<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>文章 · Henson的个人站</title><description>朱亚威（Henson）：开放平台产品经理，写 AI Agent、API 生态、开发者体验与产品判断。</description><link>https://zhuyawei.com</link><language>zh-CN</language><atom:link href="https://zhuyawei.com/rss.xml" rel="self" type="application/rss+xml"/><item><title>用 Vibe Coding 完成 Personal Agent 一期</title><link>https://zhuyawei.com/blog/personal-agent-phase-one/</link><guid isPermaLink="true">https://zhuyawei.com/blog/personal-agent-phase-one/</guid><description>从模型与编码智能体路由、单一时间线与自动会话、受治理的财务 MCP，到真实 iPhone、ECS、备份与恢复，我用 16 天完成了一个只服务自己的 Personal Agent。本文记录最重要的产品决策、生产事故、评测补洞，以及为什么模型说“已经记录”从来不算完成。</description><pubDate>Fri, 07 Aug 2026 00:00:00 GMT</pubDate><category>AI 协作</category></item><item><title>我为什么决定把业余时间 all in 到一个大项目：Personal Agent</title><link>https://zhuyawei.com/blog/all-in-personal-agent/</link><guid isPermaLink="true">https://zhuyawei.com/blog/all-in-personal-agent/</guid><description>从 ChatGPT、豆包、Claude Code 和 Codex，到 iCloud 云盘、GitHub 与多个个人数据源：我为什么不再继续寻找更好用的 AI 工具，而要开始搭建一套自己拥有、能够跨设备延续并对真实行动负责的 Personal Agent。</description><pubDate>Mon, 03 Aug 2026 00:00:00 GMT</pubDate><category>AI 协作</category></item><item><title>大模型 API 是怎么设计的：一次传统 API 视角的逐项拆解</title><link>https://zhuyawei.com/blog/llm-api-design-teardown/</link><guid isPermaLink="true">https://zhuyawei.com/blog/llm-api-design-teardown/</guid><description>我对大模型 API 的认知一度停在「两个字符串进、一段文本出」。真去把 Anthropic、OpenAI 和几家国产平台的文档逐一翻完，才发现今天的大模型 API 是一台无状态的对话回放机，外挂一套内置的工具调用协议。这篇用「传统 API 设计」的眼睛逐项拆解：无状态回放与缓存四模式、stop_reason 与内容块、工具调用的挂起-恢复循环、采样参数被平台收回、API 归拢的四种哲学、协议对象从代码变成模型，以及一场很少被点破的 Anthropic 兼容端点协议战。结论是：外围全是传统 API 的老套路，真正的新物种只有核心那一小块。</description><pubDate>Tue, 21 Jul 2026 00:00:00 GMT</pubDate><category>开放平台</category></item><item><title>从 Prompt 到 Loop：为什么研发更容易自动闭环，产品工作需要让人回来</title><link>https://zhuyawei.com/blog/prompt-context-loop-engineering/</link><guid isPermaLink="true">https://zhuyawei.com/blog/prompt-context-loop-engineering/</guid><description>Prompt 决定这一轮怎么做，Context 决定这一轮知道什么，Harness 决定允许怎样行动，Loop 决定这一轮结束后发生什么。研发拥有测试、编译等硬验证器，产品工作的关键反馈却常在用户、业务和组织现场之外。本文从一次多 Agent 假绿和一个 iCloud 到小红书的个人工作流出发，讨论 Loop Engineering 的逻辑、适用边界，以及人在循环里的正确位置。</description><pubDate>Sat, 11 Jul 2026 00:00:00 GMT</pubDate><category>AI 协作</category></item><item><title>当开放平台面对的 Agent，从开发者变成服务调用方</title><link>https://zhuyawei.com/blog/agent-as-service-caller-open-platform/</link><guid isPermaLink="true">https://zhuyawei.com/blog/agent-as-service-caller-open-platform/</guid><description>当 Agent 开始代表用户直接执行高风险服务操作时，开放平台要补上的不只是权限控制，而是平台可验证的本次意图确认。本文讨论的核心不是 OAuth 失效，而是第三方 Agent 的执行界面对平台变成了不透明信道。</description><pubDate>Wed, 01 Jul 2026 00:00:00 GMT</pubDate><category>开放平台</category></item><item><title>我为什么把 Agent 的记忆搬出聊天窗口</title><link>https://zhuyawei.com/blog/agent-memory-knowledge-base/</link><guid isPermaLink="true">https://zhuyawei.com/blog/agent-memory-knowledge-base/</guid><description>从 Claude Chat、Claude Cowork、Claude Code 到 Codex 双持之后，我越来越确信：Agent 时代真正需要沉淀的不是某个产品里的“记忆”，而是一套可迁移、可审计、可被不同 Agent 共同读取的个人上下文基础设施。我的做法是把知识库和项目 git 拆成两个事实源：知识库负责长期上下文，git 负责执行上下文。</description><pubDate>Sun, 28 Jun 2026 00:00:00 GMT</pubDate><category>AI 协作</category></item><item><title>拆开飞书同一份 API 的两套 Agent 连接方式：MCP 与 CLI</title><link>https://zhuyawei.com/blog/mcp-vs-cli-agent-encapsulation/</link><guid isPermaLink="true">https://zhuyawei.com/blog/mcp-vs-cli-agent-encapsulation/</guid><description>飞书在同一份 OpenAPI 之上，同时做了两套面向 Agent 的连接方式：自动生成 1271 个工具的 lark-mcp，和人工策展十几个域的 lark-cli。我把两套源码、工具描述、REST 映射和 shortcut 层逐段拆开看完之后，结论不是谁替代谁，而是它们分别站在同一条「覆盖 × 质量」前沿的两个位置：MCP 负责低边际成本的广覆盖，CLI + skills 负责高摩擦任务的人工消解。而有趣的是，在飞书这个样本里，20 倍的 star 差距给出了一个清晰的开发者注意力信号：第一层自动封装很难单独形成护城河。</description><pubDate>Thu, 25 Jun 2026 00:00:00 GMT</pubDate><atom:updated>2026-07-10T00:00:00.000Z</atom:updated><category>开放平台</category></item><item><title>从传统开放平台到 AI 时代：当调用方从人变成 Agent</title><link>https://zhuyawei.com/blog/traditional-to-ai-open-platform/</link><guid isPermaLink="true">https://zhuyawei.com/blog/traditional-to-ai-open-platform/</guid><description>开放平台过去十年都在优化「一个有经验的人会怎么用」。当调用方从开发者变成 Agent，衡量什么、设计什么、怎么验证都要换一套。这是在得物国际开放平台（POIZON Open Platform）上，把 Agent 真正放上去端到端跑过之后，踩出来的三段转换——以及一个不浪漫的结论：真正的门槛不在技术。</description><pubDate>Wed, 24 Jun 2026 00:00:00 GMT</pubDate><atom:updated>2026-06-25T00:00:00.000Z</atom:updated><category>开放平台</category></item></channel></rss>