Skip to content
Back/Product Trends

Build, then align — Agent 时代的产品流程颠覆

View in Graph
Updated 2026-11-11
4 min read
805 words

Build, then align

What it is

Build, then align is a product-development loop that aligns the team on the user problem first, builds a working solution at an appropriate fidelity, gathers feedback from the shared artifact, and then aligns on the solution. It moves human-to-human agreement about implementation after an initial build; clear instructions between people and agents can still come before coding.

Why it matters

当 agent 降低构建原型的时间与成本,过早争论实现方案可能比先做出可体验版本更慢。Zach Lloyd、Matt Pocock 和 Paweł Huryn 的帖子都强调以具体产物或真实交互收集反馈;@every 另称其团队一个周末为 Checks 评测平台开出 75 个 PR,并把 backlog 视为尚未作出的产品决策。后者提示实现吞吐上升时,产品判断可能成为约束;这些实践者自述没有对照数据,不能证明普遍规律。

Key points

  • 先对齐问题与用户目标,再用可运行的版本让团队讨论同一个对象;方案选择可在初次构建后进行。
  • 原型应达到足以检验关键判断的 fidelity。Matt Pocock 认为,agent 让做原型和收集早期反馈更容易;文字规划无法展示真实实现。
  • Paweł Huryn 把人或 agent 使用真实产品与真实数据视为重要证据,并认为会议、实验和文档不能完全替代这种交互。此为作者判断,未提供对照研究。
  • 对人和 agent 的任务边界、约束与验收标准仍要在执行前说清;后置的是人和人对解决方案的对齐。
  • 吞吐增加会暴露决策待办:@every 报告一个周末为 Checks 评测平台开出 75 个 PR,并称 backlog 越来越像尚未作出的产品决定。该公司单一案例未衡量 PR 质量、周期或用户结果。

旧世界的对齐为什么必要

写代码是最贵、最慢的环节,所以在动手前要确保所有 stakeholder(开发、设计、PM、senior 工程师、Leadership)对"要做什么"达成一致。两个目的:

  1. 第一次就尽可能做对
  2. 避免"以为在做 X 结果做出来是 Y"——这种 stakeholder surprise 会带来巨大延迟

为了对齐,需要会议、Slack 长串、Loom 录屏——昂贵,但比"做错了重来"便宜。

新世界为什么这套流程崩了

编码瓶颈正在消失。当 build 变得极快极便宜,前置对齐的成本反而成了瓶颈。

"Once something is built, you are no longer guessing. You are experimenting. What's more, you are all experimenting on the same thing, not on some hypothetical implementation of a thing which every stakeholder is imagining working in a slightly different way."

关键洞察:脑子里想象的产品和实际跑起来的产品永远不一样。每个 stakeholder 脑中那个"实现"都略有不同——前置对齐其实是在对齐一堆幻觉。

流程对比

旧流程(Warp 的 How We Work guide)

  1. 明确问题和用户目标
  2. 团队探索解决方案空间(product jam)
  3. 设计多个方案 mockup(设计瓶颈)
  4. 团队对齐选哪个方案(团队对齐瓶颈)
  5. 写代码(编码瓶颈,正在消失)
  6. review 是否符合 spec(验证瓶颈)
  7. 内部 dogfood 迭代
  8. Bug bash
  9. Ship

新流程

把第 4 步(团队对齐)移到第 5、6 步之后。

仍然要做问题定义、product jam、设计第一版猜测方案——但目标是尽快启动一个基于实物的反馈循环,而不是争论哪个方案最好。

旧:  问题 → 方案争论 → 对齐 → 写代码 → 验证 → 迭代
新:  问题 → 初版设计 → 写代码 → 体验 → 对齐 → 迭代
                              ↑
                       低成本 build = 真实样本

实施方式

  • 不要纠结第一版是否完美 — it doesn't matter
  • 想做几个版本就做几个版本,让用户试
  • 别浪费 cycle 猜,build it 然后体验真实方案
  • "Alignment should come after we have solutions in-hand"

Warp 的反例与正例

反例(2023):做 Warp Drive(协作平台)时,花了无数小时在 mockup、Slack、PRD 上对齐才动手。Zach 自评:"I would not do this again."

正例(2026 最近):发布 coding agent 套件时,只在"问题"上对齐(让 Warp 成为用任何 coding agent 的最佳地方),但不 bikeshed 方案。设计团队出初版设计 → 团队 build → 一起迭代 → 上线时自然对齐。

一个重要边界

Zach 强调:他说的 alignment 是人和人对齐(humans aligning with each other on how a product should work),不是人和 agent 对齐(humans aligning with agents on how to build)。后者他完全支持,参见他自己的 Spec & Verify 文章。

换句话说:

  • 人对人 alignment → 后置(先 build)
  • 人对 agent alignment → 前置(写好 spec)

Evidence across sources

Source Key Claim Relevance
Build then align — Zach Lloyd, 2026-05-07 Align on the problem, build a first version, then align on a concrete solution Origin of the workflow and its people-to-people boundary
AI Briefing 2026-11-10 Evening — @mattpocockuk, tweet 2120100383527682165 At an appropriate fidelity, working prototypes and early feedback can outperform extended text-only planning Adds agent-enabled prototype speed as a reason to move solution discussion later
AI Briefing 2026-11-10 Evening — @PawelHuryn, tweet 2120147957525700919 Real people or agents using a product with real data provide evidence that meetings, experiments, or documents alone cannot supply Adds use in context as a feedback source; the post is a practitioner judgment
AI Briefing 2026-11-11 Evening — @every, tweet 2120501127884275739 Reports 75 PRs for Checks over one weekend and frames the backlog as product decisions still to make A single-company signal that implementation capacity may shift attention toward product judgment; not a benchmark

Counterpoints & Gaps

  • The supporting material is a small set of practitioner posts and company self-reports; it does not independently verify the 75-PR count or measure cycle time, product quality, or user outcomes against a document-first process.
  • The posts do not establish that every prototype should precede planning. Regulated, safety-critical, or costly-to-change work may need more analysis before implementation.
  • Zach Lloyd distinguishes human-to-human solution alignment from human-to-agent specifications; the latter can remain explicit before a build.

Open questions

  • Does "build then align" apply to regulated industries where post-build changes carry compliance costs?
  • How does team size affect the viability of this workflow? Warp is a startup; does it scale to 500-person product teams?
  • When coding is no longer the bottleneck, what becomes the new bottleneck — design judgment, user research, or distribution?

Prompts for witness

  • Which current project has a decision that could be tested with a small working prototype instead of more text planning?
  • What minimum fidelity would let a real user or agent test the central assumption safely?
  • Which decisions must be specified to an agent before it starts, even if the team can align on the implementation afterward?

Pending Jean review

  • product-trends/build-then-align-warp — ai_draft; assess whether a real project backlog contains decisions that can be tested with a small artifact, and track user outcomes rather than PR volume alone.

相关主题入口

Sources

Synthesized from 4 sources
  • Build, then align — Zach Lloyd, 2026-05-07Supporting source listed by this page.Whole pagemediumbody
  • AI Briefing 2026-11-10 EveningSupporting source listed by this page.Whole pagemediumbody
  • AI Briefing 2026-11-11 EveningSupporting source listed by this page.Whole pagemediumbody
  • Build then align - Zach Lloyd WarpSupporting source listed by this page.Whole pagemediumabsorb log

Evolution

1 event
  1. absorbed

    Derived from source material

    This page is currently synthesized from 4 sources.

    From Build, then align — Zach Lloyd, 2026-05-07, AI Briefing 2026-11-10 Evening, AI Briefing 2026-11-11 Evening, Build then align - Zach Lloyd WarpTo Build, then align — Agent 时代的产品流程颠覆
    Sources: raw/twitter/x-articles/Build then align - Zach Lloyd Warp.md · raw/briefing/AI Briefing/2026-11-10-22-58.md · raw/briefing/AI Briefing/2026-11-11-23-00.md · raw/to-learn/Build then align - Zach Lloyd Warp.md

Linked from