})

完整模式请参阅 @integration/amplihack-integration.md。

---

快速入门工作流

1. 安装：pip install agent-framework-core --pre（Python），或 dotnet add package Microsoft.Agents.AI --prerelease（C#）

2. 创建基础 Agent：

   from agents_framework import Agent, ModelClient
   
   agent = Agent(
       name="assistant",
       model=ModelClient(model="gpt-4"),
       instructions="You are a helpful assistant"
   )
   
   response = await agent.run(message="Hello!")
   

3. 添加工具：

   @function_tool
   def calculate(expr: str) -> float:
       return eval(expr)
   
   agent = Agent(model=model, tools=[calculate])
   

4. 构建工作流：

   workflow = GraphWorkflow()
   workflow.add_node("agent1", agent1)
   workflow.add_node("agent2", agent2)
   workflow.add_edge("agent1", "agent2")
   result = await workflow.run(message="Task")
   

5. 添加遥测：

   from opentelemetry import trace
   tracer = trace.get_tracer(__name__)
   with tracer.start_as_current_span("agent-run"):
       response = await agent.run(message="Hello")
   

---

参考文档

详细信息请参阅：

• @reference/01-overview.md - 架构、组件、使用场景
• @reference/02-agents.md - Agent 创建、生命周期、高级功能
• @reference/03-workflows.md - 工作流模式、执行器、检查点
• @reference/04-tools-functions.md - 工具定义、审批工作流、错误处理
• @reference/05-context-middleware.md - 上下文提供程序、中间件模式、身份验证
• @reference/06-telemetry-monitoring.md - OpenTelemetry、日志记录、调试
• @reference/07-advanced-patterns.md - 多 Agent 模式、流式传输、DevUI

可运行示例

• @examples/01-basic-agent.py - 简单的对话式 Agent
• @examples/02-tool-integration.py - 支持函数调用的 Agent
• @examples/03-simple-workflow.py - 多 Agent 工作流
• @examples/04-basic-agent.cs - C# Agent 实现
• @examples/05-tool-integration.cs - C# 工具集成
• @examples/06-simple-workflow.cs - C# 工作流示例

维护

检查框架的新鲜度：python @scripts/check-freshness.py

当前版本跟踪信息：@metadata/version-tracking.json

---

Token 数量：约 4,200 个（低于 4,800 个上限）