multi-agent-architecture（多 Agent 架构）

Multi-Agent Architecture — Run Multiple AI Agents on One Server

What You Get

• 💰 5 agents on one €47/mo VPS — no need for separate servers
• ⚡ 60% fewer rate limit hits with multi-token strategy across agents
• 📉 75% less token burn with optimized AGENTS.md (target <5KB per agent)
• 🔄 Self-healing — crashed services auto-restart in under 5 minutes
• 🧠 Shared knowledge — agents access each other's discoveries via knowledge graph
• 🎯 Specialized agents — each does one thing well instead of one bot doing everything badly

When to Use

• You need more than one agent with different specializations
• One agent is hitting rate limits and you want to split the load
• You want agents for different tasks: ops, trading, security, freelancing
• You need isolated workspaces so agents don't interfere with each other
• You want to route different Telegram groups/topics to different agents

Architecture Overview

┌─────────────────────────────────────────────────┐
│                 OpenClaw Gateway                  │
│          (single process, multiple agents)        │
├──────────┬──────────┬──────────┬────────────────┤
│  Agent 1 │  Agent 2 │  Agent 3 │   Agent N      │
│  (main)  │  (ops)   │  (trade) │   (custom)     │
├──────────┼──────────┼──────────┼────────────────┤
│workspace │workspace │workspace │  workspace     │
│          │  -ops    │  -trade  │  -custom       │
├──────────┴──────────┴──────────┴────────────────┤
│              Shared Infrastructure               │
│     (Docker, monitoring, LightRAG, backups)      │
└─────────────────────────────────────────────────┘

Quick Start

Step 1: Plan agents

Step 2: Create workspaces

mkdir -p ~/.openclaw/workspace-ops/{skills,memory,scripts,state}
mkdir -p ~/.openclaw/workspace-security/{skills,memory,scripts,state}

Step 3: Configure openclaw.json

{
  "agents": {
    "main": {
      "name": "Main",
      "model": "anthropic/claude-sonnet-4-6",
      "workspace": "workspace",
      "channels": { "telegram": { "botToken": "TOKEN_1" } }
    },
    "ops": {
      "name": "Ops",
      "model": "anthropic/claude-sonnet-4-6",
      "workspace": "workspace-ops",
      "channels": { "telegram": { "botToken": "TOKEN_2" } }
    }
  }
}

Step 4: Write AGENTS.md per agent

Keep each AGENTS.md under 5KB. Every byte loads into context every message. Smaller = cheaper.

Step 5: Share skills via symlinks

ln -s ~/.openclaw/workspace/skills/self-improving \
      ~/.openclaw/workspace-ops/skills/self-improving

Share utilities. Don't share specialized skills.

Rate Limit Management

Multi-token strategy

Split agents across 2+ Claude subscriptions:

{
  "auth-profiles": {
    "primary": { "token": "TOKEN_A" },
    "secondary": { "token": "TOKEN_B" }
  }
}

Model priority

• Critical agents: Opus or Sonnet
• Background agents: Sonnet only
• When approaching limits: downgrade heavy agents mid-week

Telegram Routing

Option A: Separate bots (recommended) — one bot per agent, cleanest.

Option B: Forum topics — one supergroup, each topic routes to a different agent.

Option C: Commands — /ops check disk → ops agent, everything else → main.

Monitoring & Self-Healing

Heartbeat per agent

Each agent has a HEARTBEAT.md — minimal checks, alert only on problems.

Self-healing script (cron every 5 min)

#!/bin/bash
if ! openclaw gateway status | grep -q "running"; then
    openclaw gateway restart
fi
for svc in langfuse n8n lightrag; do
    if docker compose -f ~/docker/$svc/docker-compose.yml ps | grep -q "Exit"; then
        docker compose -f ~/docker/$svc/docker-compose.yml restart
    fi
done

Agent watchdog

Track last response time per agent. Silent > 15 min → alert.

Shared Memory

Three-tier architecture:

1. Hot: MEMORY.md — in context, instant, free, per-agent
2. Warm: memory_search — vector search, instant, free
3. Deep: Knowledge graph (LightRAG) — cross-agent, ~3-8 sec, small cost

See lightrag-knowledge-base skill for deep memory setup.

Backup

Back up daily: openclaw.json, auth-profiles, all MEMORY.md, all memory/ dirs. Back up weekly: skills/, scripts/, docker configs. Keep 7 days. Automate via cron.

Production Checklist

• [ ] Each agent: own workspace + AGENTS.md + MEMORY.md
• [ ] AGENTS.md < 5KB per agent
• [ ] Dedicated Telegram bot per agent (or topic routing)
• [ ] allowedChatIds on all bots
• [ ] Rate limits distributed across tokens
• [ ] Heartbeat configured per agent
• [ ] Self-healing cron running
• [ ] Daily backups automated
• [ ] Bot tokens secured (not in git/chat)

Common Mistakes

1. Too many agents too fast — start with 2, add as needed
2. Giant AGENTS.md — 20KB = wasted tokens every message
3. No rate limit plan — 5 agents, one token = rate limits by Wednesday
4. Shared workspace — agents overwrite each other's memory
5. No monitoring — agent dies silently, nobody notices for hours