Sub-Agents: Stop Getting Slop From Your AI

Overview

Sub-agents are one of the most vital features in OpenClaw, dramatically improving work quality and reducing "slop" output. By distributing tasks across parallel agents with focused contexts, you achieve better results, faster execution, and more reliable outputs.

Why Sub-Agents Matter

The Problem with Single-Agent Workflows

When one agent handles everything:

• Context overload - Agent tries to remember too much
• Rushed execution - Pressure to deliver results quickly leads to shortcuts
• Generic output - Lack of specialization produces boilerplate responses
• Higher error rates - No cross-checking or validation

The Sub-Agent Solution

Massive improvement in quality - Specialized agents produce focused, accurate work
Parallel execution - Multiple tasks run simultaneously, saving time
Context optimization - Each sub-agent has minimal context, maximizing intelligence
Built-in validation - Multiple agents can cross-check each other's work
Reduced hallucination - If one agent hallucinates, others can correct it

How Sub-Agents Work

Architecture

┌─────────────────────────────────────┐
│   Main Agent (Orchestrator)         │
│   - Knows everything about you      │
│   - Plans and coordinates           │
│   - Synthesizes results             │
└──────────────┬──────────────────────┘
               │
       ┌───────┴───────┐
       │               │
┌──────▼──────┐ ┌─────▼───────┐
│ Sub-Agent 1 │ │ Sub-Agent 2 │
│ Research    │ │ Graphics    │
│ (Minimal    │ │ (Minimal    │
│  context)   │ │  context)   │
└─────────────┘ └─────────────┘

Key Principles

1. Orchestrator knows you - Main agent has full context about your preferences
2. Sub-agents are specialized - Each focuses on one specific task
3. Minimal context per sub-agent - They don't need your life story
4. Parallel execution - Work happens simultaneously
5. Results aggregation - Main agent synthesizes outputs

Enabling Sub-Agents

Default Behavior

Important: Sub-agents are NOT enabled by default. You must explicitly request them.

Basic Invocation

Use sub-agents to research this topic and create a presentation

Advanced Invocation

Can you make a presentation on [topic]?
- Send sub-agents to research why it's important
- Use another sub-agent to make the presentation
- Use other sub-agents to do the SVG graphics

Context Window Optimization

Why Sub-Agents Save Context

Main agent context:

• Your preferences: 50K tokens
• Project history: 30K tokens
• Current conversation: 20K tokens
• Total: 100K tokens

Sub-agent context:

• Task instructions: 5K tokens
• Relevant data only: 10K tokens
• Total: 15K tokens per sub-agent

Performance Impact

Lower-end models (MiniMax, Qwen):

• Perform significantly better with smaller context
• Sub-agents keep them in the "smart zone" (under 40% context)
• Cost-effective for parallel workflows

Higher-end models (Opus, Sonnet):

• Still benefit from specialization
• Faster execution through parallelism
• Better quality through focused attention

Practical Examples

Example 1: Research and Presentation

Prompt:

Create a presentation on OpenClaw sub-agents. 
Use parallel sub-agents for research, then synthesize 
into a presentation with graphics.

What happens:

1. Main agent spawns:

   • Research Agent 1: Reads OpenClaw documentation
   • Research Agent 2: Searches web for examples

2. Research agents work in parallel:

   • Agent 1 returns: Official documentation insights
   • Agent 2 returns: Community use cases and patterns

3. Main agent spawns:

   • Presentation Agent: Creates slides from research
   • Graphics Agent: Generates SVG illustrations

4. Main agent synthesizes:

   • Combines all outputs
   • Delivers final presentation

Result: Comprehensive, well-researched presentation with custom graphics

Example 2: Code Review

Prompt:

Review this codebase using sub-agents:
- One for security analysis
- One for performance review
- One for code style consistency

Benefits:

• Each agent specializes in one aspect
• Parallel execution saves time
• Cross-validation catches more issues

Example 3: Content Creation

Prompt:

Write a blog post about AI trends.
Use sub-agents to:
- Research latest developments
- Analyze competitor content
- Generate outline
- Write sections in parallel

Result: Higher quality content with diverse perspectives

Model-Specific Considerations

MiniMax and Chinese Models

Critical for success:

• Sub-agents are essential for good results
• Main agent should explicitly instruct sub-agent usage
• May need to remind agent daily to use sub-agents

Configuration tip: Add to your agent's memory or skills:

When handling complex tasks, always use parallel sub-agents 
to optimize context and improve quality.

Claude Opus

Natural sub-agent usage:

• Understands when to spawn sub-agents automatically
• Better at orchestration
• Handles complex multi-agent workflows

Still benefits from:

• Explicit instructions for critical tasks
• Parallel execution for time savings

Sub-Agent Improvements (Latest Version)

Error Reporting

Old behavior:

• Sub-agent fails silently
• No notification
• Main agent waits indefinitely

New behavior:

• Failure notifications delivered immediately
• Error details provided
• Can retry failed sub-agents

Status Updates

You now receive real-time updates:

• "Sub-agent 1 started: Researching documentation"
• "Sub-agent 2 started: Searching web"
• "Sub-agent 1 completed: Found 15 relevant sources"
• "Sub-agent 2 failed: Retrying with adjusted parameters"

Cost Considerations

Token Usage

More work = More tokens:

• Each sub-agent consumes tokens
• Parallel execution means simultaneous API calls
• Total cost is higher than single-agent approach

But:

• Each sub-agent uses fewer tokens (smaller context)
• Better results mean less rework
• Time savings offset cost increase

Cost Optimization

For MiniMax users:

• Subscription includes generous prompt allowance
• Sub-agents are cost-effective within plan limits

For Opus users:

• Monitor usage for expensive workflows
• Use sub-agents for high-value tasks
• Consider cheaper models for sub-agent tasks

Configuring Sub-Agent Models

Mixed Model Strategy

You can configure sub-agents to use different models:

Use Opus for main orchestration, but spawn 
sub-agents using Sonnet for cost efficiency

Benefits:

• Main agent has full intelligence for coordination
• Sub-agents use cheaper models for focused tasks
• Significant cost savings on large workflows

Configuration: (Advanced topic - see model configuration guide)

Best Practices

1. Explicit Task Breakdown

Good:

Create a presentation using sub-agents:
- Research agent: Find latest data
- Analysis agent: Identify key trends
- Writing agent: Draft content
- Graphics agent: Create visuals

Bad:

Make a presentation (hoping agent uses sub-agents)

2. Parallel When Possible

Efficient:

Spawn all research agents in parallel, then synthesize

Inefficient:

Research topic 1, then topic 2, then topic 3 sequentially

3. Clear Success Criteria

Good:

Research agent should return:
- Key findings
- Notable opinions
- Links to sources
- Patterns across sources
- Gaps nobody is talking about

Bad:

Research this topic

4. Validate Results

Main agent should:

• Cross-check sub-agent outputs
• Identify contradictions
• Synthesize coherent final result
• Flag low-confidence findings

5. Retry Failed Sub-Agents

With new error reporting:

Sub-agent 2 failed. Retry with more specific instructions.

Don't accept partial results - ensure all sub-agents complete successfully.

Common Pitfalls

Pitfall 1: Not Requesting Sub-Agents

Problem: Agent does everything itself, produces slop

Solution: Explicitly request sub-agent usage in prompt

Pitfall 2: Too Much Context Per Sub-Agent

Problem: Sub-agents receive full context, defeating the purpose

Solution: Main agent should send only task-relevant information

Pitfall 3: Sequential Execution

Problem: Sub-agents run one after another, wasting time

Solution: Request parallel execution explicitly

Pitfall 4: No Result Synthesis

Problem: Sub-agents return raw data, no coherent output

Solution: Main agent must synthesize and format results

Troubleshooting

"Sub-agents not being used"

Cause: Not explicitly requested, or agent doesn't understand

Solution:

1. Add to prompt: "Use parallel sub-agents for this task"
2. Add to agent memory: "Always use sub-agents for complex tasks"
3. Upgrade to latest OpenClaw version

"Sub-agent failed with no output"

Cause: Old OpenClaw version, or task too vague

Solution:

1. Update to latest version for error reporting
2. Provide more specific sub-agent instructions
3. Check sub-agent logs for details

"Results are inconsistent"

Cause: Sub-agents have different context or instructions

Solution:

1. Ensure main agent provides consistent instructions
2. Have main agent validate and reconcile differences
3. Use structured output formats

"Too expensive"

Cause: Too many sub-agents or wrong model selection

Solution:

1. Use sub-agents only for high-value tasks
2. Configure cheaper models for sub-agent work
3. Reduce number of parallel sub-agents

Advanced Patterns

Research Synthesis Pattern

1. Spawn 3-5 research agents with different sources
2. Each returns structured findings
3. Main agent identifies:
   - Common themes
   - Contradictions
   - Unique insights
   - Gaps in coverage
4. Synthesize into executive summary

Validation Pattern

1. Main agent completes task
2. Spawn validation sub-agent
3. Validator checks:
   - Accuracy
   - Completeness
   - Consistency
4. Main agent incorporates feedback

Iterative Refinement Pattern

1. Sub-agent produces draft
2. Main agent reviews
3. Spawn refinement sub-agent with specific feedback
4. Repeat until quality threshold met

Measuring Success

Quality Indicators

• Reduced hallucinations - Cross-validation catches errors
• More detailed output - Specialized agents go deeper
• Consistent formatting - Structured outputs maintained
• Faster completion - Parallel execution saves time

Before/After Comparison

Without sub-agents:

• Generic presentation
• Missing sources
• Boilerplate content
• 30 minutes execution time

With sub-agents:

• Detailed, researched presentation
• Cited sources with links
• Unique insights and analysis
• 15 minutes execution time (parallel)

Related Resources

• Context Window Management
• Skills Optimization
• Memory Management

---

Duration: 9 minutes
Difficulty: Intermediate
Video Reference: OpenClaw Sub-Agents EXPLAINED