OpenClaw Multi-Agent Orchestration: Running Multiple AI Agents

One AI agent is powerful. Multiple specialized agents working together is transformative. Here's how to set up multi-agent orchestration with OpenClaw.

Why Multiple Agents?

A single agent trying to do everything is like one employee handling sales, engineering, marketing, and support. It works, but specialized agents perform better:

• Research Agent — optimized for web search, data extraction, summarization
• Writing Agent — tuned for content creation, email drafting, copywriting
• Coding Agent — configured for development, code review, debugging
• Personal Agent — handles scheduling, reminders, personal tasks
• Business Agent — manages clients, invoicing, project tracking

Each agent has its own personality, memory, model configuration, and skill set.

Architecture Patterns

Pattern 1: Router Agent

One primary agent receives all messages and delegates to specialists:

User → Router Agent → Writing Agent (for content tasks)
                   → Research Agent (for research tasks)
                   → Coding Agent (for dev tasks)

The router analyzes each message and forwards to the most appropriate specialist.

Pattern 2: Independent Agents

Each agent has its own channel:

Telegram DM → Personal Agent
#writing Slack channel → Writing Agent
#dev Slack channel → Coding Agent

You choose which agent to talk to based on the channel.

Pattern 3: Hierarchical Delegation

Agents can call other agents:

User: "Research competitors and write a blog post about our advantages"
Personal Agent → delegates research to Research Agent
Research Agent → returns findings
Personal Agent → delegates writing to Writing Agent
Writing Agent → returns draft
Personal Agent → sends draft to user

Setting Up Multiple Agents

Option 1: Multiple Containers

Run separate OpenClaw instances, each configured as a specialist:

# Personal assistant on port 3000
docker run -d --name openclaw-personal -p 3000:3000 ...

# Coding assistant on port 3001
docker run -d --name openclaw-coding -p 3001:3000 ...

# Research assistant on port 3002
docker run -d --name openclaw-research -p 3002:3000 ...

Each gets its own SOUL.md, skills, and model configuration.

Option 2: Persona Switching

A single OpenClaw instance with multiple personas:

openclaw config set personas.writer.systemPrompt "You are a professional copywriter..."
openclaw config set personas.writer.model "claude-opus-4"

openclaw config set personas.researcher.systemPrompt "You are a thorough researcher..."
openclaw config set personas.researcher.model "claude-sonnet-4-5"

Switch with: > "Switch to writer mode" or > "/persona writer"

Option 3: Skill-Based Routing

Configure automatic routing based on message content:

openclaw config set routing.rules '[
  {"match": "code|bug|PR|deploy", "persona": "developer"},
  {"match": "write|draft|email|post", "persona": "writer"},
  {"match": "research|find|compare", "persona": "researcher"}
]'

Model Optimization per Agent

Different agents benefit from different models:

Memory Across Agents

Shared Memory

All agents access the same memory store. Information learned by one agent is available to all.

Scoped Memory

Each agent maintains its own memory context. The research agent remembers research topics, the writing agent remembers style preferences.

Hybrid

Shared facts (your name, job, preferences) + scoped context (agent-specific knowledge).

Cost Management

Multiple agents can increase costs. Optimize by:

1. Use cheap models for routing — Haiku for the router agent
2. Use expensive models only where quality matters — Opus for writing
3. Set per-agent token limits — prevent any single agent from running up costs
4. Disable unused agents — only run what you actively use

Real-World Example

A solopreneur setup:

1. Morning: Personal agent sends daily briefing
2. Client email arrives: Router detects email → sends to Writing Agent → drafts reply
3. Need research: "Research the top 5 CRM tools" → Research Agent handles it
4. Blog time: Writing Agent drafts the weekly blog post using Research Agent's findings
5. Bug report: Coding Agent triages the issue and creates a fix PR
6. Evening: Personal Agent summarizes the day

All coordinated through Telegram conversations.

On ClawTank

ClawTank currently provides one container per user — perfect for single-agent or persona-switching setups. For multi-container setups, each agent needs its own deployment. Contact us for team/enterprise multi-agent configurations.

Get Started

Start with a single agent on ClawTank. Once you outgrow it, add specialized agents for your most common tasks. The multi-agent future is here.