🦞 OpenClaw Multi-Agent System

8-agent AI system with shared memory, smart model routing, auto-failover, and Telegram integration.

Architecture

graph TD
    TG[📱 Telegram] --> A1

    subgraph Pairs["Agent Pairs (Primary ↔ Backup)"]
        A1[🎯 Agent 1\nGemini 2.5 Flash\nOrchestrator] <--> A5[🛡️ Agent 5\nClaude Haiku\nBackup Orchestrator]
        A2[🎨 Agent 2\nDeepSeek Chat\nCreative] <--> A6[Agent 6\nQwen 3.5\nBackup Creative]
        A3[🔍 Agent 3\nDeepSeek Reasoner\nAnalytical] <--> A7[Agent 7\nQwen3 Max\nBackup Analytical]
        A4[💻 Agent 4\nClaude Opus\nTechnical] <--> A8[Agent 8\nQwen3 Coder\nBackup Technical]
    end

    subgraph Workspace["📁 Shared Workspace (/root/.openclaw/workspace/)"]
        MEM[MEMORY.md]
        USER[USER.md]
        SOUL[SOUL.md]
        HEALTH[health-state.json]
    end

    A1 & A2 & A3 & A4 & A5 & A6 & A7 & A8 --> Workspace

    subgraph Failover["🔄 Failover Daemon"]
        FD[agent_failover.py]
        FD --> HEALTH
        FD -->|"restart gateway"| GW[OpenClaw Gateway\n:18789]
    end

    GW --> A1

Model Routing

flowchart LR
    MSG[User Message] --> R{Router\nagent1}

    R -->|"Chat / Q&A\nAnalisis file"| G[🟢 Google Gemini 2.5 Flash\nDirect API — no markup]
    R -->|"Generate panjang\nLaporan / Artikel"| DS[🔵 DeepSeek Chat\nOutput 2.8x lebih murah]
    R -->|"Reasoning\nResearch"| DR[DeepSeek Reasoner\nagent3]
    R -->|"Coding\nDevOps"| CL[Claude Opus\nagent4]

    G -->|"Rate limit / error"| FB1[DeepSeek Chat]
    FB1 -->|"Gagal"| FB2[ModelStudio Qwen]
    FB2 -->|"Last resort"| OR[OpenRouter\n⚠️ markup 10-20%]

Model per Agent

Shared Memory Structure

/root/.openclaw/workspace/
├── SOUL.md          ← Agent personality & identity
├── USER.md          ← User profile (mas Aris)
├── MEMORY.md        ← Long-term shared memory
├── AGENTS.md        ← Agent roles & routing rules
├── SYNC.md          ← Multi-agent sync protocol
├── TOOLS.md         ← Tool configs & credentials
├── health-state.json ← Live agent health (written by daemon)
├── memory/          ← Daily logs per date
│   └── YYYY-MM-DD.md
├── diary/           ← Agent reflections
├── tasks/           ← Lessons learned
└── scripts/         ← Utility scripts

Each agent reads the same shared workspace but has its own identity files:

/root/.openclaw/agents/agentN/agent/
├── SOUL.md          ← Agent-specific personality
├── AGENTS.md        ← Agent-specific routing rules
├── IDENTITY.md      ← Name, emoji, vibe
├── USER.md          ← User profile copy
└── auth-profiles.json ← API keys for this agent

Failover System

sequenceDiagram
    participant D as Failover Daemon
    participant H as health-state.json
    participant OC as OpenClaw Gateway
    participant TG as Telegram

    loop Every 60s
        D->>D: Check provider health
        D->>H: Write status
    end

    Note over D: agent1 fails 2x
    D->>OC: Update bindings → agent5
    D->>OC: Restart gateway
    TG->>OC: Message
    OC->>OC: Route to agent5 (backup)
    Note over D: agent1 recovers 3x
    D->>OC: Restore bindings → agent1

Pair System

Installation

Prerequisites

# Node.js v22+
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.0/install.sh | bash
nvm install 22 && nvm use 22

# Install OpenClaw
npm install -g openclaw

# Python 3 (for failover daemon)
python3 --version

1. Clone repo

git clone https://github.com/arist130194/openclaw-agents.git ~/openclaw
cd ~/openclaw

2. Setup API Keys

cp .env.example .env
nano .env   # isi dengan key asli

ANTHROPIC_API_KEY=sk-ant-...
DEEPSEEK_API_KEY=sk-...
OPENROUTER_API_KEY=sk-or-v1-...
DASHSCOPE_API_KEY=sk-...
GOOGLE_API_KEY=AIzaSy...

3. Setup OpenClaw

# Init wizard
openclaw wizard

# Copy agent configs
for i in 1 2 3 4 5 6 7 8; do
  mkdir -p ~/.openclaw/agents/agent$i/agent
  cp agents/agent$i/SOUL.md ~/.openclaw/agents/agent$i/agent/
  cp agents/agent$i/AGENTS.md ~/.openclaw/agents/agent$i/agent/
done

# Copy workspace
cp -r workspace/* ~/.openclaw/workspace/

4. Mulai failover daemon

chmod +x agents/start_failover.sh
./agents/start_failover.sh start

# Cek status
./agents/start_failover.sh status

# Lihat log
./agents/start_failover.sh log

5. Jalankan OpenClaw TUI

openclaw tui

6. TUI Watchdog (opsional)

Untuk recovery otomatis kalau TUI hang / stale lock setelah Ctrl+C:

chmod +x tui_watchdog.sh

# One-time cleanup
./tui_watchdog.sh clean

# Background monitor (check tiap 30 detik)
./tui_watchdog.sh watch 30

Environment Variables

⚠️ Gunakan API key berbeda untuk setiap server (lokal vs VPS) agar tidak berebut rate limit.

Scripts

Web Panel

Management agent, API key, model routing, dan creative config via browser.

# Install & jalankan
pip install flask flask-cors
python3 panel/app.py

# Atau via systemd
systemctl start openclaw-panel

Akses: http://[IP-SERVER]:7842

Lupa token?

panel/reset-token.sh               # reset ke default
panel/reset-token.sh token-baru    # set token baru

Docs lengkap: docs/PANEL-GUIDE.md · docs/PANEL-TOKEN-RESET.md

References

• OpenClaw Multi-Agent System
• Multi-Agent Shared Memory