OpenClaw Agent Teams: How to Build a Multi-Agent Crew
A single agent can do a lot. A team of agents can run your entire workflow. This guide shows you how to build, configure, and manage an OpenClaw agent team from scratch.
Why Use Agent Teams?
A single AI agent is like hiring one person to do every job at your company. It works for small tasks, but it breaks down when you need specialized knowledge, parallel work, or quality checks. Agent teams solve this by giving each agent a specific role with focused instructions.
Each agent masters one skill instead of being mediocre at everything
Multiple agents work simultaneously on different parts of a task
One agent creates, another reviews — built-in checks and balances
Step 1: Design Your Team Structure
Before writing any configuration, decide which agents you need. Here are three proven team structures:
Content Team (5 agents)
Agents: PM, Writer, SEO Analyst, Researcher, Social Media
Flow: PM plans content calendar → Researcher gathers data → Writer creates draft → SEO optimizes → Social Media creates distribution posts
Best for: Blogs, newsletters, social media management
Ops Team (4 agents)
Agents: PM, Metrics Analyst, Alert Monitor, Report Generator
Flow: Alert Monitor watches KPIs → triggers PM on anomalies → Metrics Analyst investigates → Report Generator creates summary
Best for: SaaS dashboards, e-commerce monitoring, DevOps
Support Team (3 agents)
Agents: Ticket Router, Knowledge Agent, Escalation Agent
Flow: Ticket Router categorizes incoming requests → Knowledge Agent handles common questions → Escalation Agent handles complex cases
Best for: Customer support, helpdesk, internal IT
Step 2: Create Your Agents
Use the openclaw agents add command to create each agent. Here is a content team setup:
# Create each agent
openclaw agents add orion # PM / Coordinator
openclaw agents add echo # Content Writer
openclaw agents add radar # SEO Analyst
openclaw agents add scout # Researcher
openclaw agents add pulse # Social Media
# Set up authentication
openclaw models auth paste-token --provider anthropic
# Verify all agents are registered
openclaw agents list
# Output:
# orion (PM)
# echo (Writer)
# radar (SEO)
# scout (Researcher)
# pulse (Social)For a detailed walkthrough of the agents add command, options, and authentication setup, see our complete guide to adding agents in OpenClaw.
Step 3: Configure agents.md
The agents.md file is the team's org chart. It lives in the root of your workspace and defines how agents relate to each other:
# Content Team
## Team Mission
Produce 3 blog posts per week. Each post should
be 1,500+ words, SEO-optimized, and published
with social media distribution.
## Agents
- @Orion — PM and coordinator. Plans content
calendar, assigns tasks, tracks progress.
- @Echo — Content writer. Writes blog posts,
emails, and landing page copy.
- @Radar — SEO analyst. Keyword research,
content optimization, meta descriptions.
- @Scout — Researcher. Gathers data, competitor
analysis, fact-checking.
- @Pulse — Social media. Creates distribution
posts for Twitter, LinkedIn, newsletters.
## Workflow
1. @Orion creates weekly content plan
2. @Scout researches topics assigned by @Orion
3. @Echo writes drafts using @Scout's research
4. @Radar optimizes drafts for SEO
5. @Pulse creates social posts from final content
## Rules
- All communication in English
- @Orion approves all final content before publish
- @Echo must include sources from @Scout's research
- @Radar must provide keyword targets before @Echo
starts writingThe agents.md file is read by the gateway and shared with all agents as context. Each agent knows who else is on the team and what their role is. For more on agent communication patterns, read our agent-to-agent communication guide.
Step 4: Define Handoffs in SOUL.md
Each agent's SOUL.md needs a Handoffs section that tells it when and how to pass work to teammates:
# Echo
## Role
You are a content marketing writer for a SaaS
company. You write blog posts based on research
briefs and keyword targets.
## Rules
- ALWAYS respond in English
- Target 1,500-2,000 words per blog post
- Include research citations from @Scout
- Follow keyword targets from @Radar
## Handoffs
- BEFORE writing: Ask @Radar for keyword targets
and search intent analysis
- BEFORE writing: Ask @Scout for research brief
and supporting data
- AFTER writing: Send finished draft to @Radar
for SEO review and optimization
- AFTER @Radar approval: Send to @Orion for
final review
- AFTER publish: Send title + URL to @Pulse
for social media distributionClear handoffs are what separate a group of independent agents from a functioning team. Without explicit handoff rules, agents either try to do everything themselves or wait indefinitely for instructions.
Step 5: Run and Monitor Your Team
Start the gateway and your team is live:
# Start the gateway
openclaw gateway start
# Kick off a workflow by messaging the PM
openclaw agent --agent orion --message \
"Plan this week's content. Topic: AI agent trends."
# Orion will:
# 1. Create a content plan
# 2. Assign research to @Scout
# 3. Send keyword targets request to @Radar
# 4. Assign writing tasks to @Echo
# 5. Queue social distribution with @Pulse
# Monitor activity
openclaw gateway statusFor always-on teams that run on a schedule, set up HEARTBEAT.md with cron-style triggers. See our OpenClaw automation guide for scheduling and monitoring setup.
Team Orchestration Patterns
There are three main ways to orchestrate your agent team:
Hub & Spoke (Coordinator)
One PM agent receives all tasks and delegates to specialists. Best for structured workflows where order matters.
User → PM → assigns to Writer/SEO/Social → PM reviews all outputPipeline (Sequential)
Each agent passes work to the next in a defined sequence. Best for content production and data processing.
Research → Write → SEO Review → Publish → Social DistributionPeer-to-Peer (Mesh)
Any agent can communicate with any other agent directly. Best for creative work and brainstorming.
Writer ↔ Researcher ↔ SEO (all communicate freely based on need)Most production teams use a hybrid: Hub & Spoke for task assignment, Pipeline for execution. For a deeper dive into orchestration strategies, read our AI agent orchestration guide.
Real Example: Our 6-Agent Content Team
Here is the exact agent team we use to run content operations. This setup produces 3 blog posts, 15 social posts, and 2 email campaigns per week:
| Agent | Role | Model | Cost/week |
|---|---|---|---|
| Orion | PM & Coordinator | Claude Haiku | ~$0.50 |
| Echo | Content Writer | Claude Sonnet | ~$3.00 |
| Radar | SEO Analyst | GPT-4o | ~$1.50 |
| Scout | Researcher | Claude Sonnet | ~$2.00 |
| Pulse | Social Media | Claude Haiku | ~$0.30 |
| Metrics | Analytics | GPT-4o | ~$1.00 |
Total cost: ~$8.30/week for a full content operation. For tips on reducing agent costs further, read how we reduced our AI agent cost by 16x.
Related Guides
Frequently Asked Questions
What is an OpenClaw agent team?
An OpenClaw agent team is a group of AI agents that work together in the same workspace. Each agent has a specialized role defined in its SOUL.md file, and they communicate through @mentions and task handoffs. A typical team has 3-10 agents — for example, a PM agent that plans tasks, a writer agent that creates content, and an SEO agent that optimizes it. The team is coordinated through the agents.md configuration file.
How do OpenClaw agents communicate with each other?
OpenClaw agents communicate through the agent-to-agent tool. When an agent mentions another agent with @AgentName in its response, OpenClaw routes the message to that agent. The receiving agent processes the message and can respond back. This enables patterns like delegation (PM assigns task to Writer), pipeline (Writer passes draft to SEO), and peer-to-peer collaboration. Communication rules are defined in each agent's SOUL.md Handoffs section.
What is agents.md and how does it work?
agents.md is a configuration file that sits in the root of your workspace and defines team-level settings. It lists all active agents, their relationships, shared context, and communication rules. While each agent's SOUL.md defines individual behavior, agents.md defines team behavior — who can talk to whom, task routing rules, and shared memory. Think of SOUL.md as a job description and agents.md as an org chart.
How many agents should I have on a team?
Start with 3-5 agents. A good starter team covers: (1) a coordinator/PM agent that plans and delegates, (2) 2-3 specialist agents that do the actual work, and (3) optionally a QA/review agent. Don't create agents for tasks that happen rarely — only create agents for recurring workflows. You can always add more agents later with 'openclaw agents add'. Teams with more than 10 agents become harder to coordinate.
Can different agents use different language models?
Yes. Each agent can use a different model provider. A common pattern is: Claude for writing and reasoning tasks, GPT-4o for code and data analysis, and Ollama (local) for simple repetitive tasks to save costs. You set the model per-agent in the agent's configuration. This lets you optimize for both quality and cost — expensive models for important tasks, cheaper models for routine work.
What is the difference between a single agent and a team?
A single agent handles all tasks itself. A team distributes tasks across specialized agents. Single agents work well for simple use cases (answering questions, writing individual pieces). Teams are better when you have complex workflows with multiple steps, different skill requirements, or need parallel processing. A team of 5 specialists typically outperforms a single generalist agent on complex tasks.
Build your agent team with CrewClaw
Skip the manual configuration. CrewClaw's team builder lets you create a coordinated agent crew visually. Pick roles, configure tools, and download a complete workspace package.