Multi-Agent Dev Team
2-agent collaborative software development workflow for OpenClaw
Build complete software projects using AI agents that work together like a real development team.
What is this?
The Multi-Agent Dev Team skill provides a PM (Project Manager) and Dev (Developer) agent that collaborate to build software projects. You describe what you want, the PM breaks it down into tasks, and the Dev agent implements them.
Perfect for:
- - Landing pages and websites
- Small web applications
- Prototypes and MVPs
- Code generation projects
- Learning multi-agent workflows
Quick Start
1. Install the skill
CODEBLOCK0
2. Set up agents
The skill includes two pre-configured agents:
PM Agent (multi-agent-pm)
- - Handles project planning
- Coordinates the Dev agent
- Reports progress to you
Dev Agent (multi-agent-dev)
- - Implements code
- Tests functionality
- Commits to Git
3. Configure in ~/.openclaw/config.yaml
CODEBLOCK1
4. Start your PM agent
CODEBLOCK2
Usage Example
Building a Landing Page
You (Director):
Build me a Next.js landing page with hero section, features, and contact form. Use Tailwind CSS and TypeScript.
PM Agent:
I'll break this down into a clear task for the Dev agent. Starting now...
[Spawns Dev agent with detailed spec]
Dev agent is working on the project. I'll monitor progress and report back when complete.
PM Agent (later):
✅ Landing page completed!
Deliverables:
- - Next.js 14+ App Router project
- Hero, Features, Contact sections
- Fully responsive design
- TypeScript + Tailwind CSS
- Git repository initialized
Location: INLINECODE3
Ready for deployment. Would you like me to deploy to Vercel?
How It Works
Architecture
CODEBLOCK3
Workflow
- 1. You describe your project to the PM agent
- PM creates a structured task specification
- PM spawns Dev agent with the spec
- Dev implements the code
- PM reviews the deliverables
- PM reports completion (or requests revisions)
- Repeat steps 4-6 if needed (max 3 iterations)
Task Specification Format
The PM agent uses this template to communicate with Dev:
CODEBLOCK4
Supported Project Types
✅ Works Great
- - Next.js landing pages & apps
- React components & SPAs
- Node.js scripts & APIs
- TypeScript projects
- Static sites (HTML/CSS/JS)
- Documentation sites
⚠️ Limited Support
- - Complex backend systems (use Pro version)
- Real-time applications
- Multi-service architectures
- Mobile apps
❌ Not Recommended
- - Large enterprise systems
- Mission-critical production code without human review
- Projects requiring specialized agents (use Pro version)
Configuration Options
Agent Models
PM Agent (orchestration):
- -
anthropic/claude-sonnet-4-5 - Best reasoning - INLINECODE5 - Fast & efficient
- INLINECODE6 - Balanced
Dev Agent (code generation):
- -
google/gemini-2.5-flash - Fast iteration (recommended) - INLINECODE8 - Higher quality
- INLINECODE9 - Budget-friendly
Workspace Configuration
Set a dedicated workspace for projects:
CODEBLOCK5
Best Practices
1. Start Small
Don't ask for everything at once. Start with an MVP:
❌ Bad:
Build a full e-commerce site with user auth, payments, admin dashboard, and inventory management.
✅ Good:
Build a simple product landing page with hero, features, and signup form.
2. Be Specific
The more specific your requirements, the better the result:
❌ Vague:
Make a nice website.
✅ Specific:
Create a Next.js landing page with:
- - Hero section with CTA button
- 3-column feature grid
- Contact form with email validation
- Tailwind CSS styling
- Dark mode support
3. Iterate Incrementally
Build in phases:
Phase 1: Basic structure
Phase 2: Add features
Phase 3: Polish & deploy
4. Review Output
Always review the generated code before deploying. The agents are good, but human oversight is important.
5. Provide Examples
If you have a specific style or pattern in mind, share examples:
Build a landing page similar to https://example.com, but for [your product].
Troubleshooting
"Dev agent didn't complete the task"
Check:
- 1. Was the task specification clear?
- Are required tools available (Node.js, Git)?
- Did the agent hit resource limits?
Solution:
- - Simplify the task
- Check PM agent logs via INLINECODE10
- Try again with clearer requirements
"Code doesn't work"
Check:
- 1. Dependencies installed? (
npm install) - Environment variables set?
- Correct Node.js version?
Solution:
- - Ask PM agent: "The code has errors. Please review and fix."
- The PM will spawn Dev again for corrections
"Task took too long"
Solutions:
- - Break into smaller tasks
- Use faster model for Dev agent (
gemini-2.5-flash) - Simplify requirements
Examples
Example 1: Simple Landing Page
CODEBLOCK6
Example 2: React Component Library
CODEBLOCK7
Example 3: API Integration
CODEBLOCK8
Upgrading to Pro
Want more power? Upgrade to Multi-Agent Dev Team Pro ($49):
Pro Features
- - 🎯 6 specialized agents: PM, Architect, Dev, QA, DevOps, BizDev
- 🔄 Lobster pipelines: Automated workflows with approval gates
- 🏗️ Architecture design: Dedicated agent for system design
- ✅ Automated QA: Code review & testing agent
- 🚀 DevOps automation: Deployment & CI/CD setup
- 💼 Business planning: Market research & strategy agent
- 📚 Comprehensive guides: English + Korean setup docs
Get Pro →
Support
License
MIT License - See LICENSE file for details
Credits
Built by UBIK Collective
Powered by OpenClaw
Ready to build with AI agents? Install now:
CODEBLOCK9
多智能体开发团队
面向OpenClaw的双智能体协作软件开发工作流
使用像真实开发团队一样协作的AI智能体构建完整的软件项目。
这是什么?
多智能体开发团队技能提供了一个PM(项目经理)和Dev(开发者)智能体,它们协作构建软件项目。你描述需求,PM将其分解为任务,Dev智能体负责实现。
完美适用于:
- - 落地页和网站
- 小型Web应用
- 原型和MVP
- 代码生成项目
- 学习多智能体工作流
快速开始
1. 安装技能
bash
npx clawhub install multi-agent-dev-team
2. 设置智能体
该技能包含两个预配置的智能体:
PM智能体(multi-agent-pm)
Dev智能体(multi-agent-dev)
3. 在~/.openclaw/config.yaml中配置
yaml
agents:
multi-agent-pm:
soul: ~/.openclaw/skills/multi-agent-dev-team/agents/pm-agent/SOUL.md
model: anthropic/claude-sonnet-4-5-20250929 # 或你偏好的模型
multi-agent-dev:
soul: ~/.openclaw/skills/multi-agent-dev-team/agents/dev-agent/SOUL.md
model: google/gemini-2.5-flash # 代码生成更快
4. 启动你的PM智能体
bash
openclaw chat --agent multi-agent-pm
使用示例
构建落地页
你(指导者):
为我构建一个包含英雄区、功能区和联系表单的Next.js落地页。使用Tailwind CSS和TypeScript。
PM智能体:
我将为Dev智能体将其分解为清晰的任务。现在开始...
[用详细规格生成Dev智能体]
Dev智能体正在处理项目。我将监控进度并在完成后汇报。
PM智能体(稍后):
✅ 落地页已完成!
交付物:
- - Next.js 14+ App Router项目
- 英雄区、功能区和联系区
- 完全响应式设计
- TypeScript + Tailwind CSS
- Git仓库已初始化
位置: ~/projects/my-landing-page/
已准备好部署。需要我部署到Vercel吗?
工作原理
架构
你(指导者)
↓
PM智能体(编排者)
↓
Dev智能体(实现者)
↓
可运行代码
工作流
- 1. 你向PM智能体描述你的项目
- PM创建结构化任务规格
- PM用规格生成Dev智能体
- Dev实现代码
- PM审查交付物
- PM报告完成(或请求修改)
- 如果需要,重复步骤4-6(最多3次迭代)
任务规格格式
PM智能体使用此模板与Dev通信:
markdown
项目:[名称]
任务:[操作]
需求:
- 1. [需求1]
- [需求2]
技术约束:
验收标准:
交付物:
支持的项目类型
✅ 效果良好
- - Next.js落地页和应用
- React组件和SPA
- Node.js脚本和API
- TypeScript项目
- 静态网站(HTML/CSS/JS)
- 文档网站
⚠️ 有限支持
- - 复杂后端系统(使用专业版)
- 实时应用
- 多服务架构
- 移动应用
❌ 不推荐
- - 大型企业系统
- 未经人工审查的关键生产代码
- 需要专业智能体的项目(使用专业版)
配置选项
智能体模型
PM智能体(编排):
- - anthropic/claude-sonnet-4-5 - 最佳推理
- google/gemini-2.5-flash - 快速高效
- openai/gpt-4o - 均衡
Dev智能体(代码生成):
- - google/gemini-2.5-flash - 快速迭代(推荐)
- anthropic/claude-sonnet-4-5 - 更高质量
- openai/gpt-4o-mini - 经济实惠
工作区配置
为项目设置专用工作区:
yaml
agents:
multi-agent-pm:
soul: ~/.openclaw/skills/multi-agent-dev-team/agents/pm-agent/SOUL.md
model: anthropic/claude-sonnet-4-5-20250929
cwd: ~/dev-projects # 所有项目在此创建
最佳实践
1. 从小处着手
不要一次性要求所有功能。从MVP开始:
❌ 不好:
构建一个包含用户认证、支付、管理后台和库存管理的完整电商网站。
✅ 好:
构建一个包含英雄区、功能区和注册表单的简单产品落地页。
2. 具体明确
需求越具体,结果越好:
❌ 模糊:
做一个好看的网站。
✅ 具体:
创建一个Next.js落地页,包含:
- - 带CTA按钮的英雄区
- 3列功能网格
- 带邮箱验证的联系表单
- Tailwind CSS样式
- 暗色模式支持
3. 增量迭代
分阶段构建:
阶段1: 基础结构
阶段2: 添加功能
阶段3: 打磨与部署
4. 审查输出
在部署前始终审查生成的代码。智能体很优秀,但人工监督很重要。
5. 提供示例
如果你有特定的风格或模式,分享示例:
构建一个类似https://example.com的落地页,但用于[你的产品]。
故障排除
Dev智能体未完成任务
检查:
- 1. 任务规格是否清晰?
- 所需工具是否可用(Node.js、Git)?
- 智能体是否达到资源限制?
解决方案:
- - 简化任务
- 通过sessions_history检查PM智能体日志
- 用更清晰的需求重试
代码无法运行
检查:
- 1. 依赖是否已安装?(npm install)
- 环境变量是否已设置?
- Node.js版本是否正确?
解决方案:
- - 告诉PM智能体:代码有错误。请审查并修复。
- PM将再次生成Dev进行修正
任务耗时太长
解决方案:
- - 分解为更小的任务
- 对Dev智能体使用更快的模型(gemini-2.5-flash)
- 简化需求
示例
示例1:简单落地页
你:为名为TaskFlow的SaaS产品构建落地页。
包含英雄区、功能(3张卡片)和定价表。使用Next.js
和Tailwind CSS。
PM:正在处理...
[2分钟后]
PM:✅ TaskFlow落地页已完成!已准备好部署。
示例2:React组件库
你:创建一个可复用的Button组件库,包含变体
(primary、secondary、outline)和尺寸(sm、md、lg)。使用
TypeScript和class-variance-authority。
PM:任务已收到。正在生成Dev智能体...
[3分钟后]
PM:✅ Button组件库已完成,附带Storybook示例。
示例3:API集成
你:构建一个获取并显示GitHub用户资料的Next.js应用。
包含搜索功能和响应式卡片。
PM:开始开发...
[4分钟后]
PM:✅ GitHub资料查看器已完成,包含搜索和错误处理。
升级到专业版
想要更多功能?升级到多智能体开发团队专业版($49):
专业版功能
- - 🎯 6个专业智能体:PM、架构师、Dev、QA、DevOps、BizDev
- 🔄 Lobster流水线:带审批门的自动化工作流
- 🏗️ 架构设计:专用于系统设计的智能体
- ✅ 自动化QA:代码审查和测试智能体
- 🚀 DevOps自动化:部署和CI/CD设置
- 💼 业务规划:市场研究和策略智能体
- 📚 综合指南:英文+韩文设置文档
[获取专业版 →