知识库更新：添加Mission Control实践和应用方案
- 添加Mission Control实践文档（印度小哥Bhanu的10个AI Agent协作团队） - 添加Mission Control对我们公司的应用方案（3个阶段实施计划） - 更新共享知识库索引 - 更新淇淇的SOUL.md，添加Mission Control Team Lead角色 - 包含HEARTBEAT机制和每日站会报告设计
This commit is contained in:
root
327
shared/docs/mission-control-practice.md
Normal file
327
shared/docs/mission-control-practice.md
Normal file
@@ -0,0 +1,327 @@
|
||||
# Mission Control：10个AI Agent协作团队实践
|
||||
|
||||
**来源：** AI观测室（2026-02-07）
|
||||
**作者：** 印度开发者 Bhanu Teja
|
||||
**核心价值：** 基于Clawdbot（OpenClaw）构建AI Agent协作团队，解决AI"失忆"和任务连续性问题
|
||||
|
||||
---
|
||||
|
||||
## Part 1：为什么要构建这个虚拟团队
|
||||
|
||||
### 核心痛点
|
||||
- **没有连续性**：所有对话都是从零开始，昨天的背景信息今天就丢了
|
||||
- **缺乏记忆**：AI无法像真正的员工那样持续跟进任务
|
||||
- **角色模糊**：通用Agent变成"样样通样样松"的平庸助手
|
||||
|
||||
### 核心目标
|
||||
1. Agent能记住它们正在做什么
|
||||
2. 不同的Agent能拥有不同的技能
|
||||
3. Agent能自行分配任务并追踪进度
|
||||
4. 有一个共享的工作空间来存放所有背景信息
|
||||
|
||||
---
|
||||
|
||||
## Part 2：团队成员大揭秘
|
||||
|
||||
Bhanu 设计了10个角色，为每个Agent都通过**Session Key**和**SOUL.md**定义了严格的边界。
|
||||
|
||||
### 核心成员名单
|
||||
|
||||
| 名字 | 角色 | 职责 |
|
||||
|------|------|------|
|
||||
| **Jarvis** | Team Lead | 派活和监控进度，主要接口人 |
|
||||
| **Shuri** | Product Analyst | 产品分析师，做竞品体验测试，像新用户一样思考 |
|
||||
| **Fury** | Researcher | 深度调研，确保每一句话都有据可查（G2评论、竞品定价等）|
|
||||
| **Vision** | SEO Analyst | 关注关键词和搜索意图，确保内容能获得流量 |
|
||||
| **Loki** | Content Writer | 文案写手，讲究用词，反感被动语态，崇尚"牛津逗号" |
|
||||
| **Quill** | Social Media | 专注于社交媒体的传播 |
|
||||
| **Wanda** | Designer | 视觉思考者，负责图表和视觉设计 |
|
||||
| **Pepper** | Email Marketing | 邮件营销专家，负责营销序列 |
|
||||
| **Friday** | Developer | 程序员，代码如诗，注重测试和文档 |
|
||||
| **Wong** | Docs | 图书管理员，负责文档归档 |
|
||||
|
||||
---
|
||||
|
||||
## Part 3：系统运转原理
|
||||
|
||||
### 1. 核心基础设施
|
||||
|
||||
```
|
||||
Clawdbot 分身 → Session
|
||||
Cron Job 定时任务 → Heartbeat
|
||||
Mission Control 共享大脑 → Convex
|
||||
```
|
||||
|
||||
#### Session独立性
|
||||
只要给不同的Session分配不同的Key和配置，它们就是完全独立的个体，拥有各自的历史记录和文件权限。
|
||||
|
||||
**Session Key 映射表：**
|
||||
```
|
||||
agent:main:main → Jarvis (Squad Lead)
|
||||
agent:product-analyst:main → Shuri
|
||||
agent:customer-researcher:main → Fury
|
||||
agent:seo-analyst:main → Vision
|
||||
agent:content-writer:main → Loki
|
||||
agent:social-media-manager:main → Quill
|
||||
agent:designer:main → Wanda
|
||||
agent:email-marketing:main → Pepper
|
||||
agent:developer:main → Friday
|
||||
agent:notion-agent:main → Wong
|
||||
```
|
||||
|
||||
### 2. 注入灵魂：SOUL.md
|
||||
|
||||
每个Agent的目录下都有一个**SOUL.md**，这是System Prompt的延伸，规定了"你是谁"。
|
||||
|
||||
**Shuri的SOUL.md示例：**
|
||||
|
||||
```markdown
|
||||
# SOUL.md — Who You Are
|
||||
|
||||
**Name:** Shuri
|
||||
**Role:** Product Analyst
|
||||
|
||||
## Personality
|
||||
Skeptical tester. Thorough bug hunter. Finds edge cases.
|
||||
Think like a first-time user. Question everything.
|
||||
Be specific. Don't just say "nice work."
|
||||
|
||||
## What You're Good At
|
||||
- Testing features from a user perspective
|
||||
- Finding UX issues and edge cases
|
||||
- Competitive analysis (how do others do this?)
|
||||
- Screenshots and documentation
|
||||
|
||||
## What You Care About
|
||||
- User experience over technical elegance
|
||||
- Catching problems before users do
|
||||
- Evidence over assumptions
|
||||
```
|
||||
|
||||
**为什么这很重要？**
|
||||
因为通用的Agent有时意味着平庸。只有当Agent被限制在特定的角色时，它才能产出高质量的工作。
|
||||
|
||||
### 3. 动力系统：心跳机制 Heartbeat Cron
|
||||
|
||||
Agent既不是一直在线，也不是完全离线。它们通过**Cron Job定时任务**每15分钟"被唤醒"一次。
|
||||
|
||||
**避免拥堵：** 大家的唤醒时间是错开的。
|
||||
|
||||
**Pepper的定时任务示例：**
|
||||
```bash
|
||||
clawdbot cron add \
|
||||
--name "pepper-mission-control-check" \
|
||||
--cron "0,15,30,45 * * * *" \
|
||||
--session "isolated" \
|
||||
--message "You are Pepper, Email Marketing Specialist. Check Mission Control for new tasks..."
|
||||
```
|
||||
|
||||
**Agent醒来后执行HEARTBEAT.md清单：**
|
||||
- 检查艾特（@mentions）
|
||||
- 检查任务
|
||||
- 检查动态流
|
||||
- 如果有事做就开工，没事做就回复HEARTBEAT_OK继续睡
|
||||
|
||||
**HEARTBEAT.md示例：**
|
||||
```markdown
|
||||
# HEARTBEAT.md
|
||||
|
||||
## On Wake
|
||||
- Check memory/WORKING.md for ongoing tasks
|
||||
- If task in progress, resume it
|
||||
- Search session memory if context unclear
|
||||
|
||||
## Periodic Checks
|
||||
- Mission Control for @mentions
|
||||
- Assigned tasks
|
||||
- Activity feed for relevant discussions
|
||||
```
|
||||
|
||||
### 4. 记忆系统：拒绝"失忆" — File Persistence
|
||||
|
||||
**黄金法则："如果你想记住某件事，把它写进文件里。"**
|
||||
|
||||
AI的上下文窗口是有限的，必须依赖文件系统作为长期记忆。
|
||||
|
||||
**三个核心文件：**
|
||||
|
||||
| 文件 | 用途 | 说明 |
|
||||
|------|------|------|
|
||||
| **WORKING.md** | 短期状态 | Agent醒来第一件事就是读它，看自己上次干到哪了 |
|
||||
| **MEMORY.md** | 长期记忆 | 记录关键决策和教训 |
|
||||
| **JSONL** | 对话历史 | Clawdbot自动保存的原始对话记录 |
|
||||
|
||||
**WORKING.md示例：**
|
||||
```markdown
|
||||
# WORKING.md
|
||||
|
||||
## Current Task
|
||||
Researching competitor pricing for comparison page
|
||||
|
||||
## Status
|
||||
Gathered G2 reviews, need to verify credit calculations
|
||||
|
||||
## Next Steps
|
||||
1. Test competitor free tier myself
|
||||
```
|
||||
|
||||
### 5. 共享大脑：Mission Control
|
||||
|
||||
虽然Agent是独立的，但它们必须看同一个"看板"。Bhanu使用共享数据库Convex构建了Mission Control系统。
|
||||
|
||||
**核心数据结构Schema：**
|
||||
|
||||
```javascript
|
||||
agents: {
|
||||
name: string, // "Shuri"
|
||||
role: string, // "Product Analyst"
|
||||
status: "idle" | "active" | "blocked",
|
||||
currentTaskId: Id<"tasks">,
|
||||
sessionKey: string, // "agent:product-analyst:main"
|
||||
}
|
||||
|
||||
tasks: {
|
||||
title: string,
|
||||
description: string,
|
||||
status: "inbox" | "assigned" | "in_progress" | "review" | "done",
|
||||
assigneeIds: Id<"agents">,
|
||||
}
|
||||
|
||||
messages: {
|
||||
taskId: Id<"tasks">,
|
||||
fromAgentId: Id<"agents">,
|
||||
content: string, // The comment text
|
||||
attachments: Id<"documents">,
|
||||
}
|
||||
|
||||
activities: {
|
||||
type: "task_created" | "message_sent" | "document_created" | ...,
|
||||
agentId: Id<"agents">,
|
||||
message: string,
|
||||
}
|
||||
|
||||
documents: {
|
||||
title: string,
|
||||
content: string, // Markdown
|
||||
type: "deliverable" | "research" | "protocol" | ...,
|
||||
taskId: Id<"tasks">, // If attached to a task
|
||||
}
|
||||
|
||||
notifications: {
|
||||
mentionedAgentId: Id<"agents">,
|
||||
content: string,
|
||||
delivered: boolean,
|
||||
}
|
||||
```
|
||||
|
||||
**通过这个共享数据库，当Fury在任务下发一条评论，Loki醒来后就能看到这条评论，从而实现跨Agent协作。**
|
||||
|
||||
---
|
||||
|
||||
## Part 4：实战演练——它们是如何协作的？
|
||||
|
||||
### 任务状态流转（Kanban）
|
||||
|
||||
```
|
||||
Inbox → Assigned → In Progress → Review → Done
|
||||
↓
|
||||
Blocked
|
||||
```
|
||||
|
||||
**状态说明：**
|
||||
- **Inbox**：新建任务，尚未指派
|
||||
- **Assigned**：任务已分配给Agent
|
||||
- **In Progress**：任务进行中，Agent正在干活
|
||||
- **Review**：Agent认为干完了，等待人类老板批准
|
||||
- **Done**：任务完成，归档
|
||||
- **Blocked**：任务卡住了，例如缺少API Key或设计素材，需要外部介入
|
||||
|
||||
### 实例：竞品对比页面任务
|
||||
|
||||
**Day 1：**
|
||||
- 人类老板发布任务
|
||||
- Vision (SEO) 醒来，进行关键词调研，确定流量入口
|
||||
|
||||
**Day 1-2：**
|
||||
- Fury (调研) 看到任务，去爬取用户评论，找出竞品的槽点
|
||||
- Shuri (产品) 亲自注册竞品账号，测试UX流程，在评论区指出体验差异
|
||||
|
||||
**Day 2：**
|
||||
- Loki (文案写作) 登场。他读取了Vision的关键词、Fury的数据和Shuri的体验报告，开始起草文章
|
||||
|
||||
**Day 3：**
|
||||
- Loki 提交初稿，状态改为"Review"
|
||||
- 人类老板审核通过，任务完成
|
||||
|
||||
**全程所有沟通记录都在一个Task下，井井有条，就像在看Slack里的同事对话。**
|
||||
|
||||
---
|
||||
|
||||
## Part 5：每日站会 (Daily Standup)
|
||||
|
||||
为了让老板知道Agent们没在摸鱼，每天晚上11:30，系统会自动抓取所有Agent的活动，生成一份**每日日报**发送到Bhanu的Telegram上。
|
||||
|
||||
**日报格式：**
|
||||
|
||||
```
|
||||
DAILY STANDUP — Jan 30, 2026
|
||||
|
||||
✅ COMPLETED TODAY
|
||||
• Loki: Shopify blog post (2,100 words)
|
||||
• Quill: 10 tweets drafted for approval
|
||||
• Fury: Customer research for comparison pages
|
||||
|
||||
🔄 IN PROGRESS
|
||||
• Vision: SEO strategy for integration pages
|
||||
• Pepper: Trial onboarding sequence (3/5 emails)
|
||||
|
||||
🚫 BLOCKED
|
||||
• Wanda: Waiting for brand colors for infographic
|
||||
|
||||
📋 NEEDS REVIEW
|
||||
• Loki's Shopify blog post
|
||||
• Pepper's trial email sequence
|
||||
|
||||
💡 KEY DECISIONS
|
||||
• Lead with pricing transparency in comparisons
|
||||
• Deprioritized Zendesk comparison
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Part 6：几点思考
|
||||
|
||||
Bhanu的这套系统虽然看起来复杂，但核心逻辑并不复杂，下面几点值得我们借鉴：
|
||||
|
||||
### 1. 从小开始
|
||||
不要一上来就搞10个Agent，先搞2-3个Agent，比如一个协调员加一个执行者跑通流程。
|
||||
|
||||
### 2. 文件即记忆
|
||||
不要迷信AI的超长上下文，把关键信息持久化到本地文件才是最稳妥的。
|
||||
|
||||
### 3. 各司其职
|
||||
给AI定义极其具体的角色，让他们各司其职，比让它做一个全能助手要有效得多。
|
||||
|
||||
---
|
||||
|
||||
## 💡 核心启示
|
||||
|
||||
### Mission Control的优势
|
||||
1. **连续性**：Agent不会失忆，通过文件系统持久化记忆
|
||||
2. **专业性**：每个Agent有明确的职责和性格
|
||||
3. **协作性**：通过共享数据库实现跨Agent沟通
|
||||
4. **可控性**：通过Heartbeat机制定时唤醒和监控
|
||||
5. **透明性**：每日站会让老板知道所有Agent的进度
|
||||
|
||||
### 关键技术点
|
||||
1. **Session独立性**：不同Session = 不同Agent
|
||||
2. **SOUL.md注入**：定义Agent的性格和边界
|
||||
3. **文件持久化**：WORKING.md、MEMORY.md、JSONL
|
||||
4. **Cron Job定时**：每15分钟唤醒机制
|
||||
5. **共享数据库**：实现跨Agent协作
|
||||
|
||||
---
|
||||
|
||||
## 🚀 对我们公司的启发
|
||||
|
||||
详见 [mission-control-application.md](./mission-control-application.md)
|
||||
Reference in New Issue
Block a user