Field notes from building AI-native software. Multi-agent systems, memory, products. Not a tutorial site — a judgment site.https://mind.pages.dev/en-ushttps://mind.pages.dev/writing/cloudflare-top-to-bottom/https://mind.pages.dev/writing/cloudflare-top-to-bottom/大多数 AI 产品跑在 AWS 或 Vercel 上 + 某处的数据库。我把每一层都跑 在 Cloudflare 上——Workers、D1、R2、KV、Vectorize、AI Gateway、Pages。 六个理由它加起来意味着更少的代码和更低的账单，以及两种我不会用它的 情况。Sun, 03 May 2026 00:00:00 GMTPatternshttps://mind.pages.dev/writing/anti-sycophancy-rule/https://mind.pages.dev/writing/anti-sycophancy-rule/LLMs are trained to be helpful, which is to say agreeable. In a multi-agent roundtable that is the failure mode. Five lines, added to every persona, brought back disagreement and made the discussions actually useful.Sun, 03 May 2026 00:00:00 GMTEurekahttps://mind.pages.dev/writing/coach-who-doesnt-flatter/https://mind.pages.dev/writing/coach-who-doesnt-flatter/大多数 AI 产品在为"用户当下感觉好"做优化。一个教练必须为"用户变得更 好"做优化。这两个目标真的会冲突。从选第二个出发的三个产品决策。Sun, 03 May 2026 00:00:00 GMTAI Productshttps://mind.pages.dev/writing/episodic-memory-and-rag/https://mind.pages.dev/writing/episodic-memory-and-rag/RAG remembers what was said. Episodic memory remembers what happened. The difference is whether your system can learn — or just recite. A recipe for adding episodic memory to a RAG-only system in a week.Sun, 03 May 2026 00:00:00 GMTAgents & Systemshttps://mind.pages.dev/writing/five-layers-of-an-agent-system/https://mind.pages.dev/writing/five-layers-of-an-agent-system/人人都把"LLM 加几个工具"叫做 agent。这个词背后藏了太多东西。我把它拆成 五层，每层缺位时的症状写在旁边，结尾给一个 30 秒电梯版。Sun, 03 May 2026 00:00:00 GMTAgents & Systemshttps://mind.pages.dev/writing/memory-as-foundation/https://mind.pages.dev/writing/memory-as-foundation/大多数 AI 产品把记忆当成"撒上去的功能"。好的产品把它当成底座。一个 真正的记忆层需要的六件事，"向量库 + RAG" 给不了你。Sun, 03 May 2026 00:00:00 GMTAgents & Systemshttps://mind.pages.dev/writing/persona-vs-specialist-split/https://mind.pages.dev/writing/persona-vs-specialist-split/Most "multi-agent" systems collapse two roles into one and end up with agents that are both bad at thinking and bad at doing. The split is worth ten frameworks.Sun, 03 May 2026 00:00:00 GMTAgents & Systemshttps://mind.pages.dev/writing/submit-and-poll/https://mind.pages.dev/writing/submit-and-poll/WebSocket 看上去是 AI 长任务的正确答案。它几乎从不是。提交+轮询模式 看上去更丑、生产里更稳定、易于扩展、优雅降级。我停止和 WebSocket 战斗之后，AI 功能的失败率下降了一个数量级。Sun, 03 May 2026 00:00:00 GMTPatternshttps://mind.pages.dev/writing/why-amber-doesnt-have-tabs/https://mind.pages.dev/writing/why-amber-doesnt-have-tabs/Tab-based AI products fragment the user. Chat-based AI products focus the user. Why an AI coach has no navigation, no settings page, no dashboard — and what it took to mean it.Sun, 03 May 2026 00:00:00 GMTAI Products