Sub-Agents Skills
Runs sub-agents from Markdown files in the .agents folder on backends such as Codex, Claude Code, Cursor CLI and Gemini.
- 100 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Sub-agents run without prompts in non-interactive mode and can edit files.
- `yolo` mode disables approvals and the sandbox.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/sub-agents-skillsDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Sub-agents run without prompts in non-interactive mode and can edit files.
- `yolo` mode disables approvals and the sandbox.
This is third-party code. Review the repository files before installing.
What it does
Sub-Agents Skills installs a `Runner` skill into the parent agent that launches sub-agents from Markdown files in the `.agents/` folder. The `run-agent` field in the file header selects the backend: Codex, Claude Code, Cursor CLI, Grok Build, Antigravity, Gemini CLI, OpenCode and more. Role, task and done criteria are written once and the executor changes with one line. The `permission` field sets the access level: `read-only`, `safe-edit` by default or `yolo`. A sub-agent starts without stdin in non-interactive mode and returns its result to the parent. Codex, Claude Code, Grok Build and Antigravity have plugins, while Cursor, VS Code and Gemini CLI use the `install.sh` script.
Who it is for. Developers with several AI coding tools who route review, implementation and verification to the backend that suits each best.
Good fit when
- Review runs in one tool and implementation in another, with no duplicated role definitions.
- Two reviewers on different models compare findings, then the fixes go to a third executor.
- A test or documentation agent is written once and launched from any parent client.
- An investigation is limited to `read-only` while the agent must change nothing.
Not a fit when
- Built-in sub-agents of one tool are enough: portability between backends is not needed.
- The task cannot split into independent parts: each sub-agent starts with a clean context.
- The environment is untrusted: `yolo` mode disables approvals and the sandbox.
Example request
Run the code-reviewer agent on the authentication changes and a second reviewer on another backend in parallel, then merge the findings.Limitations
Python 3.9 or later and at least one installed CLI from the backend list are required. The same `permission` mode gives different isolation guarantees on different backends, and `read-only` is not a security boundary. With `mcode`, `read-only` still allows direct edits. A sub-agent runs without stdin, so there is no dialogue. Name both the agent and the task in the request: an agent name alone sets no task. An agent must not refer to other agents or earlier context: every run starts from a clean slate.
How to disable. Remove the `runner` plugin with the client's tools (the `/plugin` menu in Claude Code, `/plugins` in Codex) or the skill folder if you used the script. Delete `.agents/` folders in projects by hand.
FAQ
Which backends are supported?
Codex, Claude Code, Cursor CLI, GLM and Kimi through `claude`, Grok Build, Antigravity, Gemini CLI, OpenCode, Command Code, MiniMax Code and Pi.
What does permission mean?
`read-only` is for checks without edits, `safe-edit` (the default) allows edits in the working folder and `yolo` bypasses approvals and the sandbox.
How do I call the skill?
In Codex after plugin install with `$runner:sub-agents`. Other clients take a request that names the agent and the task, for example reviewing the authentication changes.
Can backends be mixed?
Yes, `.agents/` can hold files with different `run-agent` values and the parent calls them in parallel.
How does this differ from built-in sub-agents?
Built-in ones are tied to their own model: Claude Code delegates to Claude and Codex to GPT. Here the role is separate from the backend.
Discussion
No questions or reviews yet.
Related
SkillsA base AGENTS.md and 12 skills: delegation to Codex, Grok, Devin and Antigravity, planning, copywriting, scoped changes.
A command-line skill manager: installs skills from a marketplace, GitHub, private Git and npm into 45 agents' folders and scans them first
The squads CLI runs teams of agents defined in markdown files in git: a lead briefs, workers act, a verifier checks.
A scaffold of personas, rules and skills in plain Markdown: Maestro hands work to different models and spares your main plan
One config for Claude Code and Codex in every repo: shared rules, skill packs, second-model review and a destructive-command guard
A catalog of nearly two hundred skills: code delivery, review, debugging, AI governance, product work and agent coordination.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Sub-Agents Skills (this card) | no | 100 | |
| Oh My Subagents | no | 180 | keeps Codex and Claude sub-agent teams in a local service; here a Markdown role and a one-line backend choice |
| pi-subagents | no | 1,449 | an extension for Pi only; here one skill runs agents on twelve backends |
README in short
The README lists clients and backends, three permission levels and rules for agent authors: one responsibility, a self-contained file, explicit boundaries. Translations exist in Russian, Chinese, German and Spanish.