Parallel execution engine — run multiple AI-agent tasks simultaneously instead of sequentially.
Stop waiting in line when you can use all lanes. Ultrawork fires independent tasks in parallel, routes each to the right model tier, and collects results — all at once.
- True parallelism — spawns subagents simultaneously for independent tasks
- Tier routing — auto-selects the right model per task (LOW/STANDARD/THOROUGH)
- PID tracking — real process management via
wait -nand PID arrays - Zero dependencies — pure bash, no npm/pip/containers needed
- OpenClaw / Hermes compatible — plugs in as a skill
# Clone to your OpenClaw skills directory
git clone https://github.com/nerudek/parallel-agent-executor.git ~/.openclaw/skills/ultrawork
# Or symlink if you keep skills elsewhere
ln -s /path/to/parallel-agent-executor ~/.openclaw/skills/ultrawork# Multiple independent tasks — all fire at once
/ultrawork "add types to auth.ts" "fix bug in utils.ts" "update README"
# With explicit count
/ultrawork 3 "task1" "task2" "task3"
# Parallel file processing
/ultrawork "analyze src/*.ts" "generate docs" "run tests"Or just say: ulw, ultrawork, run in parallel, do these at once
Sequential (slow): Parallel (fast):
┌─────────┐ ┌─────────┐
│ Task A │ ──10s──> │ Task A │ ──┐
├─────────┤ ├─────────┤ │
│ Task B │ ──10s──> │ Task B │ ──┼──10s──> Done
├─────────┤ ├─────────┤ │
│ Task C │ ──10s──> │ Task C │ ──┘
└─────────┘ └─────────┘
Total: 30s Total: 10s
| Tier | Model | Use For | Speed |
|---|---|---|---|
| LOW | qwen3-coder | Simple lookups, type exports | ⚡ Fast |
| STANDARD | qwen3.5-35b | Normal implementation | 🚀 Normal |
| THOROUGH | kimi-k2.5 | Complex analysis, security | 🐢 Thorough |
Use it when: tasks are independent and you want to cut total execution time. Skip it when: tasks have dependencies (B needs A) or you need guaranteed sequential verification.
Full docs: SKILL.md | SKILL.hermes.md
Current version: 1.0.0 | License: MIT | Author: nerudek
If this saved you time: ☕ PayPal.me/nerudek