Claw Multi Agent

---
name: claw-multi-agent
description: Multi-agent parallel orchestration for OpenClaw. Spawn AI agents as a team — parallel research, multi-model comparison, code pipelines. Proven 50-65% time savings. Trigger words: multi-agent, parallel agents, swarm, spawn multiple agents, parallel research, compare models, deep research, comprehensive research, detailed investigation, thorough analysis, research multiple topics, 多智能体, 多个Agent, 并行调研, 并行搜索, 同时搜索, 同时调研, 深度调研, 详细调研, 全面调研, 深度研究, 详细检索, 多角度分析, 全面分析, 多个模型, 让多个AI, 分别搜索, 同时搜索, 组建团队, Agent小队, 多Agent.
---

# claw-multi-agent 🐝

> **Replace one AI with a team of AIs. Turn serial into parallel. Turn hours into minutes.**

---

## What can it do?

| Scenario | Example | Speedup |
|----------|---------|---------|
| **Parallel research** | Search 5 frameworks simultaneously, each writes a report | ~65% ⚡ |
| **Multi-model compare** | Ask Claude, Gemini, Kimi the same question at the same time | ~50% ⚡ |
| **Code pipeline** | Plan → Code → Review, auto hand-off in sequence | Quality ↑ |
| **Batch processing** | Translate / analyze / summarize multiple docs in parallel | Scales linearly |

---

## ⚡ Get started in 30 seconds

Just say something like:

- "Research LangChain, CrewAI, and AutoGen in parallel"
- "Have multiple agents search these topics and write a combined report"
- "Compare how Claude and Gemini answer this question"
- "Use multi-agent mode to do this research"

---

## 🎭 Interaction Style — How to Talk to the User

This is the recommended pattern. Every multi-agent run must follow this interaction pattern.

### Step 0 — Announce skill activation FIRST

**⚠️ Iron rule: The activation announcement must be your FIRST reply after receiving the task — before reading any files, before investigating, before spawning.**

**Why this matters**: Reading files, researching background, and spawning all take time. If you do those first, users see long silence. Worse: context compression can happen during that time, and the announcement will never be sent.

**Correct order**: Receive task → Send announcement immediately → Then read files / spawn / wait

The very first thing to say when this skill is triggered — before any planning or spawning:

```
🐝 **claw-multi-agent 已唤醒**
多智能体并行模式启动，我来组建 Agent 小队处理这个任务。
```

This tells the user the skill is active and sets expectations for what's about to happen.

### Before spawning — announce the plan

Right after the activation announcement, present the plan BEFORE calling sessions_spawn:

```
🚀 [N]个方向同时开搞，全面覆盖你的问题。

📋 任务规划：
🔍 研究员A（GLM）— [一句话任务描述]
🔍 研究员B（GLM）— [一句话任务描述]
📊 分析师（Kimi）— 先等前[N]个结果，单独召唤（note when sequential）

模式：🎯 指挥官模式（联网搜索）
预计耗时：~[X]s（[N] Agent 并行[，分析师串行跟进]）
正在派出 Agent 小队...
```

**Role emoji reference:**
| Role | Emoji | Example |
|------|-------|---------|
| Researcher | 🔍 | 🔍 研究员A（GLM）— Research XX |
| Analyst | 📊 | 📊 分析师（Kimi）— Deep comparison |
| Writer | ✍️ | ✍️ 写作者（Gemini）— Draft the report |
| Coder | 💻 | 💻 程序员（Kimi）— Implement the logic |
| Reviewer | 🔎 | 🔎 审核员（GLM）— Quality check |
| Planner | 📋 | 📋 规划师（Sonnet）— Break down tasks |

**Key rules:**
- ✅ Always list each agent with: emoji + role + **model name** + one-line task
- ✅ State the mode (指挥官/流水线/混合) and estimated time
- ✅ End announcement with: `正在派出 Agent 小队...`
- ✅ Note sequential agents as: "先等前N个结果，单独召唤"
- ❌ Never silently call sessions_spawn without announcing

### While waiting — brief note

After spawning, say one line:
```
⏳ 子 Agent 已全部出发，等结果回来...
```

### After results — structured output (not raw dump)

**Never paste sub-agent raw output directly.** Always digest and restructure by content logic — NOT by agent order.

**Recommended output order:**

```
1. 执行统计卡 ← 先让用户知道跑了什么
2. 核心结论（3-5条最重要发现）← 最有价值的放最前面
3. 分主题展开细节（按内容逻辑组织，不按子Agent顺序）← 读起来是一篇完整文章
4. 下一步行动建议 ← 落地结尾
```

**统计卡格式：**
```
## 📊 执行统计
| Agent | 模型 | 耗时 | 状态 |
|-------|------|------|------|
| 🔍 研究员A | GLM | 58s | ✅ |
| 🔍 研究员B | GLM | 62s | ✅ |
| 📊 分析师  | Kimi | 45s | ✅ |
串行需要约 165s → 并行实际 62s，节省 **62%** ⚡
```

**❌ Wrong — agent order:**
```
子Agent1的结果...
子Agent2的结果...
子Agent3的结果...  ← 读者要自己拼图，体验差
```

**✅ Right — content logic:**
```
## 核心结论
1. 最重要发现A（来自多个Agent综合）
2. 最重要发现B
...

## 详细分析：[主题1]
...（整合所有相关Agent的内容）

## 详细分析：[主题2]
...

## 下一步建议
...
```

**The main agent rewrites everything in its own words.** Sub-agent outputs are raw material, not the final answer.

### After results — deliver the report (channel-aware)

**Always save to file first. Then deliver based on the current channel.**

```python
# Step 1: Always save to file first
write("/workspace/projects/{topic-slug}/report.md", content)
```

**Then choose delivery method by channel:**

| Channel | Delivery method |
|---------|----------------|
| **feishu** + has `feishu-all-operations` skill | Create Feishu doc → send link (best UX) |
| **feishu** + no Feishu skill | `message(filePath=..., filename="report.md")` — send as attachment |
| **Discord / Telegram / Slack** | `message(message=...)` — Markdown renders normally |
| **Other / unknown** | Save file + tell the user the path |

**Why this matters:** Feishu chat does NOT render Markdown. Sending raw Markdown text shows `##`, `|---|` symbols. Always use attachment or doc link on Feishu.

```python
# Feishu (no Feishu doc skill): send as attachment
message(action="send", filePath="/workspace/projects/{topic-slug}/report.md", filename="report.md")

# Discord/Telegram: send markdown directly
message(action="send", message=report_content)
```

**End with one line:**
```
需要调整某个方向，或推送到飞书文档吗？
```

**Rules:**
- ✅ Always save `.md` file first — regardless of channel
- ✅ Check current channel before deciding how to send
- ❌ Never paste >300 words of Markdown text on Feishu — it won't render
- ❌ Never just say "报告已保存至 /path/xxx" — user can't open server paths
- ❌ Never ask "要不要我帮你整理成文档？" — just do it

### Sequential vs parallel — analyst must wait for researchers

**Critical:** Agents spawned in the same round run in parallel and share NO context with each other.

```
❌ Wrong: spawn researcher-A + researcher-B + analyst all at once
          → analyst has no data, returns empty

✅ Right: 
  Round 1: spawn researcher-A + researcher-B (parallel, independent)
  Wait for both to return...
  Round 2: main agent consolidates research results
           → then either: main agent writes analysis itself
           → or: spawn analyst with research results injected as context
```

**Best practice: Any agent that depends on another agent's output should be spawned in a later round, after collecting the dependency.**

---

## 🤖 Model Selection Guide — Which Model for Which Role

Always pick the right model for each agent. State the model explicitly in the announcement.

### Model roster

| 模型 | 别名 | 特点 | 适合角色 |
|------|------|------|---------|
| `glm` | GLM | 便宜、速度快、中文好 | 搜索、简单调研、状态检查 |
| `kimi` | Kimi | 长上下文（128k）、代码强 | 深度分析、代码、长文整合 |
| `gemini` | Gemini | 创意好、多模态 | 写作、文案、图像理解 |
| `sonnet` | Claude Sonnet | 均衡、工具调用稳 | 复杂推理、规划、审核 |
| `opus` | Claude Opus | 最强推理 | 极复杂分析、架构设计 |

### Role → Model mapping (default)

| 角色 | 默认模型 | 原因 |
|------|---------|------|
| 🔍 研究员 / Researcher | **GLM** | 轻量搜索，够用且便宜 |
| 📊 分析师 / Analyst | **Kimi** | 长上下文，处理大量资料 |
| ✍️ 写作者 / Writer | **Gemini** | 创意写作效果最好 |
| 💻 程序员 / Coder | **Kimi** | 长上下文代码理解 |
| 🔎 审核员 / Reviewer | **GLM** | 简单判断，不需重炮 |
| 📋 规划师 / Planner | **Sonnet** | 结构化规划能力强 |
| 🧐 批评者 / Critic | **Sonnet** | 逻辑严谨，挑战假设 |

### When to override defaults

- 任务很简单 → 降级到 GLM（省成本）
- 需要最高质量 → 升级到 Opus
- 用户明确指定模型 → 照用户说的来
- 多模型对比场景 → 每个 Agent 用不同模型，在公告里说明

### Always announce the model

In the pre-spawn announcement, every agent line must include the model:
```
✅ 这样：🔍 研究员A（GLM）— 调研 LangChain
❌ 这样：🔍 研究员A — 调研 LangChain
```

---

## Step 0: Always plan first (dynamic agent count)

**Never hardcode how many agents to spawn.** The right number depends on the task complexity. Always start with a planning step:

```
1. Analyze the task → identify subtopics / dimensions
2. Decide: how