OpenClaw Agent-to-Agent Communication: How AI Agents Collaborate
A single AI agent is useful. A team of agents that talk to each other is powerful. This guide explains how OpenClaw agents communicate, delegate tasks, and collaborate — and how to configure your own multi-agent team.
Why Agents Need to Talk to Each Other
A content writer agent can produce blog posts, but it cannot optimize them for search engines. An SEO agent can analyze keywords, but it cannot write compelling prose. When these agents communicate, the writer can ask the SEO agent for keyword suggestions before drafting, and the SEO agent can review the draft for optimization opportunities. The result is better than what either agent produces alone.
OpenClaw enables this through a message-passing system. Agents do not share memory or state directly. Instead, they send text messages to each other through the gateway, just like team members communicating via chat. This keeps agents independent and modular — you can add, remove, or replace any agent without affecting the others.
How Agent Communication Works
The communication flow follows a simple pattern: request, process, respond. Here is what happens when one agent delegates a task to another:
1. User sends "Write a blog post about AI agents" to Coordinator
2. Coordinator reads agents.md → knows Writer and SEO agents exist
3. Coordinator delegates "Research top keywords for AI agents" to SEO Agent
4. SEO Agent processes the request → returns keyword list
5. Coordinator delegates "Write 1500-word post using these keywords: ..." to Writer
6. Writer produces the draft → returns to Coordinator
7. Coordinator combines results → sends final output to UserThe gateway handles all message routing. Agents never connect to each other directly.
Setting Up agents.md
The agents.md file is the key configuration that enables communication. It tells each agent who their teammates are and what each teammate can do. Place this file in the agent's workspace directory alongside SOUL.md.
agents.md for a Coordinator Agent
# Team Members
## Echo (Content Writer)
- Role: Writes blog posts, articles, social media content
- Strengths: Creative writing, storytelling, adapting tone
- When to delegate: Any content creation task
## Radar (SEO Analyst)
- Role: Keyword research, content optimization, competitor analysis
- Strengths: Data analysis, search trends, technical SEO
- When to delegate: Keyword research, SEO audits, content scoringagents.md for a Specialist Agent
# Team Members
## Orion (Project Manager / Coordinator)
- Role: Receives tasks from the user, breaks them down, coordinates team
- When to escalate: If the task is unclear or requires multiple agents
- Note: Report completed work back to OrionSpecialist agents typically only need to know about the coordinator. The coordinator knows about everyone and handles task routing.
Three Communication Patterns
1. Coordinator Pattern (Recommended)
One central agent receives all tasks and delegates to specialists. This is the simplest and most reliable pattern. The coordinator acts as a project manager — it understands the big picture and assigns subtasks to the right agent.
User → Coordinator → Writer Agent
↘ SEO Agent
↘ Research AgentBest for: Most use cases. Content teams, research workflows, customer support systems.
2. Pipeline Pattern
Tasks flow through agents in a fixed sequence. Agent A processes first, passes to Agent B, which passes to Agent C. Each agent adds value to the output before passing it along. No coordinator is needed because the order is predetermined.
User → Research Agent → Writer Agent → Editor Agent → UserBest for: Content production, data processing, sequential workflows with clear stages.
3. Peer-to-Peer Pattern
Every agent knows about every other agent and can delegate freely. This is the most flexible but hardest to control. Without clear rules in SOUL.md, agents can create delegation loops where tasks bounce between agents endlessly.
Agent A ↔ Agent B
Agent A ↔ Agent C
Agent B ↔ Agent CBest for: Small teams (2-3 agents) with well-defined roles and strong boundary rules.
Complete Setup Example: Content Team
Here is a complete three-agent team setup with a coordinator, a writer, and an SEO analyst. Copy these commands to set up the team from scratch.
# Create workspaces
mkdir -p agents/orion agents/echo agents/radar
# Register all three agents
openclaw agents add orion --workspace ./agents/orion --non-interactive
openclaw agents add echo --workspace ./agents/echo --non-interactive
openclaw agents add radar --workspace ./agents/radar --non-interactive
# Start the gateway
openclaw gateway start
# Send a task to the coordinator
openclaw agent --agent orion --message "Write an SEO-optimized blog post about AI automation"Orion (the coordinator) will break this task down, ask Radar for keyword research, delegate writing to Echo with the keyword data, and return the finished blog post. For ready-to-use agent configurations, use the CrewClaw SOUL.md Generator to build a complete team workspace.
Best Practices for Agent Communication
Define clear role boundaries
Each agent should have a distinct specialty. Overlapping roles cause confusion about who handles what. Write specific "When to delegate" rules in agents.md so agents know exactly when to ask for help.
Keep agent teams small
Start with 2-3 agents and add more only when you identify a clear gap. Each agent adds token consumption and coordination overhead. A focused three-agent team outperforms a sprawling ten-agent team in most scenarios.
Add anti-loop rules to SOUL.md
Include explicit rules like "Never delegate a task back to the agent that sent it to you" and "If you receive a delegated task, complete it yourself rather than re-delegating." This prevents infinite delegation chains.
Test with simple tasks first
Before running complex multi-step workflows, test each agent individually with direct messages. Then test a simple two-agent delegation before scaling to the full team.
Frequently Asked Questions
How do OpenClaw agents communicate with each other?
OpenClaw agents communicate through the gateway, which acts as a message router. When Agent A needs help from Agent B, it sends a delegation request through the gateway. The gateway delivers the message to Agent B, collects the response, and returns it to Agent A. All communication is text-based and asynchronous. You configure which agents know about each other through the agents.md file in each agent's workspace.
What is agents.md and why do I need it?
agents.md is a configuration file that tells an agent which other agents exist and what they do. It lives in each agent's workspace directory alongside SOUL.md. Without agents.md, an agent operates in isolation and cannot request help from other agents. A typical agents.md lists each teammate's name, role, and capabilities so the agent knows who to delegate to for specific tasks.
How many agents can communicate with each other?
There is no hard limit on the number of agents in an OpenClaw team. Most practical setups use 2-5 agents because each additional agent adds coordination overhead. A three-agent team (for example: a coordinator, a writer, and a reviewer) handles most workflows effectively. If you need more specialized roles, you can add them, but keep in mind that every agent consumes LLM tokens when processing messages.
Can agents from different model providers talk to each other?
Yes. Agent communication is handled by the gateway at the text level, so the underlying model does not matter. You can have a Claude-powered coordinator delegating tasks to a GPT-powered writer and an Ollama-powered researcher. Each agent processes messages independently using its own model provider. This lets you optimize cost and quality per role.
What is the coordinator pattern?
The coordinator pattern uses one central agent (typically called a project manager or coordinator) to receive all incoming tasks, break them down, and delegate subtasks to specialized agents. The coordinator collects results, combines them, and returns a final output. This is the most common multi-agent pattern in OpenClaw because it simplifies routing and prevents agents from sending tasks in circles.
How do I prevent agents from creating infinite loops?
OpenClaw's gateway includes built-in loop detection that limits the number of delegation hops per conversation. You can also set explicit rules in SOUL.md like 'Never delegate a task back to the agent that delegated it to you' and 'Complete tasks yourself if you can — only delegate when the task clearly falls outside your expertise.' Clear role boundaries in agents.md also prevent circular delegation.
Deploy a Ready-Made AI Agent Team
CrewClaw offers pre-configured team packages with coordinator, writer, and analyst agents — complete with SOUL.md, agents.md, and Docker deployment files.