📄 中文摘要
随着用户为 AI 代理安装越来越多的技能，系统提示中每个请求都携带了大量的技能描述信息，导致性能下降。具体来说，安装第 53 个技能后，系统提示中包含了 25KB 的技能描述，约 6200 个 token 的开销，造成在实际对话前的负担加重。为了解决这一问题，团队尝试了四种不同的方法，其中三种未能成功，最终确定了一种新的架构来优化性能和减少不必要的上下文信息。
我们如何解决 AI 代理技能生态系统中的上下文窗口膨胀问题
出处: How We're Solving Context Window Bloat in an AI Agent Skill Ecosystem
发布: 2026年3月18日
📄 English Summary
How We're Solving Context Window Bloat in an AI Agent Skill Ecosystem
As users install more skills on the AI agent, each request carries a significant amount of skill description data, leading to performance degradation. Specifically, after the 53rd skill is installed, the system prompt contains 25KB of skill descriptions, resulting in approximately 6,200 tokens of overhead before any actual conversation occurs. To address this issue, the team explored four different approaches, three of which failed. Ultimately, they settled on a new architecture designed to optimize performance and reduce unnecessary context information.