OpenClaw Sub-Agents: How to Run Parallel Tasks Efficiently

By David Park | February 7, 2026

OpenClaw Sub-Agents: How to Run Parallel Tasks Efficiently

The single biggest bottleneck with AI agents is waiting.

You ask your agent to research a topic, and your entire conversation freezes. You request five independent tasks, and they execute one at a time. This sequential approach wastes hours.

OpenClaw's sub-agent architecture solves this by letting you spawn independent workers that run in the background while your main conversation continues.

---

What Sub-Agents Actually Do

Sub-agents are completely isolated worker sessions spawned from your main agent. Each sub-agent:

Runs in its own session (agent::subagent:)

Has its own context window and conversation history

Cannot see your main conversation

Reports results back when complete

Operates independently without blocking

Think of it like a construction site:

Your main conversation is the project manager

Sub-agents are specialized workers

Each worker handles one task independently

The manager coordinates and synthesizes results

---

When to Use Sub-Agents

Use sub-agents for:

Research tasks — Multiple topics simultaneously

Content creation — Multiple pieces at once

Data processing — Split large datasets

Background monitoring — Watch for changes while you work

Parallel web scraping — Gather data from multiple sources

Don't use sub-agents for:

Quick one-off questions

Tasks requiring main context

Single sequential steps

Trivial operations (overhead not worth it)

---

Basic Sub-Agent Usage

Spawning a Sub-Agent

openclaw sessions spawn --task "Research three competitors in the EV market. List their strengths, weaknesses, and market share." --label "EV Competitor Research"

Via the agent (in conversation):

You: "Can you research EV competitors in the background while we plan our marketing strategy?"

The agent will spawn a sub-agent automatically for suitable tasks.

Managing Sub-Agents

Use the /subagents slash command:

/subagents list          # Show all running sub-agents /subagents info      # Detailed info about one /subagents log  100  # View last 100 messages /subagents stop      # Terminate a sub-agent /subagents send     # Send additional instructions

---

Configuration Options

Default Model for Sub-Agents

Set sub-agents to use cheaper models:

{   "agents": {     "defaults": {       "subagents": {         "model": "claude-haiku",         "thinking": false       }     }   } }

Cost impact: Using Haiku instead of Opus can reduce sub-agent costs by 90%.

Concurrency Limits

Control how many sub-agents run simultaneously:

{   "agents": {     "defaults": {       "subagents": {         "maxConcurrent": 4       }     }   } }

Default is 8. Adjust based on your API limits and budget.

Auto-Archive Settings

Sub-agent sessions auto-cleanup after completion:

{   "agents": {     "defaults": {       "subagents": {         "archiveAfterMinutes": 60       }     }   } }

---

Practical Parallelization Patterns

Pattern 1: Research Fan-Out

Before (sequential):

Research Apple → Research Google → Research Microsoft Total time: 30 minutes

After (parallel):

Spawn sub-agent for Apple Spawn sub-agent for Google   Spawn sub-agent for Microsoft Total time: 10 minutes

Prompt:

You are a market researcher. Analyze [COMPANY] and provide: 1. Company overview 2. Recent product launches 3. Market position 4. Key strengths and weaknesses Report back with a concise summary for each section.

Pattern 2: Content Batching

Before (sequential):

Write blog post → Write social updates → Write email → Write newsletter Total time: 4 hours

After (parallel):

Spawn sub-agent: Blog post Spawn sub-agent: Social updates Spawn sub-agent: Email copy Spawn sub-agent: Newsletter Total time: 1 hour

Prompt:

Create [CONTENT_TYPE] about [TOPIC]: [TYPE-SPECIFIC REQUIREMENTS] Target audience: [AUDIENCE] Tone: [TONE] Length: [LENGTH]

Pattern 3: Data Pipeline Split

Before (sequential):

Process 10,000 rows → Aggregate results → Generate report Total time: 2 hours

After (parallel):

Spawn sub-agent: Process rows 1-2500 Spawn sub-agent: Process rows 2501-5000 Spawn sub-agent: Process rows 5001-7500 Spawn sub-agent: Process rows 7501-10000 Spawn main agent: Aggregate and report Total time: 45 minutes

---

Tool Access for Sub-Agents

Sub-agents have restricted tool access by default:

Tools Sub-Agents CAN Use:

read, write, exec (with approvals)

web_fetch, browser

memory_get, memory_search

Tools Sub-Agents CANNOT Use:

sessions_list, sessions_history

sessions_send, sessions_spawn

Other session manipulation tools

Customizing Tool Access

{   "tools": {     "subagents": {       "tools": {         "deny": ["gateway", "cron", "process"],         "allow": ["read", "write", "exec", "web_fetch"]       }     }   } }

---

Cost Optimization

Each sub-agent consumes tokens independently. Optimize with:

1. Use Cheaper Models

{   "agents": {     "defaults": {       "model": "claude-opus-4",       "subagents": {         "model": "claude-haiku"       }     }   } }

2. Limit Run Time

{   "agents": {     "defaults": {       "subagents": {         "maxRuntimeSeconds": 300       }     }   } }

3. Tight Task Scoping

Vague tasks → more iterations → more tokens

Bad: "Research AI agents" Good: "Find the top 5 AI agent frameworks by GitHub stars and last update date"

---

Results and Announcements

When a sub-agent completes:

1. Announce step runs inside the sub-agent session 2. Results post to your main conversation 3. Statistics included: - Runtime duration - Token usage (input/output/total) - Estimated cost (if pricing configured) - Session key and transcript path

Example announcement:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 📋 Sub-agent Complete: EV Market Research ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Status: ✅ Success Runtime: 7m32s Tokens: 12,451 in | 3,892 out | 16,343 total Cost: ~$0.12 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Summary Tesla (TSLA) Market cap: $975B Recent: Cybercab announced Q4 2025 Strength: Charging network Weakness: Price sensitivity

[Full report in transcript]

---

Limitations to Understand

No Nested Spawning

Sub-agents cannot spawn other sub-agents. This prevents runaway recursion.

Isolated Context

Each sub-agent starts fresh. Provide all necessary context in the initial task.

Best-Effort Announce

If the gateway restarts during a sub-agent run, the announce may be lost. Use /subagents log to recover results.

Shared Resources

Sub-agents share the same gateway process. Too many concurrent sub-agents can slow everything down.

---

Advanced: Custom Sub-Agent Models

Override the model per spawn:

openclaw sessions spawn \   --task "Write technical documentation" \   --model "claude-sonnet-4" \   --label "Tech Docs"

Useful when a task requires more reasoning but shouldn't use your main agent's premium model.

---

Monitoring and Control

Real-Time Monitoring

# Watch all sub-agent activity openclaw gateway logs --follow | grep subagent
Check specific sub-agent
 openclaw sessions history

Resource Management

# Stop all sub-agents from main session /stop
View sub-agent costs
 openclaw costs --subagents

---

Example Workflow: Market Analysis

Main conversation:

You: "I need a comprehensive analysis of the electric vehicle market. Research competitors, regulations, and trends." Agent: Spawning three sub-agents for parallel research... ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 📋 Sub-agent spawned: Competitor Analysis 📋 Sub-agent spawned: Regulations Research 📋 Sub-agent spawned: Trend Analysis ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ You: "While those run, let's discuss our go-to-market strategy." [Main conversation continues independently] [30 minutes later - announcements arrive] 📋 Competitor Analysis: ✅ Complete (12m, $0.08) 📋 Regulations Research: ✅ Complete (8m, $0.05) 📋 Trend Analysis: ✅ Complete (15m, $0.11)

You: "Great, now synthesize all three reports into a strategy recommendation."

---

Best Practices

1. Start simple — One sub-agent, observe behavior 2. Scope tightly — Narrow tasks = predictable costs 3. Monitor actively — Use /subagents list regularly 4. Kill when needed — Don't let runaway sub-agents accumulate 5. Use cheaper models — Haiku is sufficient for most sub-agent work 6. Provide full context — Sub-agents don't see your main conversation

---