Anthropic与Linux Foundation旗下的Agentic AI Foundation（AAIF）在10月22日分别宣布捐赠与接纳Agent Skills。移交范围包括项目名称、标识、GitHub组织和agentskills.io域名，AAIF表示，将由来自多家机构的维护者以个人身份参与技术指导委员会治理。Skills于2025年10月推出，同年12月成为开放标准；这次应理解为治理迁移，而不是新模型或一套刚刚诞生的功能。

On October 22, Anthropic and the Agentic AI Foundation, or AAIF, under the Linux Foundation separately announced the donation and acceptance of Agent Skills. The transfer includes the project name, its marks, the GitHub organization and the agentskills.io domain. AAIF said a technical steering committee would be governed by maintainers from several organizations participating as individuals. Skills launched in October 2025 and became an open standard in December of that year. This development should therefore be understood as a transfer of governance, rather than a new model or a capability that has only just been created.

技能目录以SKILL.md描述名称、用途与工作指令，并可附上脚本、范例、参考资料和模板。可以把它想成团队可搬走的操作手册：例如报告格式、审阅步骤和设计规范，不必每次从头向智能体解释。它给模型提供做事的方法和材料，不是在重新训练模型权重。

A skill directory uses SKILL.md to describe its name, purpose and working instructions, and it can include scripts, examples, reference materials and templates. Think of it as an operating manual that a team can take with it: report formats, review procedures and design standards do not have to be explained to an agent from scratch each time. It provides a model with methods and materials for doing a task; it does not retrain the model's weights.

规范描述的渐进式加载，是先让智能体看到名称和简介，决定需要某项技能后再读完整步骤，必要时继续读取资源。这样可以避免把所有材料一次塞入上下文，但本轮没有实际运行测试，不能给出节省多少token、时间或成本的比例。最终效果还取决于客户端是否正确识别、加载和执行这些步骤。

The specification describes progressive loading: the agent first sees a name and short description, reads the full instructions after deciding that it needs the skill, and accesses additional resources when necessary. This can avoid putting all available materials into the context at once. However, this review did not include a practical execution test, so it cannot establish a percentage saving in tokens, time or cost. The outcome also depends on whether the client correctly identifies, loads and carries out those instructions.

技能格式负责封装工作方法，工具调用协议负责连接能够调用的服务。即使拿到同一个技能包，不同客户端的工具接口、依赖库、文件权限和模型行为也可能不同。规范中的环境兼容声明能帮助使用者判断条件，不能替代运行验证；实验性的allowed-tools字段也不能被当成所有实现都一致执行的权限保障。

A skill format packages a way of working, while a tool-calling protocol connects services that can be invoked. Even with the same skill package, clients may differ in their tool interfaces, software dependencies, file permissions and model behavior. An environment-compatibility declaration in the specification can help users assess prerequisites, but it cannot replace execution tests. The experimental allowed-tools field should not be treated as a permissions safeguard that every implementation enforces consistently.

标准结构校验可以检查文件命名或元数据，却不证明脚本安全、业务结果正确或已经获得操作授权。基金会托管也不等于每份技能都经过安全认证。正式版本流程、提案机制和安全指引仍有拟议工作，应以实际落地文件为准。

Structural validation of a standard can check file naming or metadata, but it does not prove that a script is safe, that a business result is correct, or that an operation has been authorized. Foundation stewardship also does not mean that every skill has been security-certified. Formal versioning, proposal procedures and security guidance still include planned work, which should be assessed against the documents that are actually put in place.

对团队而言，值得尝试的是能否在兼容工具间保留自己的流程知识，并用具体任务验证迁移效果。下一步看版本兼容测试、技能来源与脚本审查、权限隔离和治理流程是否落实。本轮没有独立性能实测，也不据采用量推算收入或投资回报。

For teams, the useful experiment is whether they can preserve their process knowledge across compatible tools and test the outcome of migration on specific tasks. The next things to examine are version-compatibility tests, the provenance of skills and review of their scripts, separation of permissions, and implementation of the governance process. This review contains no independent performance measurements and does not infer revenue or investment returns from adoption counts.