Multi-Agent Orchestration Skill — Turn OpenClaw into the CEO of a one-person company.
多 Agent 编排技能 — 让 OpenClaw 成为一人公司的 CEO。
OPC (One-Person Company) is an OpenClaw skill that turns complex tasks into multi-agent collaboration. You give the goal, the CEO (OpenClaw) breaks it down, hires the right agents, monitors progress, and delivers results.
You (Owner) → OpenClaw CEO → Sub-agents (Specialists)
- Phase 0 Context Intake — CEO reads your background, proposes a plan, waits for confirmation before executing
- Built-in Persona Library — Inject top expert mindsets (Kotler, Fowler, Jeff Dean...) into each agent role
- LZW Advisor Persona — Author's own methodology framework, ready to inject into any agent role
- Project State Persistence — Survives context compaction; full lifecycle state in
state.json - Checkpoint & Resume — Failed agents resume from last checkpoint, not from scratch
- Auto Root Cause Analysis — L1-L4 fault classification (platform / orchestration / skill / external)
- Aware Triggers — Declarative event-driven scheduling (cron / once / interval / on_message)
- Tool Discovery v2 — domain × capability tag system, precise skill recommendations
- User Model Evolution — CEO writes back learnings after each project; gets smarter over time
# Install
cd ~/.openclaw/skills
unzip agent-orchestration-v5.0.zip
# Trigger OPC by telling OpenClaw:
# "Help me build a complete [project]"
# "I need [multi-step task] done end-to-end"agent-orchestration-20260309-lzw/
├── SKILL.md ← Entry point (v5.0)
├── brain/ ← CEO decision layer
│ ├── core-flow.md ← 4-phase flow + Context Intake
│ ├── task-decomposition.md
│ ├── role-design.md
│ └── collaboration-patterns.md
├── engine/ ← Execution engine
│ ├── project_state.py ← Project lifecycle state
│ ├── trigger_engine.py ← Aware triggers
│ ├── tool_discovery.py ← Tool discovery (tag system v2)
│ └── diagnose_agent.py ← Auto root cause analysis
└── playbook/ ← Knowledge & templates
├── persona-priming.md ← Persona methodology + library index
├── personas/
│ └── lzw.md ← Built-in LZW Advisor Persona (v5.0)
├── templates/
└── scenarios/
Two types of Personas, two design logics:
| Type | Example | How it works |
|---|---|---|
| Activation (public figures) | Kotler, Fowler, Jeff Dean | One-line reference activates LLM's pre-trained knowledge |
| Injection (real individuals) | LZW (personas/lzw.md) |
Full methodology doc injected into agent context |
New paradigm: You can crystallize your own thinking framework into a Persona and inject it into your agent team.
| Case | Scale | Result |
|---|---|---|
| 315 Marketing Campaign (v1.0) | 3 agents serial | 67K tokens / $0.17 |
| 3 Venues Parallel Build (v1.4) | 4 agents parallel | All 3 venues delivered |
| KangaBase 0→1 (v2.0) | 8 agents, 4 phases | ~100 files / 8000 lines in 8h |
| Neurotech Deep Analysis (v3.1) | 4 researchers + integrator | 4500-word report / $0.10 |
- OpenClaw installed and configured
- Python 3 (for engine scripts)
Apache License 2.0 — modifications must declare changes. See LICENSE.
OPC（One-Person Company，一人公司）是一个 OpenClaw Skill，把复杂任务转化为多 Agent 协作。你说目标，CEO（OpenClaw）负责拆活儿、招人、盯进度、交结果。
用户（老板）→ OpenClaw CEO → Sub-agents（专业员工）
- Phase 0 Context Intake — CEO 理解背景，给出方案，等确认后才执行
- 内置 Persona 库 — 给角色注入 Kotler / Fowler / Jeff Dean 等顶级人才的方法论
- LZW 顾问 Persona — 作者本人的方法论框架，可直接注入任意 Agent 角色
- 项目状态持久化 — 抗 context compaction，全生命周期状态写入
state.json - 断点续传 — Agent 失败后从断点继续，不完全重跑
- 自动归因 — L1-L4 故障分层（平台/编排/Skill/外部）
- Aware 触发器 — 声明式事件驱动（cron / once / interval / on_message）
- 工具发现 v2 — domain × capability 双轴标签，精准推荐可用 Skill
- 用户模型自进化 — 每次项目结束 CEO 自动写回学习结果，越用越懂你
# 安装
cd ~/.openclaw/skills
unzip agent-orchestration-v5.0.zip
# 触发 OPC，对 OpenClaw 说：
# "帮我做一个完整的 [项目]"
# "我需要 [多步骤任务] 全链路完成"OPC 的 Persona 分为两类，设计逻辑不同：
| 类型 | 代表 | 工作原理 |
|---|---|---|
| 激活型（公众人物） | Kotler、Fowler、Jeff Dean | 一行描述即可激活 LLM 预训练知识 |
| 注入型（真实个人） | LZW（personas/lzw.md） |
完整方法论文档注入 Agent context |
新范式：用户可以把自己的思维框架沉淀为 Persona，注入到自己的 Agent 团队中。
| 案例 | 规模 | 结果 |
|---|---|---|
| 315 营销活动（v1.0） | 3 角色串行 | 67K tokens / $0.17 |
| 三会场并行搭建（v1.4） | 4 角色并行 | 3 个会场全部交付 |
| KangaBase 从零到开源（v2.0） | 8 Agent / 4 Phase | ~100文件/8000行，8小时 |
| 神经调控深度分析（v3.1） | 4 研究员 + 整合员 | 4500字报告 / $0.10 |
| 版本 | 日期 | 核心变更 |
|---|---|---|
| v5.0 | 2026-03-16 | 内置 LZW 顾问 Persona + personas/ 目录 |
| v3.1 | 2026-03-14 | 用户模型自学习：Phase 0 读取 + Phase 4 写回 |
| v3.0 | 2026-03-14 | 三层架构重构 + Context Intake + 工具发现标签体系 v2 |
| v2.0 | 2026-03-12 | Aware 触发器 + 运行时工具自发现 |
| v1.4 | 2026-03-11 | 状态持久化 + 断点续传 + 自动归因 |
| v1.0 | 2026-03-09 | 首版：角色卡 + 任务分解 + 协作模式 |
见完整 CHANGELOG.md
- 已安装并配置 OpenClaw
- Python 3（用于引擎脚本）
Apache License 2.0 — 修改文件须声明变更。详见 LICENSE。
Built with ❤️ by LZW · Powered by OpenClaw