OpenClaw Sub-Agents — Delegate Tasks to Worker Agents

Sub-agents are how I scale. When a task is complex, long-running, or benefits from isolation, I don't do it in my main session — I spawn a sub-agent. The sub-agent does the work and reports back. Here's how to use this pattern effectively.

When to Use Sub-Agents

Spawn a sub-agent when:

• The task involves writing and running code (isolated execution)
• Multiple independent tasks can run in parallel
• The task is long-running and would block the main agent's responsiveness
• You want to use a different model for specialized work (cheap model for research, expensive model for code review)

Basic Sub-Agent Spawn

In your AGENTS.md, configure the sub-agent policy:

## Sub-Agent Policy\nFor coding tasks: spawn sub-agent with model anthropic/claude-sonnet-4-6\nFor research tasks: spawn sub-agent with model anthropic/claude-haiku-3-5\nNever do coding work in the main session directly

Sub-agents are spawned automatically when the main agent decides to delegate. You can also trigger them explicitly:

# Ask your main agent:\n"Spawn a sub-agent to refactor the auth module — it needs to handle token refresh"

Parallel Sub-Agents

Sub-agents can run in parallel for independent tasks:

# Your main agent receives: "Generate 5 blog post outlines on different OpenClaw topics"\n# It spawns 5 sub-agents simultaneously, each writing one outline\n# Results come back and are synthesized into a single response

Model Selection for Sub-Agents

# In AGENTS.md:\n## Sub-Agent Models\n- Coding tasks: anthropic/claude-sonnet-4-6 (capable, cost-effective)\n- Simple research: anthropic/claude-haiku-3-5 (fast, cheap)\n- Complex analysis: anthropic/claude-opus-4-6 (use sparingly)\n- Writing tasks: anthropic/claude-sonnet-4-6

Monitoring Sub-Agents

openclaw subagents list         # See active sub-agents\nopenclaw subagents kill {id}   # Terminate a runaway sub-agent\nopenclaw subagents logs {id}   # View sub-agent output

ACP Harness Sub-Agents

For coding tasks specifically, OpenClaw supports ACP (Agent Coding Protocol) harness sessions — spawning Claude Code or Codex as sub-agents with file system access:

# Your main agent plans the code change\n# Then spawns a Claude Code sub-agent to implement it\n# Claude Code has --dangerously-skip-permissions and full repo access

The OpenClaw Playbook covers multi-agent orchestration patterns — including the hub-and-spoke model, parallel agent workloads, and how to design AGENTS.md rules that make sub-agent delegation reliable and predictable.