multi-agent-deployment（多 Agent 部署）

What This Skill Does

Guides you through deploying 3-5 specialized AI agents in OpenClaw that work as a coordinated fleet. Based on a real production setup running on a Hostinger VPS with Docker.

Included Files

Step-by-Step Setup

1. Create Workspace Structure

python3 agent_setup.py --agents pat scout publisher builder --base /data/.openclaw

Creates workspace-{agent}/ with SOUL.md, MEMORY.md, drafts/, skills/, .claude/settings.json, .claudeignore.

2. Define Each Agent's Role

Edit each workspace-{agent}/SOUL.md:

• Set the agent's mission and responsibilities
• Define which tools it uses
• Add hard limits and escalation rules

3. Generate Routing Config

# Preview output
python3 routing_config.py --agents main scout publisher builder

# Write directly to openclaw.json
python3 routing_config.py --agents main scout publisher builder \
  --output /data/.openclaw/openclaw.json

Configures model routing with OpenRouter fallbacks (minimax → deepseek → kimi).

4. Set Up Cron Jobs

Add to your cron/jobs.json for each agent:

{
  "name": "Agent: Daily Run",
  "agentId": "scout",
  "schedule": { "expr": "0 10 * * *" },
  "enabled": true
}

5. Deploy to VPS

bash deploy.sh --vps root@your-vps-ip --key ~/.ssh/your_key

6. Sync Agent Memory

Run nightly or manually to propagate cross-agent intelligence:

python3 memory_sync.py --base /data/.openclaw --agents pat scout publisher builder

Architecture Pattern

Coordinator (main) — always-on Telegram, approval queue, briefings
    ├── Scout       — market intel, inbound monitoring, trends
    ├── Publisher   — content drafts for Twitter/LinkedIn/video
    └── Builder     — skill development, marketplace research

Each agent has:

• Isolated workspace with its own SOUL.md and memory
• Separate cron schedule
• Model routing with fallbacks via OpenRouter
• Shared memory sync via Cross-Agent Intel

Requirements

• OpenClaw running on a VPS (Docker)
• OpenRouter API key (for model routing)
• SSH access to your VPS

What Makes This Different

• Real production patterns — not examples, this is a live setup
• Isolation by design — each agent has its own workspace and memory
• Fallback routing — agents keep running if a model goes down
• Memory persistence — agents remember context across sessions and compaction