AI 原生工程笔记AI Native Engineering
AI 原生工程笔记AI Native Engineering

关于About

不再亲手写代码之后After I stopped writing code by hand

现在在做Now

变化是彻底的。现在 95% 的时间我不再亲手写代码，日常重心放在工程架构、模块设计和边界把控上：需求怎么拆、上下文怎么给、结果怎么验收。The change was total. Today I don't write code by hand 95% of the time. My days go into architecture, module design and boundaries: how to split the requirement, what context to give, how to accept the result.

最近在关心的事What I'm focused on lately

  • Agent 工程Agent engineering
  • 企业工作流Enterprise workflow
  • 上下文与评估Context and evaluation
  • 开源运行时Open-source runtimes

走过的路The road so far

  1. 传统后端Backend

    实习期As an intern

    做传统后端开发，做过高并发系统，也系统学过 DDD。I built conventional backends, worked on a high-concurrency system and studied DDD properly.

    Java · 分布式系统 · 可靠性Java · distributed systems · reliability

  2. AI 辅助AI-assisted

    过渡期In between

    AI 先以工具的身份进到日常开发里：写 Prompt、接 RAG、用编程副驾补代码。AI came into daily work as a tool first: writing prompts, wiring up RAG, letting a coding copilot fill in code.

    Prompt · RAG · 编程副驾Prompts · RAG · coding copilots

  3. AI 原生AI-native

    正式工作一年多后A year or so into my first job

    编程语言变得不那么重要，前后端的分工也是；也不再需要花大量时间去追框架和工具的更新。Languages matter less, and so does the front-end/back-end split; I no longer spend much time chasing framework and tool updates.

    上下文 · 运行时 · 评估 · 系统工程Context · runtime · evaluation · systems

这段经历让我确信一件事：真正沉淀下来的不是某门语言或某个框架的熟练度，而是可迁移的工程判断：怎么划模块、怎么定边界、怎么让系统可验证。What this taught me: what lasts isn't fluency in one language or framework but engineering judgment that transfers: where to draw modules, where to put boundaries, how to make a system verifiable.

工具会换，这些不会。The tools will change. These won't.

这个博客记录的就是这些判断。它们来自真实项目，也应该能直接用于下一个项目。This blog is where I write those judgments down. They come from real projects, and they should be usable on the next one.

一个任务是怎么跑完的How a task gets done

意图、上下文、计划、构建、验证，五步走完一个任务。点任意一步，它就停下来等你看。Intent, context, plan, build, verify: five steps through one task. Pick any step and it stops there for you.

图 1Fig. 1一个示例任务：从一句需求，走到能拿去发布的证据。A sample task: from a one-line request to evidence you can ship on.示例数据Sample data
Agent 工作台的五步The five steps of the Agent Workbench

把“做个智能审批”压成可验收目标Turn “build a smart approval flow” into goals you can check

先把业务结果、人工边界和失败条件写清楚。交给 Agent 的，是一组能被证伪的工程假设。Pin down the outcome, the human boundaries and the failure conditions first. The agent gets falsifiable hypotheses, not a wish.

  1. 营销审批 P50 从 11 小时降到 4 小时内Approval P50 from 11 h to under 4 h
  2. 高风险决策必须保留人工确认High-risk decisions keep a human sign-off
  3. 每次建议都能回放上下文与依据Every suggestion can be replayed with its context and reasons

当前证据Evidence so far3 条验收约束已锁定3 acceptance checks locked

brief.mdmarkdown

goal: 缩短营销审批周期goal: shorten the approval lead timeguardrail: 不覆盖人工结论guardrail: never override a human decisionsuccess:  lead_time_p50: < 4h  trace_coverage: 100%

第一步，共 5 步Step 1 of 5：意图: Intent

自动播放Autoplay

在别处Elsewhere