Calling-Agent-Squad：多 Agent 協作框架

💡 解決的問題

單一 AI Agent 有能力限制。當任務太過複雜時，單一 Agent 往往會：

遺漏重要細節
在分析的深度和廣度之間失衡
缺乏獨立的品質把關機制

多個 Agent 協作可以突破這些限制。每個 Agent 專注自己的領域，透過結構化的協作流程，完成單一 Agent 無法勝任的複雜任務。

🏗️ 架構設計

框架以Architect → Coder → Reviewer 為核心流程，形成一個完整的「規劃→執行→審查」循環。

Architect Agent（架構師）

角色：分析需求、規劃架構、制定技術路線

擅長：系統性思考、技術全景、風險識別

輸出：技術規格文檔、架構圖、實作計劃

Coder Agent（工程師）

角色：負責具體實作、編寫代碼、解決技術問題

擅長：代碼生成、演算法實現、細節執行

輸出：可運行的代碼、單元測試、技術文檔

Reviewer Agent（審查員）

角色：審查代碼或方案、把關品質、提出改進建議

擅長：邏輯分析、漏洞識別、程式碼優化

輸出：審查報告、改進建議、品質評估

⚡ 工作流程

User/LLM Router
     ↓
Architect Agent (分析需求)
     ↓
生成技術規格
     ↓
Coder Agent (執行實作)
     ↓
生成代碼
     ↓
Reviewer Agent (品質把關)
     ↓
[如果需要修訂] → Coder Agent (修訂)
     ↓
[通過審查] → 輸出最終結果

🔄 與傳統 Single Agent 的差異

維度	Single Agent	Multi-Agent (Squad)
處理複雜度	有限，容易超載	可以分工處理複雜任務
專業分工	單一角色，什麼都做但都不精	每個 Agent 有專門角色和專長
審查機制	自我檢查，容易有盲點	獨立的 Reviewer 把關
產出品質	取決於 Agent 能力上限	透過協作提升整體品質
可擴展性	受限於單一 Agent	可根據需求新增 Agent 角色

💼 適用場景

大型專案架構規劃 — 需要多方考量的大型系統
複雜系統開發 — 需要多個專業領域的整合
需要多方專業的綜合任務 — 例如同時需要後端、前端、數據庫設計
重要決策分析 — 多角度分析降低風險
長文件生成 — 架構師規劃大綱，專家填充內容，審查員把關品質

🎯 使用範例

範例：建立一個電商網站

User: 建立一個電商網站

Architect Agent:
  分析需求後輸出：
  - 系統架構：前端、後端、數據庫、支付
  - 技術棧建議：Next.js + Node.js + PostgreSQL
  - 開發時程：4 個 sprint
  - 潛在風險：支付安全、高並發處理

Coder Agent:
  根據規格實作：
  - 用戶認證系統
  - 商品管理系統
  - 購物車邏輯
  - 訂單處理流程

Reviewer Agent:
  審查後提出：
  - 安全建議：SQL注入防護、XSS防護
  - 效能建議：加入緩存機制
  - 代碼品質：重構重複代碼

❓ 常見問題 FAQ

Q: 這個框架適合什麼規模的任務？

A: 建議用於中等以上複雜度的任務。簡單任務用 Single Agent 就足夠了，不需要過度工程。

Q: 如何決定需要多少個 Agent？

A: 可以從最基本的 3 個角色開始（Architect、Coder、Reviewer）。根據任務需求，可以新增：Data Agent、Security Agent、Test Agent 等專業角色。

Q: Agent 之間如何通訊？

A: 透過結構化的 prompt 和輸出格式。每個 Agent 知道自己的角色和輸入輸出規範，確保資訊傳遞的結構性。

Q: 這個框架和 AutoGen、CrewAI 有什麼不同？

A: 那些是通用框架，這個是針對軟體開發優化的。架構更強調：需求分析→實作→審查的流程，更適合工程任務。

Q: 如果審查不通過怎麼辦？

A: 框架支持迭代。Coder 會根據 Reviewer 的反饋進行修訂，直到通過審查或達到最大迭代次數。

🛠️ 技術細節

框架語言：Python
License：Apache 2.0
依賴：OpenAI API / Claude API / 本地模型（皆可）

🚀 開始使用

詳見 GitHub：arbiger/calling-agent-squad

License: Apache 2.0

🔗 相關技能

OpenClaw — 多 Agent 協作框架的底層系統
Coaching-Me — 單一主題的深度教練
Learn-With-Coach — 系統化學習技能

🇺🇸 English Version

💡 The Problem This Solves

Single AI Agents have capability limits. When tasks are too complex, single Agents often:

Miss important details
Imbalance between depth and breadth of analysis
Lack independent quality control mechanisms

Multiple Agents collaborating can break through these limitations. Each Agent focuses on its own domain, and through structured collaboration, they complete complex tasks that a single Agent cannot handle.

🏗️ Architecture

The framework uses Architect → Coder → Reviewer as the core flow, forming a complete “Plan → Execute → Review” cycle.

Architect Agent

Role: Analyze requirements, plan architecture, define technical roadmap

Specializes in: Systematic thinking, technical overview, risk identification

Outputs: Technical specification documents, architecture diagrams, implementation plans

Coder Agent

Role: Handle specific implementation, write code, solve technical problems

Specializes in: Code generation, algorithm implementation, detailed execution

Outputs: Runnable code, unit tests, technical documentation

Reviewer Agent

Role: Review code or solutions, ensure quality, suggest improvements

Specializes in: Logic analysis, vulnerability identification, code optimization

Outputs: Review reports, improvement suggestions, quality assessments

⚡ Workflow

User/LLM Router
     ↓
Architect Agent (Analyze requirements)
     ↓
Generate technical specifications
     ↓
Coder Agent (Execute implementation)
     ↓
Generate code
     ↓
Reviewer Agent (Quality control)
     ↓
[If revision needed] → Coder Agent (Revise)
     ↓
[Passed review] → Final output

🔄 Single Agent vs Multi-Agent Comparison

Dimension	Single Agent	Multi-Agent (Squad)
Complexity Handling	Limited, easily overloaded	Can分工 handle complex tasks
Specialization	Single role, good at everything but not expert at anything	Each Agent has specialized role and expertise
Review Mechanism	Self-checking, blind spots common	Independent Reviewer for quality control
Output Quality	Depends on Agent capability ceiling	Improved through collaboration
Scalability	Limited by single Agent	Can add Agent roles as needed

💼 Use Cases

Large Project Architecture Planning — Requires multi-faceted considerations
Complex System Development — Requires multiple professional domain integration
Comprehensive Tasks Requiring Multiple Expertise — e.g., requiring backend, frontend, and database design simultaneously
Important Decision Analysis — Multi-angle analysis reduces risk
Long-form Document Generation — Architect plans outline, experts fill content, Reviewer ensures quality

🎯 Example: Building an E-commerce Website

User: Build an e-commerce website

Architect Agent:
  After analysis, outputs:
  - System architecture: Frontend, Backend, Database, Payment
  - Tech stack suggestions: Next.js + Node.js + PostgreSQL
  - Development timeline: 4 sprints
  - Potential risks: Payment security, high concurrency handling

Coder Agent:
  Implements based on specs:
  - User authentication system
  - Product management system
  - Shopping cart logic
  - Order processing flow

Reviewer Agent:
  After review, suggests:
  - Security recommendations: SQL injection protection, XSS protection
  - Performance recommendations: Add caching mechanism
  - Code quality: Refactor duplicate code

❓ FAQ

Q: What size tasks is this framework suitable for?

A: Recommended for medium-to-high complexity tasks. Simple tasks are fine with Single Agent—no need for over-engineering.

Q: How do I decide how many Agents I need?

A: Start with the most basic 3 roles (Architect, Coder, Reviewer). Based on task needs, you can add: Data Agent, Security Agent, Test Agent, etc.

Q: How do Agents communicate with each other?

A: Through structured prompts and output formats. Each Agent knows its role and input/output specifications, ensuring structured information transfer.

Q: How is this different from AutoGen or CrewAI?

A: Those are general-purpose frameworks; this is optimized for software development. The architecture emphasizes: requirements analysis → implementation → review process, making it more suitable for engineering tasks.

Q: What if the review doesn’t pass?

A: The framework supports iteration. The Coder will revise based on Reviewer’s feedback until the review passes or maximum iterations are reached.

🛠️ Technical Details

Framework Language: Python
License: Apache 2.0
Dependencies: OpenAI API / Claude API / Local models (any works)

🚀 Getting Started

See GitHub: arbiger/calling-agent-squad