Agent Skills智能体技能

Agent Skills are a way to equip your agent with knowledge and company context. An Agent Skill is a folder containing a SKILL.md file, with scripts/, references/, and assets/directories. Section 2 covers the anatomy in detail.

Agent Skills are becoming the standard for cross-platform portability. But why the sudden adoption velocity? We believe Agent Skills tackle four main friction points in AI agent development:

1. Too many instructions, worse results. Dumping every instruction you can think of into a single system prompt inevitably degrades Large Language Model (LLM) performance, a problem known as context rot. Skills solve this by loading exclusively on demand. Section 5 unpacks the research behind this.

2. Knowing how, not just knowing what. LLMs already have reasonable analogs for remembering what happened (episodic memory) and remembering facts (semantic memory). What they've lacked is a way to remember how to do things step by step, which is called procedural memory. Agent Skills can be seen as the first credible procedural memory primitive for LLM Agents.

3. Multi-agent overload. The ecosystem was flooded with complex multi-agent systems that are notoriously hard to build and maintain. While still necessary for certain tasks, Skills allow a single general-purpose agent to seamlessly flex into many specialist roles. Section 3 develops this argument in depth, with a worked example.

4. Portability. A folder with a markdown file is a remarkably lightweight primitive. Any agent with filesystem access can use them, making them perfectly portable across a multi-vendor AI landscape.

In this whitepaper, we cater to two personas: Builders (those using Skills) and Developers (those creating, versioning, and managing them). We’ll gently walk through what a Skill is and how to use it, before diving into complex Developer topics like evaluation, production readiness, meta-skills, and composition. 智能体技能是一种为你的智能体配备知识和公司上下文的方式。一个智能体技能是一个包含 SKILL.md 文件的文件夹，以及 scripts/、references/ 和 assets/ 目录。第 2 节详细介绍了其结构。 智能体技能正成为跨平台可移植性的标准。但为什么突然被广泛采用？我们认为智能体技能解决了 AI 智能体开发中的四个主要痛点： 1. 指令过多，效果更差。将所有能想到的指令塞进一个系统提示中，不可避免地会降低大语言模型（LLM）的性能，这个问题被称为上下文腐烂。技能通过按需加载来解决这个问题。第 5 节揭示了这背后的研究。 2. 知道如何做，而不仅仅是知道是什么。LLM 已经有合理的类比来记住发生了什么（情景记忆）和记住事实（语义记忆）。它们缺乏的是记住如何逐步做事情的方式，这被称为程序性记忆。智能体技能可以被视为 LLM 智能体的第一个可信的程序性记忆原语。 3. 多智能体过载。生态系统充斥着复杂的多智能体系统，这些系统难以构建和维护。虽然在某些任务中仍然必要，但技能允许一个通用的单智能体无缝地灵活扮演多个专家角色。第 3 节通过一个实例深入探讨了这一论点。 4. 可移植性。一个包含 markdown 文件的文件夹是一个非常轻量级的原语。任何具有文件系统访问权限的智能体都可以使用它们，使它们在多供应商 AI 环境中完全可移植。 在本白皮书中，我们针对两种角色：构建者（使用技能的人）和开发者（创建、版本管理和管理技能的人）。我们将逐步介绍什么是技能以及如何使用它，然后深入探讨复杂的开发者主题，如评估、生产就绪、元技能和组合。