# Agent Native

> Make an out-of-session agent AgentOps-native via skills + ao CLI + CI, not hooks.

- Skill: `majiayu000/agent-native-2` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds add majiayu000/agent-native-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/majiayu000/agent-native-2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: majiayu000 (https://skillmd.com/u/majiayu000)
- Updated: 2026-09-09
- Page: https://skillmd.com/skills/majiayu000/agent-native-2

---


# $agent-native — Make Out-of-Session Agents AgentOps-Native (Codex Native)

> **Quick Ref:** Run a Claude/Codex loop *outside* an interactive session (Managed Agent, Agent SDK, or self-hosted sandbox) under the same AgentOps guardrails — hooklessly. Guardrails = skills + the `ao` CLI + CI, never ported hooks. Bundle skills into the agent definition, expose `ao` as a callable tool so the loop self-bootstraps + self-validates, and gate the output through the SAME CI as interactive work.

## Codex/NTM path

Codex (gpt-5.3-codex) has NO Managed Agents API, NO Workflow tool, NO Task subagent. It orchestrates via NTM (tmux pane swarms + agent-mail) + skills + the `ao` CLI + ssh to bushido. So an out-of-session Codex loop becomes AgentOps-native the same way: load the AgentOps skills, call `ao session bootstrap` / `ao inject` / `ao validate` directly (no MCP needed — Codex shells out), and gate outputs through CI (`agent-output-validate.yml`). `ao` does not wrap `gc`; a whole loop needing a mayor/refinery still routes through `gc`.

## Instructions

Load and follow the skill instructions from the sibling `SKILL.md` — OR read `skills/agent-native/SKILL.md` in the host repo for the canonical specification. Honor the Critical Constraints: this is a reframe of "port hooks", NOT a hook revival; no skill fork (load the same `skills/` files); Managed Agents are NOT ZDR (no holdout/PII in the agent definition); CI is the enforcement boundary, not the optional in-loop adapter.