一种标准化、开放的格式，为人工智能代理赋予明确的角色、背景和能力。

什么是安全带？

束是一个目录，为 AI 代理提供履行角色所需的一切信息。会话开始时，代理加载 HARNESS.md 以确定其身份;任务到达时，它只捕获每个任务所需的内容——逐步披露在束束层面应用。

顶层目录名称由约束者选择。skills/ 和 references/ 是常见的惯例，但并非必须——任何符合该域的结构都适用。

关键概念

HARNESS.md

所需的入口点。使用YAML前言，后跟简短路由正文：

保持身体最小化——每次激活都会负荷，并且与任务情境竞争。

路由文件

每个顶层目录都使用一个以其命名的全大写路由文件：TOOLS.md 代表 tools/，DATA.md 代表 data/，依此类推。这一惯例贯穿整个子树——tools/ 内的每个分组子目录也使用 TOOLS.md。路由文件允许代理判断分支是否相关，而无需打开其中所有文件。

终止（离开）

Agent Harnesses

A standardized, open format for giving AI agents a defined role, context, and capabilities.

What is a Harness?

A harness is a directory that gives an AI agent everything it needs to fulfill a role. At session start the agent loads HARNESS.md to establish its identity; as tasks arrive, it pulls in only the content each task requires — progressive disclosure applied at the harness level.

my-harness/
├── HARNESS.md          # Required: identity + routing overview
├── .leaf-detectors     # Optional: leaf boundary patterns
├── tools/              # Top-level dirs are named by the harness author
│   ├── TOOLS.md        # Routing file — describes what's in this branch
│   └── query-db/
│       └── SKILL.md
└── data/
    ├── DATA.md
    └── schema.md

Top-level directory names are chosen by the harness author. skills/ and references/ are common conventions but not required — any structure that fits the domain works.

Key Concepts

HARNESS.md

The required entry point. Uses YAML frontmatter followed by a brief routing body:

---
name: Marketing Assistant
description: Helps create and review content across blog, social, and visual channels.
---

You are a marketing assistant. Produce content that is on-brand and channel-appropriate.

- `tools/` — content creation and research capabilities (see TOOLS.md)
- `brand/` — tone, typography, and visual guidelines (see BRAND.md)
- `campaigns/` — active campaign briefs and goals (see CAMPAIGNS.md)

Keep the body minimal — it is loaded on every activation and competes with task context.

Routing Files

Each top-level directory uses a routing file named after it in all-caps: TOOLS.md for tools/, DATA.md for data/, and so on. This convention propagates through the entire subtree — every grouping subdirectory within tools/ also uses TOOLS.md. Routing files let an agent decide whether a branch is relevant without opening every file inside it.

Termination (Leaves)

Two mechanisms mark a directory as a leaf — a terminal point the routing layer does not recurse into:

• .harnessleaf — a file placed in any directory to signal an explicit boundary
• .leaf-detectors — a root-level file defining keyword patterns; e.g. skill=SKILL.md automatically marks any directory containing SKILL.md as a skill leaf

Leaf type names are meaningful: a skill leaf is loaded and executed as an Agent Skill; other types are treated as opaque boundaries.

Loading Model

Relationship to Agent Skills