If you use OpenClaw for more than a few minutes a day, you have probably felt the bottleneck: one task at a time, and everything else waits. You ask your agent to research a topic, and while it works, you just sit and watch the logs. It does not have to be that way. Sub-agents let your main agent spin up separate, isolated sessions that do real work in the background, so you can keep moving while a research sweep, a draft, or a content audit runs hands-off.
In this guide, I will show you what OpenClaw sub-agents are, how to set one up in about ten minutes, and the exact tasks I delegate to them every week. You will also get the common mistakes list so you do not waste tokens learning them the hard way. No coding required. If you are new to OpenClaw, start with our OpenClaw setup guide first, then come back here.
What Are OpenClaw Sub-Agents?
A sub-agent is a child session that your main OpenClaw session spawns to complete a specific job. It gets its own context window, its own task prompt, and its own model if you choose one. It runs independently, and when it finishes, the result is pushed back to you. You do not babysit it.
Think of it like a small team: your main agent is the project lead, and each sub-agent is a contractor who takes one brief and comes back with the finished work. The lead keeps talking to you, answering your questions, and moving the project forward. The contractors work in parallel and deliver when they are done.
Why this matters
The biggest reason to use sub-agents is that they stop you from being the bottleneck. Instead of running one long task in your main session and waiting, you delegate it and keep going. There are three practical benefits:
- Parallel work. You can research, draft, and audit at the same time instead of one after another.
- Context isolation. A messy or long-running task does not bloat your main session’s context, which keeps the main agent fast and focused. We wrote a whole post about why memory and context hygiene matter.
- Cost control. You can send routine background work to a cheaper model while your main session keeps the powerful one. This is the same trick behind our token cost fix guide.
Before You Start: What You Need
Setting up sub-agents does not require coding, but a few things need to be in place first:
- A working OpenClaw install. If OpenClaw is not running yet, follow the 30 minute setup guide.
- A task that can run without your input. Sub-agents work best for jobs with a clear brief and a clear output. “Look into this and get back to me” works. “Help me think through my life decisions” does not.
- Tool access (optional). If the task needs the web, file access, or API calls, make sure those tools are enabled for sub-agent sessions.
- API keys (optional). Tasks that call external services need their keys configured in OpenClaw. Routine text tasks need nothing extra.
That is it. You do not need a separate machine, a second account, or any special plan. Sub-agents are built into OpenClaw.
How to Spawn a Sub-Agent: Step by Step
Here is the exact flow I use to delegate work to a sub-agent. It takes about ten minutes the first time, and about thirty seconds after that.
Step 1: Write a clear task brief
The quality of the result depends almost entirely on the brief. A good brief includes the context, the goal, the output format, and where to save the result. For example:
“You are a research assistant. Fetch these five URLs and summarize each in 3 bullet points covering key claims, data points, and limitations. Save the summary to research-summary.md in the workspace. Return a 2 sentence overview when done.”
Vague briefs produce vague work. Be specific about the format and the deliverable.
Step 2: Choose the mode
OpenClaw offers two sub-agent modes:
- Run mode (one-shot). The sub-agent does the job and exits. Results come back to you once. This is what you want for research, drafts, and audits.
- Session mode (persistent). The sub-agent stays alive in its own thread and can receive follow-up messages. Use this when the task needs multiple rounds, like a long content series.
Start with run mode. It covers 90 percent of background work.
Step 3: Choose the context
By default, a sub-agent starts with a clean, isolated context. That is usually the right call. Only fork the current conversation into the sub-agent when it genuinely needs the full transcript to do its job, because forking copies your whole context and costs tokens.
Step 4: Pick a model
For routine background work, assign a cheaper or faster model. Drafting, summarizing, and formatting do not need the flagship model. Save the expensive one for complex reasoning in your main session. This is the single easiest cost lever you have.
Step 5: Launch and let it run
Start the sub-agent and end your turn. OpenClaw pushes the result back to you when the sub-agent finishes, so you do not need to poll or check in. Go do something else. That is the whole point.
Step 6: Collect results and clean up
When the result arrives, check the deliverable file, then clean up the finished session. Keep the sessions list tidy so you always know what is running and what is done.
Three Background Tasks That Work Great as Sub-Agents
To make this concrete, here are three tasks I run hands-off every week.
1. Batch research sweeps
I hand a sub-agent a list of URLs or a research question, and it returns a structured summary file. While it reads, I keep writing or answering messages. One sub-agent can chew through a dozen sources in the time it takes me to draft an outline.
2. Draft generation
I give the sub-agent an outline, a word count, and a tone note, and it produces a full draft saved to a file. I review and edit later. The draft does not need to be perfect, it needs to exist so my editing time is short.
3. Content audits
Sub-agents are great for repetitive checks: fetch published posts, verify links, flag missing images, or check that internal links point to real URLs. It is tedious work for a human and perfect for a background agent. This is also how I keep the site’s internal linking healthy, the same way the multi-agent orchestration guide describes building bigger agent teams.
Common Mistakes (and How to Avoid Them)
I have made all of these mistakes. Here is how to skip them:
- Forking context when you do not need it. Forking copies the entire conversation into the sub-agent and burns tokens. Start isolated and fork only when the task genuinely depends on the transcript.
- Writing vague briefs. “Do some research” returns a mess. Specify the sources, the output format, and the file to save to.
- Spawning sub-agents for quick lookups. A fast answer that takes one tool call should stay in your main session. Sub-agents add overhead; use them for real jobs, not trivia.
- Polling instead of letting results push. You do not need to keep checking on a sub-agent. Let it deliver. Polling wastes your time and tokens.
- Running expensive models on routine work. If the task is drafting or summarizing, assign a cheap model. Keep the flagship model for complex reasoning.
- Skipping cleanup. Finished sessions pile up and clutter your workspace. Delete or archive them when the deliverable is in hand.
Check out our OpenClaw vs n8n comparison, see which platform keeps your automations running without breaking every week.
FAQ
No. Sub-agents are launched with plain language task briefs, not code. You write what you want done, choose the mode, and OpenClaw handles the rest.
Sub-agents use the same token-based billing as your main session. There is no per-sub-agent fee, but you do pay for the tokens they consume. Assigning a cheaper model to background tasks keeps the cost low.
Run mode is a one-shot job: the sub-agent completes the task and delivers results once. Session mode keeps the sub-agent alive in its own thread so you can send follow-up messages and iterate. Start with run mode for most work.
Yes. Sub-agents can access the tools and services you have configured in OpenClaw, including web search, file access, and external APIs. Make sure the required tools are enabled for sub-agent sessions before you launch.
You can run multiple sub-agents in parallel. Each one gets its own isolated context, so they do not interfere with each other or with your main session. Just keep the task briefs clear and clean up finished sessions.
Final Thoughts
Sub-agents are the difference between an AI assistant you watch and an AI assistant that works while you do. The setup is simple: write a clear brief, pick run mode, isolate the context, assign a cheap model, and let the results push back to you. Start with one batch research task today, and once you see how effortless it is, you will wonder why you ever ran long tasks in your main session at all.
If you want to go further, our multi-agent orchestration guide covers how to combine sub-agents into larger teams, and the complete OpenClaw guide has everything else you need to master the platform.
About The Author
