📄 中文摘要
大型语言模型（LLMs）在特定时间点进行训练，因此其知识是固定的。软件工程实践快速变化，新库每天推出，最佳实践也在不断演变。这使得语言模型无法单独弥补知识差距。在Google DeepMind，这种情况表现为模型在训练时对自身缺乏了解，并且对最佳实践的微妙变化（如思维循环）或SDK的更改并不总是知晓。虽然存在多种解决方案，如网络搜索工具和专门的MCP服务，但最近出现了代理技能等新方法，以帮助填补这一知识空白。
通过代理技能缩小知识差距
出处: Closing the knowledge gap with agent skills
发布: 2026年3月31日
📄 English Summary
Closing the knowledge gap with agent skills
Large language models (LLMs) are trained at a specific point in time, resulting in fixed knowledge. The fast-paced nature of software engineering, with new libraries launched daily and evolving best practices, creates a knowledge gap that language models cannot bridge alone. At Google DeepMind, this is evident as models lack awareness of their own capabilities during training and may not recognize subtle changes in best practices, such as thought circulation or SDK updates. While various solutions exist, including web search tools and dedicated MCP services, recent developments like agent skills have emerged to help address this knowledge gap.