Agent Skills
Agent Skills is a capability encapsulation mechanism for intelligent agent systems, designed to abstract specific tasks or workflows into reusable and composable capability units. Key characteristics include modular encapsulation that allows skills to be developed and invoked independently, support for capability reuse and orchestration across agent scenarios, and interoperability with tool-calling protocols such as MCP, collectively enhancing the flexibility and scalability of intelligent agent systems.
Core Facts
A Skill is a modular encapsulation that packages specialized capabilities, domain knowledge, and operational norms into reusable units, enabling Agents to invoke them automatically at the appropriate moment.
A complete Skill contains a Skill.md main document, a Reference folder, a Script folder, and an Assets folder.
Agent development is moving from an experimental phase to an engineering phase.
Timeline (last 90 days)
Agent Skills are designed with progressive disclosure: Claude first sees a list of skills with brief descriptions, judges whether they are relevant, and only then reads the full content, to avoid occupying too much context at once.
Agent Skills belong to the lowest-cost category of domain adaptation: they do not modify model weights and only guide behavior through structured context.
Although Agent Skills can improve efficiency, developers still need to retain the ability to review generated code and should version-control and peer-review key skills to prevent outdated specifications from misleading the agent.
Developers should choose matching skills based on the project's tech stack and team conventions; for example, when a project uses React and Tailwind CSS, they should prioritize finding skills defined for these technologies.
Developers should place Agent Skills' Markdown files in the project-level rules directory or explicitly reference them in the conversation, so that the agent follows the established component specifications and style conventions when generating interface code.
The tutorial's subsequent livestreams cover topics including LangChain 1.2, LangGraph 1.0, DeepAgent, and Agent Skills, and mention that LangChain 1.2 has almost no changes compared with 1.0.
The researchers propose the Conflict of Skills Attack (CoSA): without modifying existing Skills or runtime permissions, adding a small number of locally admitted new Skills to the skill pool can significantly raise the policy violation rate.
Progressive disclosure means that the Agent gradually loads and discloses relevant skill information at runtime according to task needs, rather than exposing all capabilities to the model at once.
When designing the skill list, developers should control the granularity of individual skills—too coarse makes them hard to reuse, while too fine leads to a bloated number of skills and higher selection costs.
Whether Agent Skills can become a standard paradigm of the Agent development ecosystem still needs time to be verified.
40 more timeline events
All Facts (20)
A Skill is a modular encapsulation that packages specialized capabilities, domain knowledge, and operational norms into reusable units, enabling Agents to invoke them automatically at the appropriate moment.
90%VerifiedA complete Skill contains a Skill.md main document, a Reference folder, a Script folder, and an Assets folder.
80%VerifiedAgent development is moving from an experimental phase to an engineering phase.
80%VerifiedAgent的核心公式为：Agent = 大模型 + 工具调用 + 自主决策能力
80%VerifiedAgent Skills 具备上下文感知、可组合性和自描述性三个关键特征
80%VerifiedSkills本质上是结构化的系统提示词（System Prompt）集合，通过标准化接口注入到AI代理的上下文中
80%VerifiedSkills are not competitors to MCP but rather a layer above MCP: MCP is responsible for connecting tools and standardizing external service interfaces, while Skills are responsible for packaging and encapsulating complete task execution capabilities.
80%VerifiedAgent Skills是Anthropic发布的能力封装标准，定义为一个包含Skill.md文件的目录，可包含说明、脚本和其他资源
80%VerifiedMCP addresses the connectivity problem (how a model reaches tools), while Skill addresses the usage problem (when and with what strategy the model combines tools).
75%VerifiedHermes Agent, OpenCode, and Claude Code all adopt the Agent Skills standard, and Skills can be used interchangeably across them.
75%VerifiedThe Skills folder may contain subdirectories such as scripts, references, and assets, and other custom folders can also be added, but they must be explicitly pointed to in SKILL.md, otherwise the agent cannot make basic judgments.
70%VerifiedAgent Skill adopts a lazy-loading mechanism whereby the Agent first reads only the metadata (name and description) of each Skill and loads the full instructions only after matching the user's need, so having multiple Skills coexisting does not increase Token consumption
70%VerifiedAgent Skills and the MCP protocol are both concepts pioneered by Anthropic, with other vendors following suit afterward.
70%VerifiedAgent Skills adopts a progressive loading mechanism that loads relevant sections on demand based on the current task, avoiding the context explosion problem.
70%VerifiedSkill's progressive loading mechanism has two levels: the first level exposes only name and description to form a minimal skill catalog; the second level reads the full SKILL.md and supporting files only after a hit.
65%VerifiedWhen the model decides whether to invoke a given Skill, it relies on semantic similarity matching and intent recognition, semantically comparing the request content against the descriptions of loaded Skills.
65%VerifiedThe description is intended for large language models, which rely on the description to discover and invoke the corresponding Skills.
65%VerifiedThe SKILL.md metadata must include at minimum two fields: name and description.
65%VerifiedThe SKILL.md file begins with a metadata section (Frontmatter) enclosed by three dashes, which is the key to the Skill being recognized by AI tools.
65%VerifiedMulti-Agent architecture is better suited to engineering scenarios where subtasks can be clearly decomposed and are relatively independent, rather than being the default architecture for all tasks.
65%