Building a single AI agent is a great start, but the real magic happens when you get multiple agents talking to each other. Instead of one overworked bot trying to do everything, you create a specialized team where each member has a specific job. You might have one agent that just does research, another that handles the writing, and a third that checks for errors. It’s like running a small company where no one ever sleeps and everyone follows instructions perfectly. This guide covers how to set up multi-agent pipelines in OpenClaw so your bots can collaborate autonomously while you focus on higher-level work.
What does a multi-agent setup actually look like?
In OpenClaw, every agent is its own separate process. They don’t share a brain, but they can share a folder. The simplest way to make them work together is through “shared memory.” One agent writes a Markdown file to a specific directory, and the next agent in the line reads that file and picks up where the first one left off. It’s a reliable, transparent way to build a pipeline without needing complex message queues or fancy databases.
The Scenario: You’re running a tech blog as a side hustle. You have a “Researcher” agent that scans Hacker News every night at midnight. It saves a list of the top three stories to a shared folder. At 1:00 AM, your “Writer” agent wakes up, reads that list, and drafts three articles. By the time you wake up, the work is done.
How do I create a specialized team of agents?
You start by spinning up multiple agents using the CLI. Give them names that reflect their jobs—don’t just call them agent1 and agent2. Once you have the folders created, you’ll need to set up a shared directory on your hard drive where they can exchange files. This folder becomes the “meeting room” where they pass notes back and forth.
openclaw agent create --name researcher
openclaw agent create --name writer
mkdir -p ~/.openclaw/shared/pipelineIn each agent’s config.yaml, you point them to this shared path. One agent gets write access, and the other gets read or readwrite access. It’s simple, effective, and easy to debug because you can see the files yourself.
The Scenario: You’re building a customer support team. Your “Triage” agent reads incoming emails and marks them as “Urgent” or “General.” Your “Support” agent only looks for the “Urgent” files and handles those first. It keeps your response times fast without you having to manually sort through the junk.
How do agents share information with each other?
Shared memory is great for scheduled tasks, but it requires some planning. You have to stagger their schedules so they don’t try to read and write the same file at the exact same time. If your Researcher runs at 10:00 PM, give your Writer at least thirty minutes to make sure the research is actually finished before it starts its own task.
- Researcher:
schedule: "0 22 * * *" - Writer:
schedule: "30 22 * * *"
This “staggered heartbeat” approach is the backbone of most OpenClaw pipelines. It’s low-tech, but it’s incredibly hard to break. If something fails, you just look in the shared folder to see where the chain stopped.
The Scenario: Your Researcher hits an API limit and fails to write the brief. Your Writer wakes up, sees the brief is missing or outdated, and decides to wait until tomorrow. No crashes, no infinite loops, just a quiet night for your CPU.
Can agents talk directly using ACP?
If you need your agents to have a real-time conversation, you use the Agent Communication Protocol (ACP). This lets one running agent send a message directly to another and wait for a reply. It’s more powerful than shared memory but also a bit more complex to set up. It’s perfect for tasks that require a back-and-forth, like a “Manager” agent asking a “Coder” agent to fix a bug and then verifying the fix.
- action: acp_send
target: coder_agent
message: "Fix the syntax error in main.js"
await_response: trueThe latest version of OpenClaw even adds cryptographic signatures to these messages. This means your agents can verify that a message actually came from a trusted teammate and wasn’t spoofed by something else.
The Scenario: You’re building a complex coding assistant. Your “Architect” agent creates a plan and sends it to the “Developer” agent. The Developer finds a flaw in the plan and messages the Architect back to ask for a change. They hash it out in seconds while you’re busy doing something else.
How do I start my entire fleet at once?
Running three separate terminal windows for three agents is a pain. OpenClaw has a fleet command that handles everything for you. It starts all your agents simultaneously and gives you a single dashboard where you can see all their activity at once. It’s the easiest way to manage a complex setup without losing track of who is doing what.
openclaw fleet startYou can see which agent is thinking, which one is waiting, and which one just finished a task. It turns your terminal into a command center for your AI team.
The Scenario: You’ve built a massive five-agent pipeline for market research. Instead of spending ten minutes starting each one manually every morning, you just run
fleet startand walk away. Your dashboard fills up with activity logs, and you can see the whole team working in perfect sync.
What are the common pitfalls of multi-agent workflows?
The biggest mistake is making your agents too chatty. If Agent A sends a message to Agent B, and Agent B responds with a question, you can end up in an infinite loop that burns through your API credits in minutes. Always set a “stop condition” or a max number of turns for any direct conversation.
- Stale Data: Make sure the downstream agent checks the file’s timestamp.
- Conflicts: Don’t let two agents write to the same file at once.
- Vague Prompts: Be clear about who is in charge of the final decision.
The Scenario: You set up a “Debate” team of two agents to brainstorm ideas. They start arguing about a minor detail and send fifty messages back and forth while you’re at lunch. You come back to a $20 API bill and no actual ideas. Always set a limit on their conversation length.
What to Read Next
- Connect to messaging: OpenClaw Integrations: WhatsApp, Slack, and Discord
- Secure your memory: How OpenClaw Memory and Privacy Work
- Install OpenClaw: How to Install OpenClaw on Ubuntu, macOS, and Windows