使用 Skills 构建 ADK Agents

APRIL 1, 2026

Lavi Nigam Developer Relations Engineer

Lavi Nigam 开发者关系工程师

Shubham Saboo Senior AI Product Manager

Shubham Saboo 高级 AI 产品经理

part1-cover

Your AI agent can follow instructions. But can it write new ones? Agent Development Kit (ADK)'s SkillToolset enables agents to load domain expertise on demand. With the right skill configuration, your agent can generate entirely new expertise at runtime. Whether you need a security review checklist, a compliance audit, or a data pipeline validator, the workflow is straightforward: generate it, load it, and use it.

您的 AI 代理可以遵循指令。但它能编写新的指令吗？Agent Development Kit (ADK) 的 SkillToolset 使代理能够按需加载领域专业知识。有了正确的技能配置，您的代理可以在运行时生成全新的专业知识。无论您需要安全审查清单、合规审计还是数据管道验证器，工作流程都很简单：生成它、加载它并使用它。

The SkillToolset achieves this through progressive disclosure. This architectural pattern allows agents to load context precisely when it is needed, rather than cramming thousands of tokens into a monolithic system prompt.

SkillToolset 通过 progressive disclosure 实现了这一点。这种架构模式允许代理在需要时精确加载上下文，而不是将数千个 tokens 塞入单一的系统提示中。

In this guide, we will walk through four practical skill patterns:

在本指南中，我们将逐步介绍四种实用的技能模式：

The inline checklist: A basic, hardcoded skill implementation.
内联清单： 一种基本的、硬编码的技能实现。
The file-based skill: Loading external instructions and resources.
基于文件的技能： 加载外部指令和资源。
The external import: Utilizing community-driven skill repositories.
外部导入： 利用社区驱动的技能仓库。
The skill factory: A self-extending pattern where the agent writes new skills on demand.
技能工厂： 一种自我扩展模式，其中代理按需编写新技能。

Each pattern builds on the previous one, culminating in an agent architecture capable of dynamically expanding its own capabilities.

每个模式都建立在上一个模式的基础上，最终形成一个能够动态扩展自身能力的代理架构。

The problem with monolithic prompts

单体式提示的问题

Most AI agents get their domain knowledge directly from the system prompt. Developers often concatenate compliance rules, style guides, API references, and troubleshooting procedures into one massive instruction string.

大多数 AI 代理直接从系统提示获取领域知识。开发人员经常将合规规则、风格指南、API 参考和故障排除程序连接成一个巨大的指令字符串。

This works fine when an agent...