# OpenClaw Sub-Agents: How to Run Parallel AI Workflows

Canonical: https://clawdocx.com/blog/openclaw-sub-agents-explained
Author: Sam Okafor
Published: 2026-02-24
Updated: 2026-02-24

> Stop doing things one at a time. Sub-agents let your OpenClaw run multiple tasks simultaneously—here's how they work.

# OpenClaw Sub-Agents: How to Run Parallel AI Workflows

What if your AI agent could clone itself?

That's essentially what sub-agents do. Instead of handling one task at a time, OpenClaw can spawn multiple agents to work in parallel—then combine their results.

## The Problem: Sequential Bottlenecks

Without sub-agents, your agent works like this:

1. Research competitor A → 5 minutes
2. Research competitor B → 5 minutes
3. Research competitor C → 5 minutes
4. Compare all three → 3 minutes

**Total: 18 minutes**

With sub-agents:

1. Research A, B, and C simultaneously → 5 minutes
2. Compare all three → 3 minutes

**Total: 8 minutes**

That's a 55% reduction. And the savings scale with task complexity.

## How Sub-Agents Work

Your main agent acts as an **orchestrator**. It:

1. **Decomposes** a complex task into smaller pieces
2. **Spawns** sub-agents to handle each piece
3. **Monitors** their progress
4. **Combines** their results into a final output

Each sub-agent runs in its own isolated session with its own context and model.

## Real Example: Parallel Code Review

Say you want a thorough PR review. Instead of one agent checking everything sequentially, spawn four specialists:

- **Security Agent:** Check for vulnerabilities, injection risks, secrets
- **Performance Agent:** Look for N+1 queries, missing indexes, bottlenecks
- **Style Agent:** Naming conventions, code organization, readability
- **Test Agent:** Coverage gaps, edge cases, missing assertions

Each runs for 2 minutes. You get a comprehensive review in 2 minutes instead of 8.

## When to Use Sub-Agents

**Great for:**
- Research across multiple sources
- Code review (security, performance, style in parallel)
- Bulk content generation
- Multi-step analysis with independent steps
- Competitive analysis

**Not worth it for:**
- Simple questions
- Tasks that need sequential context
- Quick one-off requests
- When budget is very tight

## Cost Considerations

Sub-agents multiply your API usage. Smart model selection keeps costs down:

| Task Type | Recommended Model | Cost/Task |
|-----------|-------------------|-----------|
| Research | Sonnet 4.6 | ~$0.15 |
| Coding | GPT-5.3-Codex | ~$0.25 |
| Summarization | Haiku 4 | ~$0.03 |
| Complex reasoning | Opus 4.6 | ~$0.75 |

**Rule of thumb:** Use the cheapest model that can handle the task. Haiku 4 for summaries, Sonnet 4.6 for most work, Opus 4.6 only for complex reasoning.

## Getting Started

The simplest sub-agent spawn:

```
"Break this research into 3 parallel tasks and use sub-agents to research them simultaneously."
```

OpenClaw handles the rest—spawning agents, monitoring progress, and combining results.

For more control, you can specify models, timeouts, and cleanup strategies in your prompts.

## What's Next?

Sub-agents are one of OpenClaw's most powerful features, but they take practice to use effectively. Common mistakes include over-parallelizing (spawning too many agents) and using expensive models for simple tasks.

**Want the complete playbook?** Our premium Sub-Agents Mastery guide covers fan-out patterns, error handling, cost optimization, and 5 real-world workflow templates you can copy today.

---

*What would you parallelize first?*