多 Agent 团队协作

交互式部署多 Agent 团队协作架构，支持自定义团队结构、预设模板和混合模式

Security Scan

The skill generally does what it claims (create team directories and generate agent config) but has several incoherences and risky behaviors (hardcoded /root paths, missing required binaries, and blind copying of auth files) that you should review before installing.

Purpose & Capability

The files and scripts match the described purpose (interactive creation of multi‑agent teams, generation of configuration snippets, directory creation). However the declared requirements list only 'bash' while the scripts invoke python3 and the 'openclaw' CLI; that mismatch is inconsistent. Also the scripts operate on /root/.openclaw (hardcoded) which is more privileged than the SKILL.md's suggested user-level install path (~/.openclaw).

Instruction Scope

The deploy script will read and copy files from /root/.openclaw/agents/main/agent/auth-profiles.json and models.json into every created agent directory. Copying authentication profiles is a sensitive action (it duplicates whatever credentials/config are stored in the main agent). The scripts also require invoking openclaw CLI commands and writing into the global OpenClaw directories; these side-effects go beyond merely generating templates.

Install Mechanism

There is no external install/download step (instruction-only plus local scripts), so there is no remote code fetch. That's lower install risk. The included scripts are the only executable content.

Credentials

The skill declares no environment variables, but it reads and copies local configuration files that likely contain credentials (auth-profiles.json). It does not request these explicitly, yet will duplicate them into many subagent directories if present. Also the metadata omits required binaries: python3 and openclaw are used but not declared.

Persistence & Privilege

The scripts create and modify system/global OpenClaw state under /root/.openclaw (agents, backups, generated snippets). This requires filesystem privileges and will increase the attack surface by proliferating auth files. 'always' is false and the skill doesn't auto-enable itself, but the actions are persistent and system‑wide if run.

What to consider before installing

This skill mostly does what it says (create team dirs and generate agent config), but review and consider the following before installing or running it: - The deploy/verify scripts assume /root/.openclaw (hardcoded). If you run as a non-root user the script may fail; running as root will operate on global OpenClaw data. Make sure you understand and consent to writes under /root/.openclaw. - The scripts copy auth-profiles.json and models.json from the main agent into every created agent directory. Inspect /root/.openclaw/agents/main/agent/auth-profiles.json first — it may contain provider credentials or tokens you do not want duplicated. Consider removing secrets or sanitizing that file before running. - The skill metadata only lists 'bash' as required, but the scripts use python3 and the 'openclaw' CLI. Ensure python3 and openclaw are present and correct, or the script will error. The metadata should be corrected to declare these dependencies. - There are minor bugs/inconsistencies in deploy.sh (array assignment syntax, reliance on undeclared variables) that could cause unexpected behavior; review the script locally and test in a safe/non-production environment first. - Because the script writes persistent config and directory structures and grants tools (the generated snippet includes 'browser' and sessions tools), prefer running this in a staging environment and backing up openclaw.json (the script already attempts a backup). Do not run blindly on production without auditing auth-profiles.json and confirming intended models/tools. If you want to proceed: (1) inspect auth-profiles.json and models.json, (2) run the deploy script in a sandbox or test instance, (3) correct the declared required binaries to include python3 and openclaw, and (4) consider modifying the script to avoid duplicating credentials or to prompt before copying auth files.

multi-agent-teams 技能

版本： v2.0
作者： CXLight
描述： 交互式部署多 Agent 团队协作架构

---

📋 功能

• 交互式配置 — 询问用户需求，创建合适的团队
• 预设模板 — 4 个标准团队快速部署
• 自定义团队 — 完全灵活的团队结构
• 混合模式 — 预设 + 自定义组合
• 自动生成 — SOUL.md/AGENTS.md 模板
• 配置验证 — 部署后自动验证

---

🚀 快速开始

安装

clawhub install multi-agent-teams

部署

cd ~/.openclaw/skills/multi-agent-teams
./scripts/deploy.sh

---

📖 使用方式

方式 1：交互式配置（推荐）

./scripts/deploy.sh

脚本会询问：

1. 配置模式（预设/自定义/混合）
2. 团队数量和名称
3. 每个团队的成员构成

方式 2：预设模板

./scripts/deploy.sh --preset

方式 3：帮助

./scripts/deploy.sh --help

---

🎯 配置模式

模式 1：预设模板

4 个标准团队：

• code (CTO) — 前端、后端、测试、产品、算法、审计
• stock (CIO) — 分析、风控、持仓、研究
• social (CMO) — 内容、排期、互动、数据分析
• flow (COO) — 工作流、定时、集成、监控

模式 2：自定义团队

创建任何你想要的团队：

• 客服团队
• 研究团队
• 设计团队
• 任何你需要的团队

模式 3：混合模式

预设 + 自定义组合

---

📁 目录结构

multi-agent-teams/
├── SKILL.md                 # 本文件
├── README.md                # 详细说明
├── templates/
│   ├── SOUL.md.tmpl         # SOUL.md 模板
│   └── AGENTS.md.tmpl       # AGENTS.md 模板
├── scripts/
│   ├── deploy.sh            # 部署脚本
│   └── verify.sh            # 验证脚本
└── generated/
    └── openclaw-snippet.json # 生成的配置

---

🧪 验证

./scripts/verify.sh

---

⚠️ 注意事项

1. 备份配置 — 部署前自动备份
2. 单一入口 — 所有请求通过 main-agent
3. 不要越级 — BOSS → main → 团队领导 → 成员
4. 重启生效 — 配置后需要 openclaw gateway restart

---

🐛 故障排查

Agent 不识别

openclaw agents list

sessions_spawn 失败

检查 subagents 配置是否包含目标成员。

---

📚 完整文档

技术文档：/root/.openclaw/docs/multi-agent-team-architecture.md

---

📝 更新日志

v2.0 (2026-03-16)

• ✨ 新增交互式配置
• ✨ 支持自定义团队结构
• ✨ 混合模式（预设 + 自定义）

v1.0 (2026-03-16)

• 初始版本