5 OpenClaw Multi-Agent Architectures That Actually Work in Production
"What are yalls architectures?" is the most asked question in the OpenClaw community. Everyone knows they want multiple agents working together, but nobody is sure how to structure them. Here are 5 proven patterns with real AGENTS.md configurations, trade-offs, and guidance on when to use each one.
The Architecture Problem
Setting up a single OpenClaw agent is straightforward. Write a SOUL.md, register it, and start chatting. The challenge starts when you add a second agent. How should they communicate? Who decides what to work on? What happens when tasks overlap?
On Reddit, architecture discussions generate more debate than any other topic. One developer spent $280 over a weekend because their 5 agents kept triggering each other in loops. Another spent 20 hours debugging a peer-to-peer setup where agents were duplicating work because no one was coordinating. A Mac Mini M4 user tried running 8 agents with Ollama and the whole system ground to a halt.
The root cause is always the same: they started building without choosing an architecture. They added agents one at a time, connected them ad hoc, and ended up with a tangled mess. Architecture is not optional for multi-agent teams. It is the difference between a system that runs autonomously and one that creates more work than it eliminates.
After studying hundreds of production OpenClaw deployments and community reports, five distinct patterns have emerged. Each one has clear trade-offs, and the right choice depends entirely on your team size, task complexity, and reliability requirements.
Pattern 1: Hub-and-Spoke
The most popular architecture in the OpenClaw community, and for good reason. One central agent (the PM) coordinates all other agents. Every task flows through the hub. Specialists never talk to each other directly. They report to the PM, and the PM delegates to the next agent in the chain.
@radar (SEO)
↑↓
@metrics (Data) ←→ @orion (PM) ←→ @echo (Writer)
↑↓
@pulse (Social)
All communication flows through @orion.
Specialists never message each other directly.# Content Team - Hub-and-Spoke
## Agents
- @orion: Project Manager (HUB). Coordinates all tasks. Only agent that delegates.
- @radar: SEO Analyst. Reports only to @orion.
- @echo: Content Writer. Reports only to @orion.
- @pulse: Social Media Manager. Reports only to @orion.
- @metrics: Data Analyst. Reports only to @orion.
## Workflow
1. @orion receives the request and creates a task plan
2. @orion delegates research to @radar
3. @radar completes and reports back to @orion
4. @orion delegates writing to @echo with @radar's research
5. @echo completes and reports back to @orion
6. @orion delegates distribution to @pulse
7. @orion delegates tracking to @metrics
## Rules
- ONLY @orion may delegate tasks using @mentions
- Specialist agents respond ONLY to @orion
- If a specialist needs input from another specialist, ask @orion to request it
- No agent-to-agent communication without @orion as intermediaryBest for
Small to medium teams (3-7 agents), content pipelines, any workflow where you want clear accountability and easy debugging. This is the default choice if you are unsure.
Strengths
Simple to understand. Easy to debug because all communication flows through one point. The PM has complete visibility into every task. No risk of agents triggering loops or duplicating work.
Weaknesses
The PM becomes a bottleneck at scale. If 5 agents all need delegation simultaneously, they queue up waiting for the PM. Also, the PM agent consumes the most tokens because it processes every handoff message.
Pattern 2: Pipeline
Agent A's output becomes Agent B's input, which becomes Agent C's input. Work flows in one direction through a chain of specialists. There is no central coordinator. Each agent knows who comes before it and who comes after it.
@radar → @echo → @editor → @pulse → @metrics
(SEO) (Write) (Edit) (Social) (Track)
Work flows left to right. Each agent processes
and passes to the next. No backward flow.# Content Pipeline
## Agents
- @radar: SEO Analyst. Entry point. Receives topic, outputs keyword brief.
- @echo: Content Writer. Receives keyword brief, outputs draft article.
- @editor: Editor. Receives draft, outputs polished article.
- @pulse: Social Media. Receives final article, outputs social posts.
- @metrics: Tracker. Receives published URLs, monitors performance.
## Workflow
1. @radar receives the topic and performs keyword research
2. @radar sends keyword brief to @echo
3. @echo writes the article and sends draft to @editor
4. @editor reviews, polishes, and sends final version to @pulse
5. @pulse creates social posts and sends published URLs to @metrics
6. @metrics begins tracking and reports weekly
## Rules
- Each agent passes output to the NEXT agent only
- No backward delegation (no sending work back up the pipeline)
- If an agent encounters an error, it flags it and passes anyway
- @radar is the only entry point for new tasksBest for
Content pipelines, DevOps workflows (build, test, deploy, monitor), data processing chains, any sequential workflow where each step transforms the output of the previous step.
Strengths
Very predictable execution order. Low communication overhead because each agent only talks to its neighbors. Easy to add or remove stages by inserting or removing an agent from the chain. No coordinator bottleneck.
Weaknesses
No error recovery. If the editor finds a fundamental problem with the draft, there is no way to send it back to the writer without breaking the pipeline pattern. Also, the whole pipeline stalls if any single agent is slow.
Pattern 3: Peer-to-Peer
Every agent can talk to every other agent. There is no hierarchy, no coordinator, no fixed workflow. Agents @mention whoever they need based on the current task. This is the most flexible pattern and also the most chaotic.
@radar ←→ @echo
↑↕↘ ↗↕↓
@metrics ←→ @pulse
Every agent can message every other agent.
No central coordinator. Fully decentralized.# Research Team - Peer-to-Peer
## Agents
- @radar: SEO Analyst. Can request data from @metrics, briefs to @echo.
- @echo: Content Writer. Can request keywords from @radar, data from @metrics.
- @pulse: Social Media. Can request content from @echo, trends from @radar.
- @metrics: Data Analyst. Can request context from any agent.
## Workflow
- No fixed workflow. Agents collaborate as needed.
- Any agent can initiate a request to any other agent.
- Agents self-organize around incoming tasks.
## Rules
- Maximum 3 @mentions per response (prevents cascade loops)
- Each agent must include a TASK_ID when delegating to track threads
- If an agent receives conflicting requests, most recent takes priority
- All agents log their actions to the shared workspaceBest for
Research teams, brainstorming, exploratory work where the workflow is not predictable in advance. Works well when agents need to ask each other questions dynamically.
Strengths
Maximum flexibility. Agents can self-organize around novel tasks. No bottleneck because there is no single coordinator. Great for creative and research workflows where the path is not linear.
Weaknesses
Can spiral into chaos. Without a coordinator, agents may trigger each other in loops, duplicate work, or conflict on tasks. Debugging is difficult because there is no single point of visibility. Requires strict rules to prevent runaway token consumption.
A word of caution: the $280 weekend horror story from Reddit was a peer-to-peer setup gone wrong. The agents kept triggering each other in feedback loops, each response generating 2-3 more @mentions, which generated more responses. Without the 3-mention limit rule, peer-to-peer can spiral fast. If you use this pattern, implement strict loop prevention.
Pattern 4: Specialist Pools
Multiple agents share the same role and are load balanced. Instead of one writer, you have three writers. A dispatcher routes incoming work to whichever writer is available. This pattern handles high volume workloads where a single agent per role would create a bottleneck.
┌→ @writer-1
@orion (Dispatcher) ├→ @writer-2 → @editor → @publisher
└→ @writer-3
Three writers share the load. The dispatcher
assigns based on availability.# High-Volume Content Team - Specialist Pool
## Agents
- @orion: Dispatcher. Routes writing tasks to available writers.
- @writer-1: Content Writer (pool member 1). Blog posts and articles.
- @writer-2: Content Writer (pool member 2). Blog posts and articles.
- @writer-3: Content Writer (pool member 3). Blog posts and articles.
- @editor: Editor. Reviews all drafts from the writer pool.
- @publisher: Publisher. Formats and publishes approved articles.
## Workflow
1. @orion receives batch of content requests
2. @orion assigns one article to each writer in round-robin order
3. Each writer completes their article and sends to @editor
4. @editor reviews and sends approved articles to @publisher
5. @publisher formats and publishes
## Rules
- @orion tracks which writer is busy and routes to idle writers
- All writers share the same SOUL.md (identical configuration)
- @editor processes articles in the order received
- If a writer is busy, @orion queues the task and assigns when available
- Maximum 2 articles per writer at any timeBest for
High-volume content production, batch processing, any workflow where a single agent per role cannot keep up with demand. Agencies producing 20+ articles per week benefit from writer pools.
Strengths
Scales horizontally. Adding more agents to the pool increases throughput linearly. One agent failing does not stop the whole pipeline because others in the pool continue working. Great for batch operations.
Weaknesses
More complex to set up and maintain. The dispatcher needs logic to track agent availability. All pool members must have identical configurations to ensure consistent output quality. Higher base cost because you are paying for multiple agents in the same role.
Pattern 5: Hierarchical
For large teams of 10 or more agents, a single coordinator cannot manage everything. The hierarchical pattern introduces team leads who manage sub-teams. The top-level PM coordinates team leads, and each team lead coordinates their specialists. This is the enterprise pattern.
@ceo (Executive PM)
/ | \
@content-lead @dev-lead @marketing-lead
/ | | \ | \
@writer @editor @frontend @backend @seo @social
3 levels. CEO manages leads. Leads manage specialists.
Specialists never talk to CEO directly.# Enterprise Team - Hierarchical
## Level 1: Executive
- @ceo: Executive PM. Coordinates team leads only. Never delegates to specialists directly.
## Level 2: Team Leads
- @content-lead: Content Team Lead. Manages @writer and @editor.
- @dev-lead: Dev Team Lead. Manages @frontend and @backend.
- @marketing-lead: Marketing Team Lead. Manages @seo and @social.
## Level 3: Specialists
- @writer: Content Writer. Reports to @content-lead only.
- @editor: Editor. Reports to @content-lead only.
- @frontend: Frontend Developer. Reports to @dev-lead only.
- @backend: Backend Developer. Reports to @dev-lead only.
- @seo: SEO Analyst. Reports to @marketing-lead only.
- @social: Social Media Manager. Reports to @marketing-lead only.
## Workflow
1. @ceo receives a project and breaks it into team-level tasks
2. @ceo delegates to relevant team leads
3. Each team lead breaks their task into specialist assignments
4. Specialists complete work and report to their team lead
5. Team leads compile results and report to @ceo
6. @ceo synthesizes the final output
## Rules
- Specialists communicate ONLY with their team lead
- Team leads communicate ONLY with @ceo and their own specialists
- Cross-team requests go through @ceo (e.g., content needing dev work)
- Each team lead sends a daily summary to @ceoBest for
Large teams (10+ agents), enterprise projects, cross-functional workflows that require multiple disciplines. Agencies, product teams, and organizations with complex delivery processes.
Strengths
Scales to any team size. Each team lead only manages 2-4 agents, keeping coordination manageable. Clear chain of command makes debugging straightforward. Cross-team dependencies are handled by the executive PM.
Weaknesses
Highest setup complexity. Requires 2-3 levels of AGENTS.md configuration. More latency because messages traverse multiple levels. The executive PM and team leads consume significant tokens for coordination overhead.
When to Use Which Architecture
| Pattern | Team Size | Complexity | Best Use Case | Avoid When |
|---|---|---|---|---|
| Hub-and-Spoke | 3-7 | Low | General purpose, first teams | High parallelism needed |
| Pipeline | 3-6 | Low | Sequential workflows | Tasks need iteration/revision |
| Peer-to-Peer | 2-5 | High | Research, brainstorming | Predictable output needed |
| Specialist Pool | 5-10 | Medium | High volume batch work | Each task is unique |
| Hierarchical | 10+ | High | Enterprise, cross-functional | Small teams, speed matters |
Common Architecture Mistakes
Too many agents too soon
Starting with 7 agents when 3 would do. Each additional agent adds coordination overhead, token costs, and debugging complexity. Start with the minimum viable team and add agents only when you have a clear bottleneck that justifies the new role.
Overlapping responsibilities
Two agents that both handle 'analysis' will fight over tasks, produce duplicate work, and confuse the coordinator. Every agent must have a clearly scoped role with no ambiguity. If two agents could reasonably handle the same task, your team structure needs refining.
No coordinator agent
Skipping the PM role to save costs and hoping agents will self-organize. In practice, without a coordinator, agents miss tasks, duplicate work, and have no way to resolve conflicts. The PM agent pays for itself by preventing wasted work across the team.
Ignoring loop prevention
In peer-to-peer and hub-and-spoke setups, agents can trigger infinite loops. Agent A mentions Agent B, which mentions Agent A, which mentions Agent B. Always set a maximum @mention limit per response and include explicit loop-breaking rules in AGENTS.md.
Wrong architecture for the task
Using a pipeline for creative brainstorming (where iteration is needed) or peer-to-peer for a predictable content workflow (where order matters). Match the pattern to the nature of the work, not the other way around.
Frequently Asked Questions
Which architecture should I start with if I am new to OpenClaw?
Start with hub-and-spoke. It is the simplest to set up, easiest to debug, and handles most use cases well. Your PM agent sits in the center and coordinates all work. Start with 3 agents (PM plus 2 specialists), get comfortable with the @mention routing, then expand from there. Most teams never need to move beyond hub-and-spoke.
Can I combine multiple architectures in one team?
Yes, and many production teams do exactly this. A common pattern is hub-and-spoke at the top level with a pipeline embedded inside it. The PM coordinates the overall workflow using hub-and-spoke, but the content creation subprocess runs as a pipeline (researcher to writer to editor). Think of architectures as patterns you can nest and combine rather than rigid frameworks.
How many agents is too many for a single team?
For hub-and-spoke and pipeline architectures, 7 agents is a practical upper limit. Beyond that, the coordinator agent struggles to track all the moving pieces and context windows get too large. For peer-to-peer, even 5 agents can get chaotic. If you need more than 7 agents, switch to the hierarchical pattern where team leads manage sub-teams of 3-4 agents each.
What is the most common mistake when setting up multi-agent teams?
Creating agents with overlapping responsibilities. When two agents both think they handle 'content,' they fight over tasks, produce duplicate work, and confuse the coordinator. Every agent must have a clearly defined scope in AGENTS.md with no gray areas. If you are unsure whether a task belongs to Agent A or Agent B, your team structure needs refining.
How does AGENTS.md differ from SOUL.md?
SOUL.md defines who an individual agent is: its personality, capabilities, rules, and model. AGENTS.md defines how the team works together: which agents exist, their roles in the team, the workflow sequence, and coordination rules. Think of SOUL.md as a job description and AGENTS.md as the team playbook. Every agent has its own SOUL.md, but the team shares one AGENTS.md.
Do I need a PM agent in every architecture?
Not necessarily, but it is strongly recommended for any team with 3 or more agents. In the pipeline pattern, the first and last agents can handle coordination without a dedicated PM. In peer-to-peer, there is no central coordinator by design. However, without a PM, debugging issues becomes much harder because there is no single agent with a complete view of the workflow. For production teams, always include a coordinator.
Skip the setup pain
190+ pre-configured agent templates with tested configs, Docker setup, and deploy packages. See our use cases for pre-built team architectures ready to deploy.
Browse Agent Templates →Deploy a Ready-Made AI Agent
Skip the setup. Pick a template and deploy in 60 seconds.