swarm
Spawn isolated Codex sub-agents for parallel task execution using the current runtime primitives. Triggers: "swarm", "spawn agents", "parallel work", "run in parallel", "parallel execution".
Best use case
swarm is best used when you need a repeatable AI agent workflow instead of a one-off prompt.
Spawn isolated Codex sub-agents for parallel task execution using the current runtime primitives. Triggers: "swarm", "spawn agents", "parallel work", "run in parallel", "parallel execution".
Teams using swarm should expect a more consistent output, faster repeated execution, less prompt rewriting.
When to use this skill
- You want a reusable workflow that can be run more than once with consistent structure.
When not to use this skill
- You only need a quick one-off answer and do not need a reusable workflow.
- You cannot install or maintain the underlying files, dependencies, or repository context.
Installation
Claude Code / Cursor / Codex
Manual Installation
- Download SKILL.md from GitHub
- Place it in
.claude/skills/swarm/SKILL.mdinside your project - Restart your AI agent — it will auto-discover the skill
How swarm Compares
| Feature / Agent | swarm | Standard Approach |
|---|---|---|
| Platform Support | Not specified | Limited / Varies |
| Context Awareness | High | Baseline |
| Installation Complexity | Unknown | N/A |
Frequently Asked Questions
What does this skill do?
Spawn isolated Codex sub-agents for parallel task execution using the current runtime primitives. Triggers: "swarm", "spawn agents", "parallel work", "run in parallel", "parallel execution".
Where can I find the source code?
You can find the source code on GitHub using the link provided at the top of the page.
SKILL.md Source
# $swarm — Parallel Agent Execution (Codex Tailoring) This override captures the Codex-native execution model for parallel work. The active runtime primitives are: - `spawn_agent` - `send_input` - `wait_agent` - `close_agent` Use `agent_type="explorer"` for read-only discovery and `agent_type="worker"` for implementation workers. ## Core Rules 1. Only parallelize tasks with disjoint file ownership. 2. Give every worker an explicit file manifest and validation command. 3. Tell each worker it is not alone in the repo and must not revert unrelated edits. 4. Wait sparingly. Keep the lead agent doing useful non-overlapping work. 5. The lead agent validates, integrates, and closes the loop. ## Codex-Native Flow ### Step 1: Prepare tasks Each task must define: - `id` - `subject` - `description` - `files` - `validation` If file ownership is unknown, spawn `explorer` agents first to map the blast radius. ### Step 2: Conflict check Do not run workers in the same wave if they claim overlapping files. Split them into sub-waves instead. ### Step 3: Spawn workers Use one `spawn_agent(...)` call per worker with: - `agent_type="worker"` - a bounded task - explicit owned files - explicit validation steps ### Step 4: Monitor Use `wait_agent(...)` only when the next integration step actually depends on the worker result. If a worker needs a correction, use `send_input(...)` with a focused follow-up. ### Step 5: Integrate The lead agent reviews worker outputs, runs repo-level validation, and decides whether to launch another wave. ## Fallback If sub-agents are unavailable, execute the same tasks sequentially in the lead session while keeping the same file-ownership and validation discipline.
Related Skills
vibe
Comprehensive code validation. Runs complexity analysis then multi-model council. Answer: Is this code ready to ship? Triggers: "vibe", "validate code", "check code", "review code", "code quality", "is this ready".
validation
Full validation phase orchestrator. Vibe + post-mortem + retro + forge. Reviews implementation quality, extracts learnings, feeds the knowledge flywheel. Triggers: "validation", "validate", "validate work", "review and learn", "validation phase", "post-implementation review".
update
Reinstall all AgentOps skills globally from the latest source. Triggers: "update skills", "reinstall skills", "sync skills".
trace
Trace design decisions and concepts through session history, handoffs, and git. Triggers: "trace decision", "how did we decide", "where did this come from", "design provenance", "decision history".
test
Test generation, coverage analysis, and TDD workflow. Triggers: "test", "generate tests", "test coverage", "write tests", "tdd", "add tests", "test strategy", "missing tests", "coverage gaps".
status
Single-screen dashboard showing current work, recent validations, flywheel health, and suggested next action. Triggers: "status", "dashboard", "what am I working on", "where was I".
standards
Language-specific coding standards and validation rules. Provides Python, Go, Rust, TypeScript, Shell, YAML, JSON, and Markdown standards. Auto-loaded by $vibe, $implement, $doc, $bug-hunt, $complexity based on file types.
shared
Shared reference documents for multi-agent skills (not directly invocable)
security
Continuous repository security scanning and release gating. Triggers: "security scan", "security audit", "pre-release security", "run scanners", "check vulnerabilities".
security-suite
Composable security suite for binary and prompt-surface assurance, static analysis, dynamic tracing, repo-native redteam scans, contract capture, baseline drift, and policy gating. Triggers: "binary security", "reverse engineer binary", "black-box binary test", "behavioral trace", "baseline diff", "prompt redteam", "security suite".
scenario
Author and manage holdout scenarios for behavioral validation. Scenarios are stored in .agents/holdout/ where implementing agents cannot see them. Triggers: "$scenario", "holdout", "behavioral scenario", "create scenario", "list scenarios".
scaffold
Project scaffolding, component generation, and boilerplate setup. Triggers: "scaffold", "new project", "init project", "create project", "generate component", "setup project", "starter", "boilerplate".