一、背景：Coordination Engineering 的演进

1. Background: The Evolution of Coordination Engineering

AI工程范式的演进，正从单智能体的"驾驭与治理"（Harness Engineering），大步迈向多智能体的"协同与进化"（Coordination Engineering）。

The evolution of AI engineering paradigm is rapidly moving from "Harness Engineering" for single agents to "Coordination Engineering" for multi-agent collaboration.

核心问题

Core Problem

Agent Team 解决了"当下怎么协作"的问题，但会话结束后经验全部消失。下次遇到同类任务，Leader 仍然要从零开始规划：

Agent Team solves the problem of "how to collaborate now", but all experience disappears after the session ends. The next time you encounter a similar task, the Leader still has to plan from scratch:

• 需要几个角色？
• How many roles are needed?
• 如何分工？
• How to divide tasks?
• 谁先谁后？
• Who goes first and who goes after?
• 什么条件算完成？
• What conditions count as completion?

Team Skills 正是为破解这些问题而来。

Team Skills is created to solve these problems.

---

二、Team Skills 是什么？

2. What is Team Skills?

定义

Definition

Team Skills = 面向多Agent协作的标准化能力包规范

Team Skills = Standardized Multi-Agent Collaboration Capability Package Specification

它将 Agent Team 一次成功的协作全链路（包括需求拆解、团队组建、任务分配、通信机制、冲突处理、交付规范等）封装为标准化的团队技能。

It encapsulates the entire collaboration chain of a successful Agent Team session (including requirement analysis, team building, task allocation, communication mechanisms, conflict handling, delivery standards, etc.) into standardized team skills.

核心价值

Core Value

概念	Concept	说明	Description
Agent Team	让团队"自主高效协作"	Enables "autonomous and efficient collaboration"
Team Skills	让团队协作能力"会沉淀、可复制"	Makes team collaboration capabilities "沉淀able and replicable"

---

三、Team Skills 文件结构

3. Team Skills File Structure

Team Skills 本质上是一个文件夹目录结构：

A Team Skill is essentially a folder directory structure:

<team-skill-name>/
├── SKILL.md                              ← 这个团队叫什么、干什么、成员有谁
├── roles/                                ← 每个成员角色各自负责什么
│   ├── <role-a>.md
│   └── <role-b>.md
├── workflow.md                           ← 大家怎么配合、执行顺序是什么
├── bind.md                               ← 遇到问题怎么处理、边界在哪里
├── dependencies.yaml                     ← 依赖哪些外部工具
└── examples/ | templates/ | assets/ ...  ← 自由扩展

最小化示例

Minimal Example

最简单的情况，一个 SKILL.md 加上 roles/ 里几个角色定义，就能组起一支可用的团队：

In the simplest case, a SKILL.md plus a few role definitions in roles/ can form a usable team:

medical-team/
├── SKILL.md
└── roles/
    ├── triage.md
    └── specialist.md

简单任务两三个文件就够，复杂任务再按需要逐步补充。

Simple tasks need only 2-3 files, and complex tasks can be gradually expanded as needed.

---

四、与 Anthropic Agent Skills 的对比

4. Comparison with Anthropic Agent Skills

你可能用过 Anthropic 提出的 Agent Skills，比如给单个 Agent 写一份"技能说明书"，让它知道遇到某类任务该怎么做。

You may have used Anthropic's Agent Skills, such as writing a "skill manual" for a single Agent to know what to do when encountering certain tasks.

Team Skills 在这个基础上往上升了一步：

Team Skills goes a step further:

层级	Level	解决的问题	Problem Solved
单 Agent Skill	"一个 Agent 怎么做事"	"How a single Agent does things"
Team Skill	"一个 Agent 团队怎么配合做事"	"How an Agent team collaborates to get things done"

---

五、核心组件详解

5. Core Components

---

六、AgentTeam 架构回顾

6. AgentTeam Architecture Overview

6.1 Leader + Teammates 模式

6.1 Leader + Teammates Pattern

角色	Role	职责	Responsibilities
Leader Agent	需求分析、团队构建、任务规划	Requirement analysis, team building, task planning
Teammate Agent	任务认领、独立执行、结果汇报、协作沟通	Task claiming, independent execution, result reporting, collaborative communication

6.2 三大核心能力

6.2 Three Core Capabilities

1. 分层自主协作：Leader 智能编排，Teammates 自主执行
2. Hierarchical Autonomous Collaboration: Leader orchestrates intelligently, Teammates execute autonomously
3. Team Workspace：团队级共享文件空间，所有成员透明访问
4. Team Workspace: Team-level shared file space, transparent access for all members
5. 全生命周期管理：双层审批机制、事件驱动、持久化团队、TeamMonitor 可观测性
6. Full Lifecycle Management: Two-layer approval, event-driven, persistent teams, TeamMonitor observability

6.3 技术原则

6.3 Technical Principles

1. 共享任务列表的一致协作：所有成员共享同一个动态任务列表
2. Consistent collaboration via shared task list: All members share the same dynamic task list
3. 消息与任务双驱动模型：任务驱动核心工作流，消息通道处理非结构化沟通
4. Dual-drive model of messages and tasks: Tasks drive the core workflow, message channels handle unstructured communication
5. 角色与工具工程：RolePolicy 定义行为规范，TeamTools 赋予协调能力
6. Role and tool engineering: RolePolicy defines behavioral norms, TeamTools endows coordination capabilities

---

七、teamskill-creator 工具

7. teamskill-creator Tool

JiuwenClaw 发布了配套的创建技能"团队技能自动生成专家"（teamskill-creator）。

JiuwenClaw released a companion creation skill called "Team Skill Auto-Generation Expert" (teamskill-creator).

使用步骤

Usage Steps

1. 在 Team Skills Hub（团队技能市场）上下载 teamskill-creator
2. Download teamskill-creator from Team Skills Hub
3. 在 JiuwenClaw 上安装该技能
4. Install the skill on JiuwenClaw
5. 输入自然语言描述即可创建团队技能
6. Enter natural language description to create team skills

示例：创建医疗专家会诊团队

Example: Creating Medical Expert Consultation Team

帮我创建一个医疗专家会诊的团队技能，
要求科目齐全，能根据用户的病情描述按需加载

结果：生成了具备23位AI医学专科专家的团队技能，可根据病情动态创建多个不同专科专家团进行会诊。

Result: Generated a team skill with 23 AI medical specialists that can dynamically create multiple specialist teams based on medical conditions.

---

八、实战案例：多学科医疗专家团队联合会诊

8. Case Study: Multi-Disciplinary Medical Expert Team Consultation

工作流程

Workflow

1. 分诊角色读取用户信息（年龄、性别、症状持续时间、关键症状表现）
2. Triage role reads user information (age, gender, symptom duration, key symptoms)
3. 判断可能涉及的科室方向
4. Determine possible department directions
5. 按需动态创建对应的专科专家成员
6. Dynamic creation on demand of corresponding specialist members
7. 为每位专家分配明确的分析重点
8. Assign clear analysis focus for each expert
9. 专科专家并行分析
10. Specialist experts analyze in parallel
11. 主任医生汇总各方意见
12. Chief physician synthesizes opinions from all parties
13. 输出完整的会诊报告
14. Output complete consultation report

关键特点

Key Features

• 动态创建：不是预先固定专家，而是按需创建
• Dynamic creation: Not pre-fixed experts, but created on demand
• 并行分析：多个专家同时工作
• Parallel analysis: Multiple experts work simultaneously
• 可见可追踪：整条协作过程可见、可追踪、可复盘
• Visible and traceable: The entire collaboration process is visible, traceable, and reproducible

---

九、跨框架兼容性

9. Cross-Framework Compatibility

Team Skills 扩展的是 Agent Skills 开放标准，不依赖特定平台框架，完全可以在 Claude Code、Cursor 等支持多智能体协同的平台上零适配运行。

Team Skills extends the Agent Skills open standard, with no dependency on specific platform frameworks, and can run zero-adaptation on platforms like Claude Code and Cursor that support multi-agent collaboration.

---

十、相关项目链接

10. Related Project Links

---

十一、参考资料

11. References

1. 量子位 - "Coordination Engineering后，JiuwenClaw发布Team Skills技能范式"
2. AIMagazine - "Beyond Harness: JiuwenClaw Leads Coordination"
3. TechGig - "JiuwenClaw's Coordination Engineering Elevates Autonomous AI Collaboration"
4. Tokention - "Next Leap to Harness Engineering: JiuwenClaw Pioneers Coordination Engineering"
5. 搜狐科技 - "openJiuwen：打造高可靠、自演进、多智能体原生、算力亲和的AgentOS"