agent-spawner — Multi-Agent Orchestration

Version: 1.0.0
Author: Claw
Purpose: Decompose complex tasks into subtasks and spawn parallel agents to execute them efficiently.

Overview

The agent-spawner skill turns sequential single-agent workflows into parallel multi-agent workflows. Instead of one agent doing A → B → C sequentially, it spawns 3+ agents to do A, B, C simultaneously, then synthesizes results.

Efficiency gain: 2-4x faster execution for multi-part tasks.

How to Use

1. Receive a complex task

Task examples:

• "Research the AI automation market in Czech Republic"
• "Compare these 5 projects: X, Y, Z, A, B"
• "Build a report on solar panel ROI for residential use"

2. Decompose into subtasks

Use scripts/spawn_planner.py or follow spawn patterns (see references/).

3. Spawn sub-agents

# For each independent subtask:
sessions_spawn \
  task="Execute subtask: \x3Cdescription>" \
  label="subtask-1" \
  mode="run" \
  runtime="subagent"

4. Yield and collect

Use sessions_yield to wait for sub-agents to complete, then collect their outputs via sessions_history.

5. Synthesize results

Combine sub-agent outputs into a coherent final deliverable. Resolve conflicts, merge findings, add context only you possess.

Spawn Patterns

Pattern A: Parallel Research

Use when: Multiple data sources need independent research. Example: "Research pricing for X across 5 competitors"

Spawn: competitor-A-price, competitor-B-price, competitor-C-price...
Collect: price data from each
Synthesize: comparison table

Pattern B: Build + Test + Document

Use when: Need code, tests, and docs simultaneously. Example: "Build a Python CLI tool with tests and documentation"

Spawn: builder (code), tester (tests), writer (docs)
Collect: source files, test results, doc files
Synthesize: complete package

Pattern C: Analyze → Summarize → Format

Use when: Raw data needs analysis, summary, and presentation. Example: "Analyze this dataset and create a visual report"

Spawn: analyzer (data processing), summarizer (insights), formatter (markdown/HTML)
Collect: analysis output, summary, formatted report
Synthesize: final deliverable

Pattern D: Review → Fix → Verify

Use when: Need code review with automated fixes. Example: "Review this codebase and fix all security issues"

Spawn: reviewer (audit), fixer (patches), verifier (tests)
Collect: findings, patches, verification results
Synthesize: reviewed code with changelog

Best Practices

1. Keep subtasks independent — no shared mutable state between agents
2. Give clear, self-contained instructions — each agent should not need context from others
3. Set timeoutSeconds — prevent runaway agents (default: 300)
4. Use descriptive labels — makes tracking and debugging easier
5. Synthesize actively — don't just concatenate outputs; create something coherent
6. One level deep — spawn agents from agents. Don't nest spawns more than 1 level.

Limitations

• Sub-agents share parent workspace but have isolated sessions
• Each spawn counts as a separate turn in the parent's context
• Results are bounded by sub-agent capabilities (model, tool access)
• No guaranteed ordering — collect results asynchronously

File Structure

agent-spawner/
  SKILL.md                    — This file
  references/
    spawn-patterns.md         — Detailed spawn patterns with examples
    model-selection.md        — When to use which model variant
  scripts/
    spawn_planner.py          — Task decomposition + spawn plan generator

Integration with OpenClaw Tools

This skill leverages:

• sessions_spawn — create parallel sub-agents
• sessions_yield — wait for results
• sessions_history — collect sub-agent outputs
• subagents — monitor and steer running sub-agents

Pricing

• Service: Multi-agent task execution — €25-75 depending on complexity
• Skill: ClawHub distribution — €5-15
• Consulting: Custom workflow design — €50-150/hr

Version History