sub-agents

What This Skill Does

The sub-agents skill is the primary orchestration tool for OpenClaw, enabling the spawning of isolated, parallel execution sessions. By delegating complex or time-consuming tasks to specialized agents, you can drastically improve system throughput and accuracy. Each sub-agent is essentially a clean-slate worker, tailored by your task instructions and model selection to complete specific objectives without being bogged down by the parent agent's state.

Installation

To integrate this orchestration capability into your workflow, run the following command in your OpenClaw terminal: clawhub install openclaw/skills/skills/bill492/sub-agents

Use Cases

The skill is ideal for multi-threaded workflows where blocking the main agent is inefficient. Common use cases include:

• Parallel research: Spawning one agent to analyze documentation while another verifies code syntax.
• Complex Model Routing: Assigning a high-thinking Opus model to debug a critical logic error while offloading routine writing tasks to a faster GPT-5.2 sub-agent.
• Sandboxed Operations: Running untrusted or experimental code in a separate session to prevent contamination of the primary working environment.
• Long-running tasks: Delegating operations expected to take over 10 seconds to ensure the main interface remains responsive for user interaction.

Example Prompts

1. "Spawn a sub-agent using the codex model to refactor the entire auth module in src/auth.ts and run the unit tests when finished."
2. "I need to research the latest trends in serverless architecture. Please delegate this to a gpt5 sub-agent and summarize the findings in a new file named research_report.md."
3. "Run a parallel analysis of these two data files. Use an opus agent for deep reasoning and provide a comparative summary once both are complete."

Tips & Limitations

• Context Blindness: Remember that sub-agents do not inherit your environment. You must explicitly include all necessary file paths or inline attachments. If a sub-agent doesn't know where to look, it cannot act.
• The 10-Second Rule: If a task takes less than 10 seconds, perform it yourself. The overhead of spawning and coordinating sub-agents should not exceed the time saved.
• Model Optimization: Use gpt5 (GPT-5.2) for the vast majority of general tasks to leverage unlimited usage on the Codex subscription. Save opus for complex reasoning and sonnet for massive context window requirements.
• Cleanup: Use the cleanup: 'delete' parameter for ephemeral tasks to keep your environment organized and logs clean, but set to keep if you need to audit the agent's work logs after completion.