大家好 我是为什么叫QQ

Hello everyone, I'm WhyCallQQ.

一个程序员 一个月往生产环境合入2500个PR

One engineer merged 2,500 PRs into production in a single month.

她睡觉时 Agent还在改代码 跑验证 合并分支

While she slept, AI agents were still modifying code, running verifications, and merging branches.

早上醒来 她再看提交历史

When she woke up in the morning, she simply reviewed the commit history.

听到这里 我第一反应是 评测又卷出新高度了

Hearing this, my initial reaction was: benchmark hype has hit an all-time high.

看完Matt Pocock和Lauren Tan这场访谈

Yet after watching this deep interview between Matt Pocock and Lauren Tan,

我反而没那么在意2500了

I found myself caring much less about the headline number 2,500.

Lauren也说 这些PR包含大量代码园艺

Lauren herself admitted that these PRs consisted heavily of code gardening—

不能换算成2500个新功能

they cannot be equated to 2,500 brand-new features.

真正值得研究的是

What truly warrants our deep study is:

她怎样把人工纠正变成代码库里的检查

how she transformed manual corrections into repository checks,

工具和约束

deterministic tools, and structural constraints.

今天这期Reaction 我想聊聊最有用的部分

In today's reaction teardown, I want to explore the most actionable takeaways—

以及我不会照抄的地方

as well as the parts I would definitely not copy blindly.

先补个人物背景

First, a quick background on the guests:

Matt Pocock是Total TypeScript和AI Hero的作者

Matt Pocock is the creator of Total TypeScript and AI Hero,

曾在Vercel做开发者倡导 也参与过XState

formerly a Developer Advocate at Vercel who also contributed to XState.

他擅长把类型系统和Agent

He excels at turning complex type systems and agentic

工作流讲成能练的东西

workflows into hands-on, practicable engineering skills.

Lauren Tan网名poteto

Lauren Tan, known online as @poteto,

做过Meta React团队 Netflix和Cursor

has worked on the Meta React team, Netflix, and Cursor.

公开身份包括React Compiler核心团队成员

Her public roles include being a core member of the React Compiler team,

以及SpaceXAI的Grok Bot和Cursor相关工作

as well as contributing to Grok Bot at SpaceXAI and Cursor tooling.

pstack是她开源的Agent工程技能库

pstack is her open-source agent engineering skill harness.

视频标题写了SpaceX 严格说是SpaceXAI

The video title mentions SpaceX, but strictly speaking it's SpaceXAI—

别把她想成在火箭部门合并代码

so don't picture her merging code in the orbital rocket department.

两人的交汇点很清楚：

The intersection between them is crystal clear:

把工程经验写进Agent的工作环境

codifying production engineering wisdom directly into the agent's operating environment.

我看访谈时带着一个问题：

I went into this interview with one core question:

团队已经能批量生成代码

Now that engineering teams can generate code in bulk,

工程师还应该把时间花在哪里？

where should human engineers actually spend their precious time?

很多人会觉得 继续开Agent

Many people think: just spin up more agents—

任务拆细 队列拉长 产量总会上去

break tasks smaller, extend queues, and throughput will naturally climb.

Lauren的路径刚好相反

Lauren's path is the exact opposite.

她先减少单个Agent需要的微操

She first minimizes the micromanagement required for each individual agent,

让它拿到运行结果 再扩大并行规模 顺序很重要

ensures it can observe real execution results, and only then scales concurrency. Sequence is everything.

缺少验证时 十个Agent会给你十份待检查的作业

Without automated verification, ten agents simply hand you ten sets of unvalidated homework.

你依然坐在屏幕前点页面 看日志 复制报错

You still sit in front of the screen clicking UIs, parsing logs, and copy-pasting errors.

代码生成变快了 验收队列也更长

Code generation accelerates, but the review and verification bottleneck lengthens.

把你的看法打在公屏上：你会多开几个Agent

Share your thoughts in the chat: would you spin up more parallel agents,

还是先补一个Agent的验证？

or first reinforce verification for a single agent?

Lauren早期优化Cursor Agents窗口性能时

When Lauren was optimizing Cursor Agents window performance in the early days,

用了一个词 meat proxy 人肉代理

she coined a vivid phrase: 'meat proxy.'

Agent改完代码 她打开Chrome DevTools

When the agent finished tweaking code, she would open Chrome DevTools,

看火焰图和堆快照 再把结果喂回去

inspect flamegraphs and heap snapshots, and feed the findings back to the AI.

很多程序员今天还在做同样的事

Many developers are still doing the exact same thing today:

AI说修好了 你启动项目；页面报错 你复制终端；

AI claims it's fixed, so you launch the app; page crashes, so you copy the terminal output;

接口失败 你截图 然后再问它一次

an API fails, so you take a screenshot and ask the model again.

你看起来在指挥AI 实际在给AI当外设

You look like you're orchestrating AI, but in reality you've become a peripheral for it.

Agent看不到真实产物 只能根据转述继续猜

The agent cannot observe real runtime outputs, so it can only keep guessing based on your hearsay.

每次猜测都占用你的注意力

Each blind guess consumes your cognitive focus.

找到这类人肉代理环节

Identifying these meat-proxy choke points

比收藏十条提示词更有价值

is infinitely more valuable than collecting ten fancy prompt templates.

我最推荐反复看16分14秒附近的verification

I highly recommend rewatching the verification discussion around 16:14.

Lauren把它形容成给Agent手和眼

Lauren describes it as giving the agent hands and eyes:

让它运行代码 像用户一样点击页面

empowering it to execute code, click through pages like a real user,

读取trace和快照 再根据结果继续修改

read traces and snapshots, and iteratively revise based on actual runtime results.

这样的loop才有反馈

Only such a closed loop provides genuine feedback.

修前端bug时 可以给Agent一条固定命令

When squashing frontend bugs, you can equip the agent with a standardized command:

启动应用 打开路由 执行复现步骤

spin up the app, navigate to the route, execute the exact reproduction steps,

保存控制台错误和截图

and capture console errors alongside screenshots.

修性能问题时 再补一份可比较的记录

When tackling performance bottlenecks, supply a repeatable before-and-after baseline profile.

目标很具体：Agent宣布完成前 必须拿出运行证据

The goal is crystal clear: before an agent declares victory, it must furnish runtime proof.

测试通过只覆盖测试写到的行为

Passing unit tests only validates the scenarios explicitly authored in those tests.

真实应用里的点击和观测能补上另一块盲区

Live UI interaction and runtime observability illuminate the remaining blind spots.

验证能增强信任 却不会自动带来好设计

Automated verification builds confidence, but it does not magically produce good architecture.

Agent可以证明按钮能点

An agent can easily prove a button is clickable,

接口能返回 性能数字变低

an API endpoint responds, or latency numbers drop—

也可能把代码塞进错误的位置

while simultaneously shoving logic into the wrong layers

把临时方案扩散到整个仓库

and sprawling quick-and-dirty patches across the entire repository.

工程验收要分开两类问题：

Engineering acceptance must cleanly decouple two distinct questions:

行为是否符合预期 结构是否方便下次修改

Does the behavior meet requirements, and is the structure maintainable for future changes?

前者可以交给脚本 测试和浏览器

The former can be delegated to scripts, test suites, and headless browsers.

后者仍需要类型设计 模块边界和人的判断

The latter still demands thoughtful type design, modular boundaries, and human engineering discernment.

一个PR能跑起来 只说明它跨过行为门槛

A PR that runs green only proves it cleared the behavioral bar.

它是否值得留下 还要看维护者半年后能不能看懂

Whether it is worth keeping also depends on if maintainers can understand it six months later.

下次需求变化时能不能安全地改

And whether it can be safely modified when requirements change next time.

第二个让我想立刻动手的地方

The second takeaway that made me want to act right away

是她对确定性的处理

is how she handles determinism.

Agent每次验证都临时写脚本

If an agent writes throwaway ad-hoc scripts for verification every time,

写完又丢掉 下个Agent再重新发明一遍

and discards them afterwards, the next agent ends up reinventing the wheel.

这既慢 也让结果难以比较

This is slow and makes results difficult to compare.

Lauren把重复步骤收进CLI和Skill Playwright

Lauren packages repetitive steps into CLIs and Playwright skills,

Chrome DevTools协议

using the Chrome DevTools Protocol,

截图和trace都由同一套工具处理

where screenshots and traces are handled by the same unified tooling.

Agent保留需要判断的部分 机械步骤交给程序

The agent reserves its reasoning for judgment calls, offloading mechanical steps to programs.

你也可以翻聊天记录：哪些命令连续贴了三周

You can check your own chat logs: which commands have you pasted for three weeks straight?

哪些检查每次都要重新解释

Which checks do you have to explain from scratch every time?

把它们变成短命令 模板或仓库脚本

Turn them into short commands, prompt templates, or repository scripts.

沉到工具里的规则

Rules codified into deterministic tooling

团队都能复用 失败也更容易定位

can be reused across the team, making failures far easier to pinpoint.

Lauren团队还有内部框架Dune

Lauren's team also has an internal framework called Dune,

功能有固定目录 lint规则很严格

with fixed feature directories and very strict linter rules.

很多改动只剩一种常见写法

Many changes are left with only one canonical idiom.

早期Grok Bot桌面端曾有八个上万行的超大文件

Early in the Grok Bot desktop app, there were eight massive files exceeding 10,000 lines each.

Agent很容易顺着坏模式继续加代码

Agents easily latch onto bad patterns and append even more code.

她会拆分结构

She refactors and decomposes the structure,

再把重复错误转成lint 类型或目录约束

then encodes recurring mistakes into lint rules, types, or directory constraints.

我觉得这比继续追加AGENTS.md更有效

I find this far more effective than endlessly appending to AGENTS.md.

提示词要求Agent记住一百条规矩

Prompting an agent to memorize a hundred rules in natural language

代码库可以直接让部分错误无法通过检查

pales in comparison to a codebase that mechanically rejects certain classes of errors.

普通项目也能禁止跨层import

Even everyday projects can ban cross-layer imports,

限制旧API 给危险状态加类型区分

restrict deprecated APIs, and use nominal typing to distinguish dangerous states.

评审里重复出现的评论 都可以考虑交给工具

Any recurring code review comment should be considered for mechanical enforcement.

单个Agent能验证 代码库也有约束 才轮到规模

Only when individual agents can verify work and the codebase enforces guardrails can you scale up.

Lauren把工作分成inner loop和outer loop

Lauren divides the workflow into an inner loop and an outer loop.

内循环负责改代码 验证和交付

The inner loop edits code, runs verification, and delivers changes.

外循环从Slack Linear和用户反馈里带入新信息

The outer loop brings in fresh context from Slack, Linear, and customer feedback.

过去她看到bug再复制给Agent 自己成了搬运工

Previously, she would spot bugs and copy-paste them to the agent, acting as a human meat proxy.

现在Grok Bot订阅外部信号

Now Grok Bot directly subscribes to external signals,

Cursor Projects里的协调Agent拆分任务

and orchestrating agents in Cursor Projects break down tasks

管理子Agent

and manage specialized sub-agents.

Cursor官方也说明 协调Agent本身不写代码

Cursor's team also clarified that the orchestrating agent doesn't write code itself;

主要规划和委派 这条链路才解释了高PR数量

it focuses on planning and delegation. This pipeline explains the high PR velocity.

触发来自真实工作 执行带有共享上下文

Triggers originate from actual work, and execution carries shared context,

人不用手动创建几千次聊天

eliminating the need for humans to manually spin up thousands of chat sessions.

访谈里有个容易略过的细节

There is an easily overlooked detail in the interview:

Lauren让Agent持续扫描React里的危险写法

Lauren has agents continuously scan for anti-patterns in React code.

扫描结果先追加到文档 不立即逐个修

Scan results are appended to a tracking document first rather than fixed piecemeal immediately.

她隔几天再读 经常发现许多记录指向同一个根因

Reviewing it a few days later often reveals that dozens of findings share a single root cause.

这个缓冲队列很反直觉

This buffer queue is highly counterintuitive.

自动化越快 人越容易把

The faster automation runs, the easier it is for humans

每个信号都当成独立工单

to treat every isolated signal as a standalone ticket,

结果重复修同一个问题 还错过更高层的模式

fixing the same underlying issue repeatedly while missing higher-level patterns.

团队可以给自动修复加一道聚合步骤

Teams can introduce an aggregation step before running automated remediation.

先收集 去重 按模块和症状分组 再决定是否开工

First collect, deduplicate, group by module and symptom, and only then schedule execution.

协调Agent也能同时看到多份报告

An orchestrator agent can also analyze multiple symptom reports simultaneously.

保留一个让人看全局的队列 反而能减少无效PR

Preserving a human-visible triage queue actually slashes redundant and conflicting PRs.

谈到2500个PR怎样评审

When asked how anyone could possibly review 2,500 PRs,

Lauren的回答是sampling 抽检

Lauren's answer was unambiguous: sampling.

她抽查已经合入的代码

She performs audit sampling on already merged code.

发现多个Agent重复走捷径

If she spots multiple agents repeatedly cutting corners,

就回去修改Skill lint 类型和目录结构

she goes back to update the Skill prompt, linter rules, type definitions, and folder architecture.

她从逐份放行

She shifted from acting as a tollbooth inspector

转向检查系统能否持续产生合格结果

to inspecting whether the generative system can reliably yield compliant outcomes.

这个争议点值得打在公屏上：

This controversial debate is worth airing in the comments:

事后看main 能不能算评审？

Does auditing main after the fact truly count as code review?

我的判断取决于改动风险和回滚成本

My assessment hinges entirely on change risk and rollback cost.

低风险 可回滚 验证充分的工作 抽检能减少阻塞

For low-risk, easily revertible, well-verified tasks, sampling eliminates deployment gridlock.

高风险改动仍需前置审查

High-risk changes still strictly require upfront human gatekeeping.

抽检也不能只看diff是否漂亮

Sampling cannot just evaluate whether the diff looks aesthetically pleasing;

还要结合线上错误 回滚记录和用户反馈

it must correlate with production telemetry, rollback logs, and user feedback.

否则你只是推迟评审 问题会在更贵的地方出现

Otherwise, you are merely postponing review, and defects will surface where remediation costs 10x more.

Lauren讲到Agent在她睡觉时自动合并

Lauren highlighted agents automatically merging branches while she slept.

单剪这一段 很容易变成全自动编程神话

Taken out of context, this clip easily morphs into a myth of fully autonomous software engineering.

放回整场访谈 门槛很高

Restored to the full interview context, the bar is exceptionally high.

项目要有验证Skill 代码库要持续清理

The project must possess robust verification skills, the repository requires continuous pruning,

常见错误要被约束 PR还会经过多个Verifier

common pitfalls must be strictly bound, and PRs pass through multiple independent verifiers,

Agent反复运行和模糊测试

undergoing repeated execution and property-based fuzzing.

pstack公开文档也要求重新检查变化后的补丁

pstack's documentation explicitly requires re-verifying rebased patches

并结合CI状态 她也直说这套流程很耗token

against green CI status—and she openly admitted this pipeline consumes immense token budgets.

我不会建议团队第一天就打开自动合并

I would never advise any team to enable auto-merge on day one.

更稳妥的顺序是先让Agent准备可审查的PR

A far safer progression is first having agents prepare review-ready PRs with rich evidence,

再开放固定类型的低风险改动

then granting auto-merge only for well-scoped, low-risk change classes,

观察回滚率和漏检 慢慢扩大权限

carefully tracking rollback rates and escaped defects before expanding permissions.

Matt追问医疗 金融等高风险场景时

When Matt probed high-stakes environments like healthcare and fintech,

用了一向门和双向门的比喻

he invoked the classic metaphor of one-way doors vs. two-way doors.

能低成本revert的改动 事后抽检比较容易接受

Changes that can be cheaply reverted through a two-way door are well-suited for post-merge sampling.

会造成数据丢失 资金变化或合规影响的改动

Changes risking irreversible data loss, ledger anomalies, or regulatory non-compliance

自动合并需要另一套证据和审批

cross a one-way door, demanding strict pre-flight proofs and explicit human approval.

Lauren也承认

Lauren readily acknowledged

难以程序化验证的领域还没有现成答案

that domains lacking programmatic verification currently have no silver bullets.

软件有些部分容易验证

Certain parts of software are trivial to verify automatically;

有些部分涉及真实用户 组织流程和监管责任

other parts entangle real users, organizational workflows, and legal accountability.

测试绿灯覆盖不了全部后果

A green CI badge cannot shield you from real-world fallout.

工程师可以给PR标注爆炸半径和回滚路径

Engineers can annotate every PR with its blast radius and explicit rollback runbook,

把权限跟风险绑定

tightly binding autonomous authority to operational risk,

让Agent多处理可逆工作

letting agents handle predominantly reversible labor

人把注意力留给不可逆决策

while humans conserve mental focus for irreversible decisions.

看到这里你可能会想 要不要直接安装pstack

At this point you might wonder: should our team just install pstack directly?

可以试 但更值得抄的是她制作Skill的方法

You certainly can try it, but the far richer goldmine is her methodology for crafting skills.

Lauren建议回看历史对话

Lauren recommends auditing your past conversation transcripts with AI:

找出反复纠正Agent的位置

pinpoint exactly where you repeatedly intervened to correct the agent,

再把纠正沉淀成流程 脚本或检查

and distill those recurring corrections into structured workflows, scripts, or linter rules.

她还做了recall

She also built 'recall'

把相似任务里的旧上下文重新整理出来

to retrieve and synthesize historical context from past analogous tasks.

我的建议是 导出最近二十次真实开发记录

My practical recommendation: export your team's last 20 real-world AI coding sessions,

标记每次人工接管：

and tag every single human intervention:

你在补背景 纠正目录 要求重跑 还是推翻设计

Were you supplying missing context, correcting folder paths, demanding reruns, or rejecting flawed designs?

出现三次以上的干预 就是改造候选

Any manual intervention recurring three times or more is an immediate candidate for automation.

别人的Skill提供起点

Someone else's skill gives you a useful starting baseline;

你的聊天记录暴露团队真正的摩擦

your own chat logs expose your team's authentic production friction.

模型能力变化后 也可以删掉多余的实现细节

As frontier models evolve, you can safely strip away transient prompt scaffolding

保留稳定的流程

while preserving the durable engineering workflow.

听完访谈 我特别想做的一个事情：

After absorbing this interview, here is the #1 initiative I want to kick off tomorrow:

挑一个每周重复五次的验收环节

Pick one manual verification workflow you repeat five times a week,

让Agent自己完成并留下证据

and require the agent to complete it autonomously while leaving verifiable evidence.

可以是一条复现命令 一组浏览器步骤

It could be a reproduction CLI command, a set of automated browser steps,

或者一份性能基线

or a deterministic performance benchmark baseline.

跑稳以后 再把重复错误写进lint和类型

Once that runs reliably, codify recurring slip-ups into linter rules and type invariants.

接着接入真实任务来源 给自动修复加聚合队列

Next, connect real incident sources and add a deduplicating triage queue for auto-remediation.

等数据证明它可靠 再讨论并行和自动合并

Only when empirical data proves reliability should you expand concurrency and discuss auto-merge.

未来程序员会在环境设计

In the future, software engineers will spend vastly more time on environment design,

验证系统和风险分级上花更多时间

verification harnesses, and risk classification tiers.

领域知识也能帮助你写清目标和验收标准

Deep domain expertise will empower you to formulate crisp goals and unambiguous acceptance criteria.

你在工作流里当过哪种人肉代理？

Which flavor of 'meat proxy' have you found yourself playing in your workflow?

欢迎把具体场景留在评论区

Share your concrete scenarios in the comments below.

感谢观看 如果这期帮你少踩一个坑

Thank you for watching! If this episode saved you from stepping into a costly trap,

记得一键三连 我是为什么叫QQ 我们下期接着聊

don't forget to like, subscribe, and share. I'm WhyCallQQ—let's keep digging in the next episode!