# Swarm

> Fan out independent slices, races, or coverage checks to Codex subagents and aggregate one evidence-backed report. Use for $swarm or explicitly parallel work.

- Skill: `vaibtan/swarm` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add vaibtan/swarm`
- Raw SKILL.md: https://api.skillmd.com/api/skills/vaibtan/swarm/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Docs & Writing
- Author: Vaibtan (https://skillmd.com/u/vaibtan)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/vaibtan/swarm

---


# Swarm

Use a swarm for separable work. Use `$arena` when every worker is producing the same artifact for synthesis.

1. State the done predicate, required slices, and whether the shape is coverage, first-pass race, rank-all race, or mixed.
2. Set the worker count from the independent work and available collaboration slots. Do not create workers whose coordination cost exceeds their task.
3. Read `$CODEX_HOME/pstack-models.toml` when present. Default ordinary workers to `general`; use `fast` for bounded mechanical inventory and `deep` for hard judgment.
4. Give each worker a self-contained brief with goal, exact scope, allowed writes, verification, and report format. Separate writable files or worktrees.
5. Launch independent workers together. Wait through the Codex collaboration mechanism and avoid rapid polling.
6. Review results. Every required coverage slice needs evidence or an explicit gap. Apply the declared race rule.
7. Return one compact report with status per slice, evidenced issues, dropouts, unresolved gaps, and integrated verification.

Subagents report `PASS`, `ISSUES`, or `BLOCKED` with evidence. Their labels are inputs to the parent's judgment, not final authority.