DEV Community: Rick

AgentCrew MCN 架构设计详解

Rick — Tue, 21 Jul 2026 09:11:16 +0000

AgentCrew MCN 架构设计详解

从一个需求说起：假如我们要搭建一个智能营销内容工厂，需要多个 AI Agent 协同工作——研究员收集信息、撰稿人撰写文章、设计师配图、审核员检查合规。如何让这些 Agent 既独立又能高效协作，同时还能被产品经理轻松调度？正是为了解决这类问题，我们设计了 AgentCrew MCN（Multi‑agent Collaboration Network）——一个可扩展、可观测、低代码的智能体协作架构。

本文将深入 AgentCrew MCN 的内核，从设计理念、核心组件、通信模型到实际代码示例，全面解析这一架构是如何让多个 Agent 像一支专业 Crew（团队）一样工作的。

1. 设计目标：为什么需要 MCN？

当前多数 AI Agent 框架侧重于单个 Agent 的能力强化（ReAct、Plan‑and‑Execute 等），但在真实业务中，往往需要多个 Agent 组成流水线或动态协作网络。AgentCrew MCN 的设计目标就是填补“单 Agent 到多 Agent 协作”之间的空白，具体包括：

解耦协作逻辑 ：将任务分配、消息传递、状态同步从 Agent 业务逻辑中剥离，让 Agent 只关注自身职责。
动态组网：Crew（团队）可以根据任务需求动态组建与销毁，支持任意拓扑（顺序、并行、条件分支）。
可靠通信：确保消息可靠投递、故障重试与超时处理，兼容同步与异步调用。
可观测性：内置分布式追踪与日志，让每次多 Agent 对话全链路可监控。
低代码编排：通过 YAML/JSON 配置或 Python SDK 快速定义协作流程。

2. 核心架构

AgentCrew MCN 的整体架构可以概括为“一个总线、两类节点、三层抽象”：

一个总线 —— CrewBus：基于消息队列的异步消息中枢，负责 Agent 之间的通信与事件分发。
两类节点 —— Agent 节点与 Crew 控制器：
- Agent 节点：挂载了特定能力（LLM、工具、记忆体）的工作单元。
- Crew 控制器：管理 Crew 生命周期、任务路由与状态机。
三层抽象：
1. Task：用户意图的最小单位，包含输入、输出类型和约束。
2. Crew：由多个 Agent 节点和一段协作拓扑（DAG）组成的执行环境。
3. Skill：可复用的能力封装，一个 Agent 可以挂载多个 Skill。

3. 消息与通信模型

AgentCrew MCN 的通信完全基于 意图消息（Intent Message），这是一个标准化的结构化数据包，包含：

{
  "intent": "write_article",
  "payload": {
    "topic": "AgentCrew MCN 架构设计",
    "style": "技术深度文章"
  },
  "sender": "orchestrator",
  "recipient": "writer_agent",
  "correlation_id": "task-123",
  "metadata": {
    "priority": "high",
    "ttl": 3600
  }
}

消息通过 CrewBus 投递，支持三种通信模式：

点对点（Direct）：明确指定接收 Agent，用于确定性任务链。
广播（Broadcast）：发送给 Crew 内所有 Agent，用于状态同步或通知。
路由匹配（Intent‑based routing）：由 Crew 控制器根据意图字段自动分配到具备对应能力的 Agent，实现灵活调度。

4. Agent 节点设计

每个 Agent 节点本质上是一个微服务，内部包含：

class AgentNode:
    def __init__(self, name: str, skills: List[Skill], memory: Memory):
        self.name = name
        self.skills = {skill.intent: skill for skill in skills}
        self.memory = memory

    async def handle_message(self, msg: IntentMessage):
        # 根据 intent 选择技能
        skill = self.skills.get(msg.intent)
        if not skill:
            return self._reject(msg, reason="unsupported_intent")
        # 利用记忆上下文增强执行
        context = await self.memory.retrieve(msg.correlation_id)
        result = await skill.execute(msg.payload, context)
        # 更新记忆
        await self.memory.store(msg.correlation_id, result)
        return result

Agent 不维护全局拓扑，只对收到的消息做出反应。这种无状态设计（除了记忆体）使得节点可以水平扩展。

5. Crew 控制器与拓扑编排

Crew 控制器是协作的大脑。它通过解析 Crew 定义文件 来构建 DAG（有向无环图），并驱动任务流转。

示例 crew.yaml：

name: article_creation_crew
entry: researcher_agent
agents:
  - name: researcher_agent
    skills: [research]
  - name: writer_agent
    skills: [write_article]
  - name: reviewer_agent
    skills: [review_compliance]

topology:
  - from: entry
    to: researcher_agent
    intent: research
  - from: researcher_agent
    to: writer_agent
    intent: write_article
    condition: "output.status == 'ok'"
  - from: writer_agent
    to: reviewer_agent
    intent: review_compliance
    condition: "output.draft != null"
  - from: reviewer_agent
    to: exit
    condition: "output.approved == true"

控制器引擎的核心逻辑：

class CrewController:
    def __init__(self, topology_def: dict):
        self.topology = self._parse_dag(topology_def)
        self.state = {}

    async def execute(self, task: Task):
        node = self.topology.entry_node
        while node != "exit":
            # 发送消息给对应 Agent
            msg = self._build_intent(task, node)
            response = await bus.dispatch(msg)
            # 记录状态
            self.state[node] = response
            # 根据条件和当前节点找到下一个节点
            node = self.topology.next(node, response)
            task.payload = response  # 数据传递
        return self.state

6. 异常处理与韧性设计

多 Agent 协作中，失败是常态。MCN 内置了多级容错策略：

超时重试：每条消息可配置 ttl 和重试次数，CrewBus 自动处理指数退避。
断路器：当某个 Agent 连续失败达到阈值时，将对其熔断，避免级联崩溃。
补偿任务：拓扑中可定义 on_failure 跳转，执行回滚或通知人工接管。
影子模式：灰度发布新 Agent 时，可同时向新旧版本发送流量并对比结果，不影响主链路。

7. 可观测性：全链路追踪

为了洞察复杂协作，MCN 为每条消息注入了 OpenTelemetry 标准的 Trace Context。开发者在控制台可以看到一次“从研究到文章发布”的完整调用链：

Task #123 (article_creation_crew)
 ├─ researcher_agent: 1.2s (ok)
 ├─ writer_agent: 2.8s (ok)
 │   └─ tool: gpt-4o 调用 1.3s
 └─ reviewer_agent: 0.5s (ok)
     └─ 输出: approved

所有日志、指标和链路由 CrewOps Dashboard 统一展示，支持按业务 ID 检索。

8. 实战示例：搭建一个营销内容流水线

下面用 Python SDK 快速搭建一个简化版的多 Agent 内容生成流水线（基于 MCN）。

from agentcrew import AgentNode, Skill, CrewController, CrewBus
import asyncio

# 1. 定义 Skill
async def research_skill(payload, context):
    # 模拟搜索资料
    return {"findings": "AgentCrew MCN v2 发布，性能提升 40%"}

async def write_skill(payload, context):
    findings = payload.get("findings", "")
    draft = f"根据研究结果撰写：{findings}"
    return {"draft": draft}

async def review_skill(payload, context):
    draft = payload.get("draft", "")
    if "性能" in draft:
        return {"approved": True, "comment": "合规"}
    return {"approved": False, "comment": "缺少数据支撑"}

# 2. 构造 Agent 节点
researcher = AgentNode(name="researcher", skills=[Skill("research", research_skill)])
writer = AgentNode(name="writer", skills=[Skill("write", write_skill)])
reviewer = AgentNode(name="reviewer", skills=[Skill("review", review_skill)])

# 3. 注册到总线并定义拓扑
bus = CrewBus()
bus.register([researcher, writer, reviewer])

topology_def = {
    "entry": "researcher",
    "topology": [
        {"from": "researcher", "to": "writer", "intent": "write"},
        {"from": "writer", "to": "reviewer", "intent": "review"},
        {"from": "reviewer", "to": "exit", "condition": "output.approved == true"}
    ]
}
controller = CrewController(topology_def, bus)

# 4. 执行任务
async def main():
    task = {"topic": "AgentCrew 架构"}
    result = await controller.execute(task)
    print(result)  # 整个流水线状态

asyncio.run(main())

9. 总结与展望

AgentCrew MCN 提供了一套轻量级、可组合的多 Agent 协作架构。它通过意图驱动的消息总线、声明式拓扑编排、以及丰富的异常处理，让开发者从“如何让 Agent 通信”的细节中解放出来，专注于业务 Skill 的开发。

未来，我们将进一步增强 MCN 的自治能力，包括：

Crew 自动优化：根据历史执行轨迹自动合并/拆分节点。
多模态 Agent 支持：不仅限文本，还可接入图像、音视频处理 Agent。
联邦协作：跨团队、跨组织的多 Crew 信任与协作协议。

如果你对多 Agent 协作感兴趣，欢迎前往 GitHub 查看项目源码，点亮 Star，也期待在 Issue 区听到你的场景与需求。让我们一同探索智能体协作的无限可能！

AgentCrew MCN: Your Open-Source AI Marketing Team That Works 24/7

Rick — Tue, 21 Jul 2026 07:00:56 +0000

AgentCrew MCN: Your Open-Source AI Marketing Team That Works 24/7

Let’s be honest—content marketing can be brutal. You need to research topics, craft technical articles, adapt the tone for every platform, schedule posts for maximum engagement, track performance, and then do it all over again. For indie hackers, open-source maintainers, and small dev teams, this often means either burning out or ignoring content altogether.

What if you had a team of AI agents that could handle the entire pipeline—from researching a topic to publishing on Juejin, Dev.to, and Zhihu—while you sleep? That’s exactly what AgentCrew MCN does. It’s an open-source multi-agent content marketing automation tool written in Python, complete with a web dashboard, RAG knowledge base, smart scheduling, and a dogfooding twist: the project promotes itself.

In this article, I’ll walk through the architecture, show you how to get started with a few commands, and demonstrate some advanced features like scheduled publishing and image generation.

What Is AgentCrew MCN?

AgentCrew MCN structures content marketing as a team of specialized AI workers, each with its own role. There are four core agents:

Writer Agent – Your copywriter. It generates technical articles, social posts, and twitter threads, injecting relevant knowledge from a RAG (Retrieval-Augmented Generation) system.
Reviewer Agent – Your quality assurance. It scans content before publication for safety, policy compliance, and basic readability checks.
Publisher Agent – Your operations person. It pushes content to platforms like Juejin, Zhihu, and Dev.to, adapting the format and style to each community.
Analyst Agent – Your data analyst. It tracks content performance, recommends optimal publish times, and learns from past results—right down to adding random jitter to avoid detection of automated scheduling.

The magic isn’t just in the agents; it’s in the Skills + Tools system. Each agent has high-level skills (like “generate a technical thread” or “batch publish”) composed of low-level tools (like search, rag, compose, devto_api). Since v0.4, the system is LLM-driven, meaning the agents can reason about which tool to use and in which order—no rigid workflows.

Quick Start: From Zero to Published in 5 Minutes

AgentCrew is a Python package available on PyPI. You’ll need Python 3.10+ and an API key for an LLM provider. It supports DeepSeek, OpenAI, Anthropic, and Ollama—switch with one config field.

# 1. Install
pip install agentcrew-mcn

# 2. Initialize configuration
agentcrew-mcn init

# 3. Edit the generated .env file with your LLM API key
#    DEEPSEEK_API_KEY=sk-...

That’s it. Now let’s generate a technical article:

agentcrew-mcn write generate \
  --topic "Python async programming" \
  --style technical

The Writer Agent will research the topic (optionally using the RAG database from your previous articles), compose a structured piece, and save it as Markdown. Before publishing, you can run a dry run to see what will happen:

agentcrew-mcn publish post \
  --file article.md \
  --platform devto \
  --dry-run

If everything looks good, remove the --dry-run flag to actually post it to your Dev.to account.

Configuration

The init command creates a config.yaml. A minimal setup looks like this:

llm:
  provider: deepseek
  api_key: ${DEEPSEEK_API_KEY}   # reads from .env
  model: deepseek-chat

platforms:
  devto:
    api_key: ${DEVTO_API_KEY}     # get it from https://dev.to/settings/extensions
  juejin:
    cookie: ${JUEJIN_COOKIE}      # export from browser

The ${VARIABLE} syntax pulls values from your .env file, keeping secrets out of your config.

The Agent Orchestra: How It All Works Under the Hood

At the center sits the Orchestrator, which is responsible for task dispatch, scheduling, and configuration. When you run agentcrew-mcn write generate, the Orchestrator hands the task to the Writer Agent, injecting the appropriate skills. The Writer might use a search tool to pull recent content from the web, a rag tool to retrieve relevant paragraphs from your knowledge base, and a compose tool to generate the final article.

The Reviewer Agent then checks the output using a series of validation rules (configurable) and can trigger a rewrite automatically if something is off.

The Publisher Agent converts the article into the platform-specific format. For example, Juejin articles use a “Golden Nugget” title convention; Dev.to uses front matter. The agent knows how to handle these differences.

Finally, the Analyst Agent runs scheduling logic. Instead of posting at a fixed time, it uses a recommendation model (powered by historical data) and adds random jitter to mimic human behavior, reducing the chance of being flagged as automation.

Advanced Features You’ll Love

1. RAG Knowledge Base

The built‑in vector store (ChromaDB) lets you ingest past articles, documentation, or any text source, so the Writer Agent can produce content that’s consistent with your tone and style.

# Ingest a previous blog post
agentcrew-mcn rag ingest \
  --file my_old_article.md \
  --source "personal_blog"

# Search the knowledge base later
agentcrew-mcn rag search \
  --query "AI Agent architecture"

The retrieved context is automatically prepended to the prompt when generating new content.

2. Scheduled Publishing

Want to keep your social media active without constant manual effort? Use the scheduler:

agentcrew-mcn schedule start \
  --topic-file topics.txt \
  --platform juejin \
  --interval 6

This will pick a topic from the file, generate and publish an article, and repeat every 6 hours. The Analyst Agent chooses the exact moment within the interval, so posting doesn’t look like an hourly cron job.

3. Multi-Provider LLM and Image Generation

Not locked into one AI provider. Set provider: openai in your config to switch to GPT-4. For cover images, enable DALL‑E 3:

image_gen:
  enabled: true
  model: dall-e-3

The Writer Agent will automatically generate a cover image for each article.

4. Dashboard and Growth Tracking

A Streamlit dashboard gives you real‑time analytics: publish history, engagement metrics, and AI-generated suggestions for next topics.

agentcrew-mcn dashboard

Additionally, a GROWTH.md file logs follower counts, views, and other KPIs over time, so you can measure progress.

Dogfooding: The Project Promotes Itself

A notable aspect of AgentCrew is dogfooding—the project uses itself to write and publish content about itself. The marketing of AgentCrew is largely automated by AgentCrew, which serves as both a proof‑of‑concept and a source of continuous improvement.

Roadmap and Community

AgentCrew is MIT‑licensed and under active development. The current version (v0.4) has 392 tests and 85% coverage. Here’s what’s coming:

v0.5 – Support for 6 new platforms: CSDN, WeChat Official Accounts, SegmentFault, X (Twitter), Xiaohongshu, and Medium.
v1.0 – REST API, plugin system, and an official community.

If you’re a Python developer, content creator, or just curious about multi‑agent systems, there are plenty of ways to contribute. Check out the GitHub repository and the documentation.

Getting started is ridiculously simple: pip install agentcrew-mcn && agentcrew-mcn init. From there, you’re five lines away from having an AI marketing team that never sleeps, never asks for a raise, and keeps your OSS project or personal brand visible across the web—without burning you out.

What would you do with a 24/7 AI writing crew? I’d love to hear your ideas in the comments. And if this project sounds useful, a ⭐ on GitHub goes a long way!

python #ai #automation #opensource #devto #marketing #multiagent #llm

AgentCrew MCN: Your Open-Source AI Marketing Team That Works 24/7

Rick — Tue, 21 Jul 2026 06:23:49 +0000

AgentCrew MCN: Your Open-Source AI Marketing Team That Works 24/7

In this article, I’ll walk through the architecture, show you how to get started with a few commands, and demonstrate some advanced features like scheduled publishing and image generation.

What Is AgentCrew MCN?

AgentCrew MCN structures content marketing as a team of specialized AI workers, each with its own role. There are four core agents:

Writer Agent – Your copywriter. It generates technical articles, social posts, and twitter threads, injecting relevant knowledge from a RAG (Retrieval-Augmented Generation) system.
Reviewer Agent – Your quality assurance. It scans content before publication for safety, policy compliance, and basic readability checks.
Publisher Agent – Your operations person. It pushes content to platforms like Juejin, Zhihu, and Dev.to, adapting the format and style to each community.
Analyst Agent – Your data analyst. It tracks content performance, recommends optimal publish times, and learns from past results—right down to adding random jitter to avoid detection of automated scheduling.

Quick Start: From Zero to Published in 5 Minutes

AgentCrew is a Python package available on PyPI. You’ll need Python 3.10+ and an API key for an LLM provider. It supports DeepSeek, OpenAI, Anthropic, and Ollama—switch with one config field.

# 1. Install
pip install agentcrew-mcn

# 2. Initialize configuration
agentcrew-mcn init

# 3. Edit the generated .env file with your LLM API key
#    DEEPSEEK_API_KEY=sk-...

That’s it. Now let’s generate a technical article:

agentcrew-mcn write generate \
  --topic "Python async programming" \
  --style technical

agentcrew-mcn publish post \
  --file article.md \
  --platform devto \
  --dry-run

If everything looks good, remove the --dry-run flag to actually post it to your Dev.to account.

Configuration

The init command creates a config.yaml. A minimal setup looks like this:

llm:
  provider: deepseek
  api_key: ${DEEPSEEK_API_KEY}   # reads from .env
  model: deepseek-chat

platforms:
  devto:
    api_key: ${DEVTO_API_KEY}     # get it from https://dev.to/settings/extensions
  juejin:
    cookie: ${JUEJIN_COOKIE}      # export from browser

The ${VARIABLE} syntax pulls values from your .env file, keeping secrets out of your config.

The Agent Orchestra: How It All Works Under the Hood

The Reviewer Agent then checks the output using a series of validation rules (configurable) and can trigger a rewrite automatically if something is off.

Advanced Features You’ll Love

1. RAG Knowledge Base

The built‑in vector store (ChromaDB) lets you ingest past articles, documentation, or any text source, so the Writer Agent can produce content that’s consistent with your tone and style.

# Ingest a previous blog post
agentcrew-mcn rag ingest \
  --file my_old_article.md \
  --source "personal_blog"

# Search the knowledge base later
agentcrew-mcn rag search \
  --query "AI Agent architecture"

The retrieved context is automatically prepended to the prompt when generating new content.

2. Scheduled Publishing

Want to keep your social media active without constant manual effort? Use the scheduler:

agentcrew-mcn schedule start \
  --topic-file topics.txt \
  --platform juejin \
  --interval 6

3. Multi-Provider LLM and Image Generation

Not locked into one AI provider. Set provider: openai in your config to switch to GPT-4. For cover images, enable DALL‑E 3:

image_gen:
  enabled: true
  model: dall-e-3

The Writer Agent will automatically generate a cover image for each article.

4. Dashboard and Growth Tracking

A Streamlit dashboard gives you real‑time analytics: publish history, engagement metrics, and AI-generated suggestions for next topics.

agentcrew-mcn dashboard

Additionally, a GROWTH.md file logs follower counts, views, and other KPIs over time, so you can measure progress.

Dogfooding: The Project Promotes Itself

Roadmap and Community

AgentCrew is MIT‑licensed and under active development. The current version (v0.4) has 392 tests and 85% coverage. Here’s what’s coming:

v0.5 – Support for 6 new platforms: CSDN, WeChat Official Accounts, SegmentFault, X (Twitter), Xiaohongshu, and Medium.
v1.0 – REST API, plugin system, and an official community.

If you’re a Python developer, content creator, or just curious about multi‑agent systems, there are plenty of ways to contribute. Check out the GitHub repository and the documentation.

What would you do with a 24/7 AI writing crew? I’d love to hear your ideas in the comments. And if this project sounds useful, a ⭐ on GitHub goes a long way!

python #ai #automation #opensource #devto #marketing #multiagent #llm

AgentCrew Quick Test

Rick — Fri, 17 Jul 2026 09:39:16 +0000

Hello Dev.to from AgentCrew 🚀