What Does This Skill Do?

multi-agent-cn is a Chinese-language multi-agent dispatch framework. Once installed, your main agent becomes a pure dispatcher — it receives tasks, assigns them to sub-agents, and reports results. It never does the actual work itself.

That sounds simple, but it's a genuine paradigm shift: from "one AI does everything" to "one CEO leads a team that does everything."

This is exactly how SFD Lab got started. Our 13-agent system is essentially an evolved version of this dispatch pattern.

Installation and Configuration

# Install via OpenClaw CLI
clawhub install multi-agent-cn

Or download SKILL.md from clawhub.ai directly
~/.openclaw/workspace/skills/multi-agent-cn/SKILL.md

First thing after install: customize your team. Open SKILL.md and find the "Custom Configuration" section:

• Dispatcher persona: Default is "Commander" — change it to CEO, Captain, Director, whatever fits your project
• Sub-agent names: Defaults to Alpha through Echo — rename to real role names like "Frontend Dev", "Backend Engineer"
• sessionKey: Each sub-agent uses a fixed sessionKey for persistent conversation state — same agent across restarts

Real-World Use Cases

Scenario 1: Content Publishing Pipeline

Boss says: "Write a science article about AI memory, publish to SFD website."

Dispatcher (Xiao Huo Long) actions:

1. → Assign to Fox🦊: Write article content (three languages: ZH/EN/TW)
2. → Assign to Butterfly🦋: Generate cover image
3. → Assign to Octopus🐙: Publish via CMS API
4. → Assign to Hedgehog🦔: Verify article displays correctly

The dispatcher writes zero content, generates zero images, calls zero APIs. It only tracks "done or not done."

Scenario 2: Code Fix Pipeline

Received QA report: frontend layout broken on a certain page.

1. → Assign to Chameleon🦎: Fix CSS (via ACP, reads whole project)
2. → Assign to Hawk🦅: Security audit (verify no new vulnerabilities)
3. → Assign to Bee🐝: Deploy (per change manifest)
4. → Assign to Hedgehog🦔: Verify layout is fixed

Lessons Learned the Hard Way

Pitfall 1: Duplicate sessionKeys causing role confusion

Without fixed sessionKeys, "Frontend Dev" sometimes got used as "Backend Engineer." Fix: assign unique, permanent sessionKeys in SKILL.md. Once set, never change them.

Pitfall 2: The dispatcher "just does it real quick"

Most common failure mode. Dispatcher thinks "this is simple, I'll just do it" and starts using exec tools. Once the dispatcher starts executing, the whole role structure breaks. Fix: hardcode in SKILL.md that the dispatcher cannot use any execution tools. Read this constraint before every heartbeat.

Pitfall 3: Sub-agents have no memory

sessions_spawn creates stateless agents unless you pass context explicitly. Fix: include necessary background in the task parameter — e.g., "read design-spec.md before starting."

Pitfall 4: Callback loss

sessions_spawn callbacks sometimes don't arrive. Fix: don't rely on callbacks. The dispatcher actively checks: file modification times, build artifacts, git log. Completion is pulled, not pushed.

How SFD Lab Uses It

We're probably among the heaviest users of this dispatch pattern globally. 13 agents, dozens of daily tasks, three channels — all running on this framework.

Our main extensions: we scaled from 5 to 13 sub-agents, each with dedicated SOUL.md and MEMORY.md. The dispatcher also handles: detecting broken handoffs and restarting them, continuing iteration when QA fails, and automated scanning every 15 minutes so no task falls through the cracks.

If you're building your own AI team, this skill is the fastest starting point.

SFD Lab Note: One AI doing work is a tool. Multiple AIs working together is a team. multi-agent-cn is the entry point. Install it. Experience what "AI managing AI" actually feels like.