微软Agent Framework工作流：多智能体编排的实战指南

In my previous post, we explored how to create your first agent using the Microsoft Agent Framework, laying the foundation for building intelligent AI systems. Now, we take the next step: getting started with workflows. Workflows are the glue that binds agents together, enabling them to collaborate and execute tasks seamlessly. In this post, we’ll dive into the basics of creating workflows, from simple sequential processes to concurrent systems, helping you build smarter and more efficient AI solutions.

在我的上一篇文章中，我们探讨了如何使用 Microsoft Agent Framework 创建你的第一个智能体，为构建智能 AI 系统奠定了基础。现在，我们迈出下一步：开始使用工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。。工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。是将智能体粘合在一起的纽带，使它们能够无缝协作并执行任务。在本文中，我们将深入探讨创建工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。的基础知识，从简单的顺序流程到并发系统，帮助你构建更智能、更高效的 AI 解决方案。

Overview: The Power of Workflows

Workflows are where the Agent Framework truly shines. They allow you to connect agents so they can work as a team to tackle larger, more complex problems—tasks that a single agent cannot accomplish alone. In my experience as a Cloud Solution Architect, I've observed many teams struggle with multi-agent setups: agents calling each other inefficiently, losing track of context, or facing significant debugging challenges. Workflows are designed to solve these exact problems, providing memory, teamwork, and orchestration capabilities within a single, cohesive SDK.

工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。是 Agent Framework 真正大放异彩的地方。它们允许你连接多个智能体，使其能够以团队形式协作，解决更大、更复杂的问题——这些是单个智能体无法独立完成的任务。根据我作为云解决方案架构师的经验，我观察到许多团队在多智能体设置上遇到困难：智能体之间调用效率低下、上下文丢失或面临严重的调试挑战。工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。正是为解决这些问题而设计，在一个统一、连贯的 SDK 中提供了记忆、团队协作和编排能力。

Core Components

The workflow framework is built upon four core layers that work in concert to create a flexible and type-safe execution environment:

Executors and Edges: Form a directed graph that visually represents the structure and flow of the workflow.
Workflows: Orchestrate the execution of executors, manage message routing between agents, and handle event streaming.
Events: Provide crucial observability into the workflow's execution, allowing for monitoring and debugging.

工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。框架建立在四个核心层之上，它们协同工作，创建一个灵活且类型安全的执行环境：

执行器工作流框架的核心组件之一，与边共同构成有向图，表示工作流的结构和执行逻辑。和边工作流中连接智能体的路径，定义消息在智能体之间的流动方向和顺序。：形成一个有向图，直观地表示工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。的结构和流程。

工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。：编排执行器工作流框架的核心组件之一，与边共同构成有向图，表示工作流的结构和执行逻辑。的执行，管理智能体之间的消息路由，并处理事件流。

事件：为工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。的执行提供重要的可观测性，便于监控和调试。

Why Workflows? A Comparative Perspective

Having worked extensively with both Semantic Kernel and AutoGen, I understand the complexities that arise when attempting to connect agents without a structured plan.

Semantic Kernel is stable and reliable, but it lacks inherent flexibility for multi-agent orchestration. Connecting agents often requires writing substantial amounts of custom, boilerplate code.
AutoGen excels at providing strong multi-agent interaction patterns—agents can debate, reflect, and collaborate effectively. However, transforming these research-focused patterns into production-ready, robust systems demands significant additional engineering effort.

The Workflows feature in the Microsoft Agent Framework builds upon the strengths of these predecessors while dramatically simplifying the developer experience. Think of workflows as blueprints for orchestration—you plug in AI agents as modular building blocks, while retaining full control over the overall process logic.

在深入使用 Semantic Kernel 和 AutoGen 之后，我理解在没有结构化计划的情况下尝试连接智能体所带来的复杂性。

Semantic Kernel 稳定可靠，但在多智能体编排方面缺乏固有的灵活性。连接智能体通常需要编写大量自定义的样板代码。

AutoGen 擅长提供强大的多智能体交互模式——智能体可以有效地辩论、反思和协作。然而，将这些以研究为中心的模式转化为可用于生产的健壮系统，需要大量的额外工程工作。
Microsoft Agent Framework 中的工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。功能建立在这些前辈优势的基础上，同时极大地简化了开发者的体验。可以将工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。视为编排的蓝图——你将 AI 智能体作为模块化构建块插入，同时保留对整体流程逻辑的完全控制。

When to Use a Workflow vs. a Single Agent

A common question is: when should you graduate from a single agent to a workflow? Here’s a practical guide:

Scenario	Single Agent	Workflow
Simple Q&A	✅ Ideal	❌ Overkill
多步骤数据处理	❌ 过于复杂	✅ 理想选择
Parallel Analysis Tasks	❌ Sequential by nature	✅ Built for concurrency
需要人工审批	❌ Difficult to implement	✅ Native support
可复用的任务流水线	❌ Requires copy-paste	✅ Composable by design

My rule of thumb: If you find yourself writing extensive if/else logic or complex state management code just to coordinate multiple agents, it's time to adopt a workflow.

我的经验法则： 如果你发现自己编写了大量的 if/else 逻辑或复杂的状态管理代码，仅仅是为了协调多个智能体，那么就是时候采用工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。了。

Essential Workflow Patterns

Let's explore the most practical and commonly used workflow patterns, moving from simple to complex.

📐 Pattern 1: Sequential Pipeline (The Most Common)

This is the foundational pattern, ideal for content generation, data processing, and review cycles where each step depends on the output of the previous one.

When to use: For any linear, step-by-step process.
Example: A documentation pipeline where one agent researches a topic, a second agent writes the draft, and a third agent reviews for quality and clarity.

何时使用： 适用于任何线性的、逐步执行的流程。
示例： 一个文档生成流水线，其中一个智能体研究主题，第二个智能体撰写草稿，第三个智能体审查质量和清晰度。

🔀 Pattern 2: Concurrent Workflow (Fan-Out/Fan-In)

This pattern enables multiple agents to work simultaneously on the same input, significantly improving efficiency for independent tasks.

When to use: When you have subtasks that are independent and can be executed in parallel for speed.
Example: Analyzing a business proposal. One agent evaluates financial viability, another assesses market fit, and a third reviews technical feasibility—all running concurrently. Their results are later combined (fanned-in) for a comprehensive assessment.

何时使用： 当你拥有可以独立执行且能并行运行以提高速度的子任务时。
示例： 分析商业提案。一个智能体评估财务可行性，另一个评估市场匹配度，第三个审查技术可行性——三者同时运行。它们的结果随后被合并，以进行综合评估。

🌳 Pattern 3: Branching Logic (Conditional Routing)

This pattern introduces decision-making into the workflow, dynamically routing tasks to the most appropriate specialist agent based on content, classification, or other criteria.

When to use: For support ticket systems, document classification, or multi-stage approval flows.
Example: A customer service workflow. An initial "Classifier" agent analyzes the incoming query. Based on its intent (e.g., "billing," "technical," "general inquiry"), the workflow conditionally routes the task to the corresponding specialist agent (BillingAgent, TechSupportAgent, GeneralAgent).

何时使用： 适用于支持工单系统、文档分类或多阶段审批流程。
示例： 一个客户服务工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。。一个初始的“分类器”智能体分析传入的查询。根据其意图（例如，“账单”、“技术”、“一般咨询”），工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。有条件地将任务路由到相应的专家智能体（BillingAgent、TechSupportAgent、GeneralAgent）。

🤝 Pattern 4: Handoff (Agent-to-Agent Context Transfer)

This pattern facilitates the seamless transfer of a task—along with its full conversation context—from one specialist agent to another.

When to use: For escalation paths in support systems or when a task requires expertise from different domains in sequence.
Example: An IT support escalation workflow. A Level 1 (L1) agent performs initial triage. If the issue is complex, it hands off the complete conversation history to a Level 2 (L2) agent for deeper investigation. The L2 agent can then further hand off to a Level 3 (L3) expert for resolution, with context preserved throughout.

何时使用： 适用于支持系统中的升级路径，或者当一项任务需要依次来自不同领域的专业知识时。
示例： 一个 IT 支持升级工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。。一级（L1）智能体执行初步分类。如果问题复杂，它会将完整的对话历史移交给二级（L2）智能体进行深入调查。然后，L2 智能体可以进一步移交给三级（L3）专家解决，整个过程中上下文都得以保留。

Practical Implementation: Building a Sequential Workflow

Let's translate theory into practice by building the most common pattern: a sequential content pipeline.

The Scenario

We'll create a Content Pipeline that:

Researches a given topic.
Writes a draft based on the research findings.
Reviews the draft for quality, tone, and accuracy.

Code Breakdown

# 1. Agent Creation — The Foundation
researcher = ChatAgent(
    chat_client=client,
    name="ResearchAgent",
    instructions="You are a research specialist. Gather comprehensive, factual information about the given topic."
)
writer = ChatAgent(
    chat_client=client,
    name="WriterAgent",
    instructions="You are a technical writer. Create clear and engaging content based on the provided research notes."
)
reviewer = ChatAgent(
    chat_client=client,
    name="ReviewerAgent",
    instructions="You are an editor. Review the draft for clarity, accuracy, tone, and grammatical errors. Provide the final polished version."
)

# 2. Workflow Construction — The Blueprint
workflow = Workflow()
workflow.add_agent(researcher)
workflow.add_agent(writer)
workflow.add_agent(reviewer)

# 3. Edge Definition — The Flow
workflow.add_edge("ResearchAgent", "WriterAgent")  # Research passes to Writer
workflow.add_edge("WriterAgent", "ReviewerAgent") # Writer passes to Reviewer

# 4. Execution — Running the Pipeline
result = await workflow.run("Write a blog post about the benefits of AI agents in enterprise automation.")

What's happening?

Agent Creation: Each agent is defined with a specific role and instructions, much like assigning jobs to team members.
Workflow Construction: The Workflow object acts as a container. Agents are added to it like building blocks.
Edge Definition: Edges establish the communication pathways, defining the sequence: Research → Write → Review.
Execution: A single run() call activates the entire pipeline. The output from one agent automatically becomes the input for the next.

发生了什么？

智能体创建：每个智能体都被赋予特定的角色和指令，就像给团队成员分配工作一样。

工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。构建：Workflow 对象充当一个容器。智能体像构建块一样被添加到其中。

边工作流中连接智能体的路径，定义消息在智能体之间的流动方向和顺序。定义：边工作流中连接智能体的路径，定义消息在智能体之间的流动方向和顺序。建立了通信路径，定义了执行顺序：研究 → 撰写 → 审查。

执行：一个简单的 run() 调用即可激活整个流水线。一个智能体的输出会自动成为下一个智能体的输入。

Visualizing and Debugging with DevUI开发用户界面，提供工作流可视化功能，帮助开发者实时查看智能体连接、数据流和执行状态。

One of the most powerful features for development and operations is the built-in Developer UI (DevUI开发用户界面，提供工作流可视化功能，帮助开发者实时查看智能体连接、数据流和执行状态。). It provides real-time, interactive visualization of your workflow's execution.

Live Graph: See your agents as nodes and message flows as edges in a dynamic graph.
Execution Tracing: Monitor the flow of data step-by-step, observe agent reasoning, and inspect inputs/outputs.
Debugging Aid: Quickly identify bottlenecks, failed steps, or unexpected routing behavior.

This transforms workflow development from a "black box" exercise into a transparent and manageable process, significantly accelerating iteration and troubleshooting.

用于开发和运维的最强大功能之一是内置的开发者用户界面（DevUI开发用户界面，提供工作流可视化功能，帮助开发者实时查看智能体连接、数据流和执行状态。）。它提供了工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。执行的实时、交互式可视化。

实时图表：在一个动态图表中，将你的智能体视为节点，消息流视为边工作流中连接智能体的路径，定义消息在智能体之间的流动方向和顺序。。

执行跟踪：逐步监控数据流，观察智能体的推理过程，并检查输入/输出。

调试辅助：快速识别瓶颈、失败的步骤或意外的路由行为。
这将工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。开发从“黑盒”操作转变为透明且可管理的过程，显著加快了迭代和故障排除的速度。

Conclusion and Next Steps

Workflows represent the natural evolution from single-agent development to sophisticated multi-agent orchestration. The Microsoft Agent Framework brings together the best of prior paradigms:

The stability and structure reminiscent of Semantic Kernel.
The powerful orchestration patterns inspired by AutoGen.
Enterprise-grade reliability, observability, and developer tooling.

The real breakthrough is not merely technological—it's a shift in mindset. When connecting intelligent agents becomes as intuitive as drawing a graph, developers can stop wrestling with "how" to wire things together and start focusing on "what" complex problems they can solve.

工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。代表了从单智能体开发到复杂多智能体编排的自然演进。Microsoft Agent Framework 融合了先前范式的优点：

让人联想到 Semantic Kernel 的稳定性和结构性。

受 AutoGen 启发的强大编排模式。

企业级的可靠性、可观测性和开发者工具。
真正的突破不仅仅是技术性的，更是一种思维方式的转变。当连接智能体变得像画图一样直观时，开发者可以不再纠结于“如何”将各个部分连接起来，而是开始专注于“可以解决什么”复杂问题。

What's Next?
In this post, we covered the fundamentals of workflows, their core benefits, and essential patterns. In the next installment of this series, I will demonstrate how to create and manage these powerful workflows directly from the Microsoft AI Foundry portal using a no-code, configuration-driven approach. You'll see how to design, deploy, and publish agent-based solutions visually, making advanced AI orchestration accessible to an even wider audience.

For hands-on code samples, templates, and the complete demo project referenced here, explore the dedicated workflow folder in my GitHub repository: My Agent Framework Journey Repo.

下一步是什么？
在本文中，我们介绍了工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。的基础知识、核心优势和基本模式。在本系列的下一篇文章中，我将演示如何直接通过 Microsoft AI Foundry 门户，使用无代码、配置驱动的方法来创建和管理这些强大的工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。。你将看到如何可视化地设计、部署和发布基于智能体的解决方案，让更广泛的用户能够使用先进的 AI 编排功能。
关于实践代码示例、模板以及本文引用的完整演示项目，请访问我的 GitHub 仓库中的工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。专用文件夹：My Agent Framework Journey Repo。

I'd love to hear your thoughts! What workflow patterns are you most excited to implement? What specific use cases would you like to see covered in future posts? Share your ideas in the comments or connect with me on LinkedIn. Let's continue the conversation!

我很想听听你的想法！你对实现哪种工作流基于图的系统，连接多个智能体和函数以执行复杂的多步骤任务，支持类型路由、嵌套、检查点和人机交互模式。模式最感兴趣？你希望未来的文章涵盖哪些具体的用例？请在评论区分享你的想法，或在 LinkedIn 上与我联系。让我们继续交流！

Helpful Resources

#AgenticAI #MicrosoftAgentFramework #Workflows #MultiAgentSystems #AIOrchestration #AzureAI #DevUI开发用户界面，提供工作流可视化功能，帮助开发者实时查看智能体连接、数据流和执行状态。 #EnterpriseAI