📄 中文摘要
模型上下文协议（MCP）被视为为 AI 代理提供工具的工具箱，而代理技能则是教授 AI 代理如何完成任务的材料。这种方法与预训练或后训练不同，后者决定了模型的整体行为和专业知识。代理技能并不对代理进行“训练”，而是通过软分叉的方式改变代理的行为，使其能够在特定任务中表现出更高的专业性。通过这种方式，AI 代理能够在不需要传统训练过程的情况下，快速适应并执行多样化的任务。
软分叉：代理技能如何在不训练的情况下创造专业化 AI
出处: Soft Forks: How Agent Skills Create Specialized AI Without Training
发布: 2026年3月9日
📄 English Summary
Soft Forks: How Agent Skills Create Specialized AI Without Training
The Model Context Protocol (MCP) serves as a toolbox that equips AI agents with tools, while Agent Skills act as materials that instruct AI agents on how to accomplish tasks. This approach differs from pre-training or post-training, which define a model's general behavior and expertise. Agent Skills do not 'train' agents; instead, they soft-fork agent behavior, enabling them to exhibit higher specialization in specific tasks without the need for traditional training processes. This allows AI agents to quickly adapt and perform a diverse range of tasks effectively.