<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Zegging&apos;s Tech Blog</title><description>国际化 saas 服务 web backend 开发 &amp;&amp; ai agent 开发 &amp;&amp; ai infra developer</description><link>https://zegging.github.io/</link><item><title>如果把 Agent 的 Context 当成一块需要 GC 的内存</title><link>https://zegging.github.io/posts/agent/theory/agent-context-gc-theory/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/theory/agent-context-gc-theory/</guid><description>从一个关于 JVM GC 的直觉出发，重新思考 Agent 如何在有限 Context Window 上维持近乎无限的运行，以及 Context GC 的对象、Root、可达性和安全边界。</description><pubDate>Sun, 28 Jun 2026 17:10:00 GMT</pubDate></item><item><title>一种 LLM Agent Runtime 的形式化模型</title><link>https://zegging.github.io/posts/agent/theory/formal-model-of-llm-agent-runtime/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/theory/formal-model-of-llm-agent-runtime/</guid><description>将 Agent 拆解为 Runtime 状态、模型上下文、厂商模型、Controller、外部环境与状态更新，给出一套可用于讨论 Agent 的形式化模型。</description><pubDate>Sat, 27 Jun 2026 01:36:00 GMT</pubDate></item><item><title>新的潮流“创建 Loop 来运行 Agent”，Loop 反而是最不重要的那一环</title><link>https://zegging.github.io/posts/agent/eval/how-to-run-a-loop/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/eval/how-to-run-a-loop/</guid><description>在 2026 年的 6 月份，Agent 和软件开发程序员中又出现了一个流行词：Loop Engineering。这个概念主要由 OpenClaw 创始人 Peter Steinberger 和 Claude Code 创作者 Boris Cherny 提出：你不再需要给 AI 写死板的提示词，而是应该设计一个可以自动循环运行、自我测试与验证的智能 Loop。</description><pubDate>Thu, 25 Jun 2026 00:23:00 GMT</pubDate></item><item><title>在 Ubuntu 裸机上部署一个 self-host 的 E2B 服务</title><link>https://zegging.github.io/posts/agent/sandbox/self-host-e2b-on-ubuntu-bare-metal/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/sandbox/self-host-e2b-on-ubuntu-bare-metal/</guid><description>一次从 0 开始最终从 Windows 远程创建 E2B sandbox 的完整部署记录。</description><pubDate>Mon, 22 Jun 2026 09:31:00 GMT</pubDate></item><item><title>作为研发直接负责具体业务的感受：SEO 业务与产品的价值</title><link>https://zegging.github.io/posts/product/seo/seo-business-and-product-value-from-engineering-view/</link><guid isPermaLink="true">https://zegging.github.io/posts/product/seo/seo-business-and-product-value-from-engineering-view/</guid><description>从研发视角理解 SEO：它不是零散的页面优化工单，而是把搜索需求持续转化为产品认知、自然流量、试用注册和商业线索的增长系统。</description><pubDate>Tue, 19 May 2026 16:00:00 GMT</pubDate></item><item><title>从 UTF-8 到 Token：你的 Prompt 进入 LLM 前发生了什么</title><link>https://zegging.github.io/posts/llm/tokenization/from-utf8-to-token-the-bpe-pipeline-of-llm/</link><guid isPermaLink="true">https://zegging.github.io/posts/llm/tokenization/from-utf8-to-token-the-bpe-pipeline-of-llm/</guid><description>把字节、码点、UTF-8、BPE、Regex 预切分、特殊词表、tiktoken 串成一条线，看清一段人类可读的 prompt 是如何被压缩、切割、编号，最终变成 LLM 唯一能消费的 token 序列。</description><pubDate>Tue, 24 Mar 2026 18:13:00 GMT</pubDate></item><item><title>谢赛宁的启示与世界模型的布道</title><link>https://zegging.github.io/posts/agent/podcast/the-idea-of-training-yourself-&amp;-why-the-world-model-is-different/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/podcast/the-idea-of-training-yourself-&amp;-why-the-world-model-is-different/</guid><description>基于2026年2月由张小珺主持、以中文发表的对谢赛宁长达七小时的访谈。在完成这段马拉松 podcast 后我有了一些混沌的启示与感悟。</description><pubDate>Sat, 21 Mar 2026 08:30:00 GMT</pubDate></item><item><title>一个 sitemap 校验工具是怎么被 Codex 做出来的</title><link>https://zegging.github.io/posts/agent/seo/sitemap-verify-with-codex/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/seo/sitemap-verify-with-codex/</guid><description>从真实的 sitemap.xml 校验需求出发，记录一个完全由 Codex 创建的项目，以及 coding agent 带来的新开发方式。</description><pubDate>Thu, 19 Mar 2026 08:30:00 GMT</pubDate></item><item><title>什么是 Agent Harness？</title><link>https://zegging.github.io/posts/agent/harness/what-is-agent-harness/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/harness/what-is-agent-harness/</guid><description>基于 Claude Code 和行业标准（Agent Skills），从零开始编写第一个 skill。</description><pubDate>Mon, 16 Mar 2026 09:30:00 GMT</pubDate></item><item><title>Langfuse 前端显示和实际存储不一致，让我白排查了一小时</title><link>https://zegging.github.io/posts/agent/langfuse/langfuse-ui-display-vs-stored-data-mismatch/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/langfuse/langfuse-ui-display-vs-stored-data-mismatch/</guid><description>合并代码后 Langfuse 测试集报告 tavily_search 失败率 100%。UI 上明明显示 tool call args 是数组 []，Pydantic 却报收到字符串 &apos;[]&apos;。最终发现是 Langfuse 前端渲染时把 JSON 字符串 &quot;[]&quot; 美化成了数组 []。</description><pubDate>Tue, 24 Feb 2026 10:00:00 GMT</pubDate></item><item><title>如何编写一个 skill</title><link>https://zegging.github.io/posts/agent/skills/skills-how-to-write-a-skill-on-your-own/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/skills/skills-how-to-write-a-skill-on-your-own/</guid><description>基于 Claude Code 和行业标准（Agent Skills），从零开始编写第一个 skill。</description><pubDate>Thu, 12 Feb 2026 16:00:00 GMT</pubDate></item><item><title>SKILL 与 MCP 的区别与应用</title><link>https://zegging.github.io/posts/agent/skills/skills-the-different-between-skills-and-mcps/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/skills/skills-the-different-between-skills-and-mcps/</guid><description>Skill 负责“怎么做”，MCP 负责“能连什么”。它们互补，不互斥。</description><pubDate>Thu, 12 Feb 2026 16:00:00 GMT</pubDate></item><item><title>opencode Skill 功能实现 - 渐进式披露</title><link>https://zegging.github.io/posts/agent/skills/skills-opencode-progressive-disclosure/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/skills/skills-opencode-progressive-disclosure/</guid><description>从实现细节到上下文虚拟化：opencode 如何用 Progressive Disclosure 把 Context 成本打到最低。</description><pubDate>Thu, 12 Feb 2026 16:00:00 GMT</pubDate></item><item><title>SKILL 的演进与终局</title><link>https://zegging.github.io/posts/agent/skills/skills-evolution-and-endgame/</link><guid isPermaLink="true">https://zegging.github.io/posts/agent/skills/skills-evolution-and-endgame/</guid><description>Skill 会从“静态死板的文档”变成“动态智能的资产”，并最终融化在 AGI 的行动本能里。</description><pubDate>Thu, 12 Feb 2026 16:00:00 GMT</pubDate></item><item><title>从一次 LinkageError 看 ClassLoader 隔离与开闭设计</title><link>https://zegging.github.io/posts/java/java-classloader-linkageerror-open-closed-design/</link><guid isPermaLink="true">https://zegging.github.io/posts/java/java-classloader-linkageerror-open-closed-design/</guid><description>一次 fat jar 部署引发的 LinkageError 排查，牵出 Java ClassLoader 机制、Flink 类隔离设计，以及平台化系统的开闭原则。</description><pubDate>Thu, 15 May 2025 16:00:00 GMT</pubDate></item><item><title>AWS S3 回源配置踩坑：从 x-oss-process 到 CloudFront Function</title><link>https://zegging.github.io/posts/aws/aws-s3-origin-x-oss-process-cloudfront/</link><guid isPermaLink="true">https://zegging.github.io/posts/aws/aws-s3-origin-x-oss-process-cloudfront/</guid><description>记录一次把对象存储同步处理能力迁移到 AWS S3 + CloudFront 时遇到的问题、排障过程、方案取舍和最终架构。</description><pubDate>Tue, 11 Mar 2025 16:00:00 GMT</pubDate></item><item><title>AWS 回源新架构：一次救火中完成的静态资源动态出海</title><link>https://zegging.github.io/posts/aws/aws-static-resource-dynamic-oversea-origin-architecture/</link><guid isPermaLink="true">https://zegging.github.io/posts/aws/aws-static-resource-dynamic-oversea-origin-architecture/</guid><description>记录一次临时救火中基于 CloudFront Origin Group、CloudFront Function 与 S3 404 日志异步消费，实现静态资源按需同步到海外 S3 的架构设计。</description><pubDate>Tue, 31 Dec 2024 16:00:00 GMT</pubDate></item></channel></rss>