中文精读
从指挥 AI 到验收结果：王文锋怎样重写 Agent 创业假设
嘉宾：王文锋，Sheet0 创始人（当期身份）｜主持：曲凯
发布：2026-03-14｜时长：57分10秒｜整理：2026-10-18
官方单集
如果工程师同时照看十几个 Agent 窗口，下一个瓶颈还在生成代码吗？王文锋这一期给出的回答，是把人的注意力从分发任务、维护配置转到最后验收，让 AI 负责管理 AI。比新概念更重要的，是他主动修正旧路线：曾追求五至十年后的完全可信，现在先解决未来三至六个月已经发生的阻塞。
以下是2026年3月的创业者判断，不是 OpenClaw 当前功能说明，也不把“所有 Agent 都会成为 Coding Agent”写成已经确定的终局。
一、能力释放与产品包装，要分开看
回听：00:56–13:46。 王文锋把 Manus 和 OpenClaw 视作两次能力释放：前者展示推理模型在研究任务中的用途，后者把编码能力接到更多任务、通信渠道与记忆。IM 接入不是全部解释；核心在于底层编码 Agent 能做什么，以及外层如何配置、组织并呈现。
他对垂直 Agent 的态度因此转变。此前相信长期垂直路线，本期却认为通用编码 Agent 加 Terminal（命令行操作环境）和 Skills（可复用的方法与说明），能复现许多垂直工作流。软件过去通过固定界面复制专家经验，现在也可能通过自然语言目标与可执行方法传播。
这是很强的判断，主持人也追问简单 Skill 是否真能承载复杂行业知识。能得到表单、报告或记录，不等于已替代 SaaS 的权限、数据治理、协作、责任与长期维护。本文不接受“已证明全面替代”的口述作为验收结果。引用报告中的49.7%用例占比，样本与分类亦未外核，不能据此推断全体用户行为。
二、文件系统让错误有机会回改，可靠性仍需另验
回听：19:55–23:07。 上期讨论长链任务时，担心每步不确定性累积，最后成功率很低。王文锋现在强调持久状态：把进度、记忆、任务数据写成文件，模型能看见当前情况，发现错误后回改，而不只沿着内存中的状态向前走。
这个机制有具体价值。长任务中断后，系统不必完全依靠对话回忆；已做什么、哪里出错、哪些内容要更新，可以变得可见。模型能力与运行方法一起进步，不能把全部变化归给更聪明的基模。
但他称修复“已经不是问题”，应保留为个人体感。文件也可能被写错，系统也可能没发现自己错了。编辑推论是：持久化要配版本、结果检查和恢复路径；先验证能否发现错误，再验证是否修对。把状态放进文件只提供了恢复条件，不能独自保证长程任务可靠。
三、主动性有层次，AI 管 AI 需要完整任务闭环
回听：23:28–31:12。 “每天八点发日报”与“发现一个未被提出的问题并建议行动”并不一样。曲凯明确指出，前者仍由人的预设触发；后者要求系统理解业务，主动探索、提出方案，再建立特定 Agent 完成任务。
王文锋把这概括为 AI 管 AI：不仅执行现成指令，还承担需求生成、配置和调度。它不是多个模型互相聊天就自然出现的能力。系统仍需先与业务、文档和工具建立连接，有磨合过程，他也承认安装即自动理解一切还有距离。
经验能否分发同样有条件。Skills 可以保存工作方法，但一家公司围绕代码仓库和流程调好的 Agent，拿到另一家公司未必有用。通用方法与私有上下文需要区分。编辑推论是：自动发现问题可以先生成建议；发消息、改数据、发布等行动则要有相应权限。主动性不是自动授权，也不能为了跨 Agent 学习而默认外传公司内部资料。
四、内部流水线说明机会，也暴露验收责任
回听：31:34–45:36。 王文锋讲述把 Linear 需求管理、GitHub 代码协作与 Slack 讨论连接给 AI：收到任务后实现、测试、提交 PR，人最后检查并合并。以前是人分发需求再指挥 AI，现在尝试让 AI 接住更多中间工作。
他称一周搭通这套流程，个人效率较去年提高十倍，开发新产品两周即可。这些是自家经验与判断，未提供可比较样本；PR 数、commit 数和代码行数不能直接折算真实生产力。早先1600次提交相当小团队一年工作的说法，同样不能证明功能与质量等价。
Harness 被他比作马鞍，是让强能力在具体团队中可控制、可协作的层，既有基础设施属性，也面向最终用户。后端难反馈的问题，可以通过状态工具与数据库只读访问改善；观察状态不需要开放任意写权限。节目中也明确提到了只读权限。
新产品面向创始人、复合型产品经理和 Builder，试图减少多终端切换。人从指挥者转成最后把关者，仍需知道需求是否完成、测试是否有效、是否引入回归。编辑推论是：一个合并按钮不等于审核过程。批准必须有足够证据，才是真正的责任点。
五、“大 C”预算来自工作价值，不能按烧 Token 排名
回听：45:57–51:45。 王文锋提出，个人或小团队若用 AI 承担过去需要雇人的工作，预算可能接近企业客户。他以1000名用户、每年各付10万美元演算1亿美元年收入。这是情景算术，并非已实现 ARR，也不代表所有个人用户都有这类支付能力。
关键在钱替代什么、创造什么。如果原本的工程预算被工具支出部分取代，用户要能看见节省或收入；高消耗也可能只是重试和低效。月约2万美元的团队账单、七人相当过去三五十人效率等，均是嘉宾当期自述。
他主张组织给员工探索空间，也使用了很强硬的淘汰表述。本文不将其写成普遍管理法。编辑推论是：预算应与任务质量、完成量、审阅负担和用户收益一起看，不能先按消耗多寡评价能力。探索可以暂时不追短期回报，但仍应明确范围、复查节点和损失上限。
六、缩短计划，是修正问题选择，不是宣布准确性无用
回听：41:49–42:35、48:24–55:15。 王文锋承认，Sheet0 原本的完全可解释、完全准确目标，模型当时做不到；用户虽然会说想要，但获得其他价值时，绝对准确还不是所有人的首要阻塞。现在看见的真实瓶颈是工程师注意力被多个终端占住，于是顺着需求多做半步。
这直接修正了上期的强可靠路线，不能只说“技术升级所以更乐观”。也不能反向推出准确性已不重要：不同任务后果不同，财务、医疗或其他高风险交付不能用一般效率体验代替质量要求。此为编辑限制。
“回到2025年3月要不要做 Genspark”是他让团队检查差异化执念的问题。预测带来下注，下注错了会提高转向成本；跟随明确信号，可以保留修正空间，但仍需证明自己的用户价值，竞品融资不是 PMF。
他还担心就业与需求可能形成负反馈，因此不愿做过远预判。这是宏观忧虑，未经本篇验证，不能作为已确立经济机制或个人职业建议。短周期应体现在更快得到证据、可逆地调整，不是免除结果责任。
练习：把“自动开发”拆成一次可审阅交付
挑一个范围小、可以回退的需求，写清验收标准，再记录需求进入、计划、实现、测试、PR、合并六个节点。每一处保留当前状态、相关产物与未决问题；故意中断一次，观察恢复后是否重复执行或丢失约束。
最后让另一个人仅凭 PR 说明判断：做了什么、为何满足需求、测到了哪里、还有什么风险。若仍要打开十几个窗口补齐事实，就说明管理层还没有真正减少注意力负担。分别测配置时间、完成时间与审核时间，别把自动生成的数量当总收益。该练习为编辑整理，不是 Sheet0 的内部规程。
来源与回听
依据完整转写、官方介绍与章节，未复听音频、读产品代码或验证效率与经营数字。55:20以后为音乐／字幕识别内容，未作为访谈观点。具体新产品未给名称，本文不臆造。
- 01:29：OpenClaw 的能力释放。
- 05:29：垂直 Agent 与 SaaS 判断的转变。
- 19:52：文件系统与长程状态。
- 23:24：定时任务与主动发现。
- 31:28：内部任务流水线。
- 34:43：Harness 与协作控制。
- 51:48：承认旧目标与当前瓶颈不匹配。
相关主题：Agent 架构、持久状态、AI 组织。可对照Hidecloud 的早期 Agent 思考和Albert 的胜率框架，比较模型能力、资源条件与产品选择如何共同变化。
English edition
From Directing AI to Accepting Results: Wang Wenfeng Revises His Agent Thesis
Guest: Wang Wenfeng, founder of Sheet0, as introduced at release | Host: Qu Kai
Published: 2026-03-14 | Duration: 57:10 | Prepared: 2026-10-18
Official episode
When engineers supervise a dozen agent windows, is generating code still the next bottleneck? Wang proposes moving human attention from distributing tasks and maintaining configurations to final acceptance, with AI managing AI. His revision of an earlier thesis matters more than the phrase: instead of pursuing complete trustworthiness five to ten years ahead, he now addresses bottlenecks already visible over the next three to six months.
These are a founder’s judgments from March 2026, not current OpenClaw documentation. The idea that every agent will become a coding agent is not presented as an established outcome.
1. Separate the release of capability from its product packaging
Listen: 00:56–13:46. Wang treats Manus and OpenClaw as two releases of capability. The former demonstrated reasoning models in research tasks; the latter connected coding to more tasks, communication channels, and memory. Messaging integration is not the whole explanation. The coding agent’s ability and the surrounding configuration, coordination, and presentation work together.
His view of vertical agents changes accordingly. Previously committed to a long-term vertical approach, he now thinks a general coding agent with a terminal and reusable Skills can reproduce many specialized workflows. A terminal provides command-line operations; Skills record reusable methods and instructions. Software once replicated expertise through fixed interfaces. Goals and executable methods may provide another route.
This is a strong claim. Qu asks whether simple Skills can really hold complex industry knowledge. Producing a form, report, or record does not establish replacement of SaaS permissions, data governance, collaboration, responsibility, and maintenance. This article does not treat Wang’s assertion of comprehensive replacement as acceptance evidence. His cited 49.7% use-case figure has no externally checked sample or classification here and cannot establish behavior across all users.
2. Files make state repairable without guaranteeing reliability
Listen: 19:55–23:07. An earlier concern was that uncertainty compounds across long action chains. Wang now emphasizes persistent state: progress, memory, and task data are written to files that the model can inspect and revise, rather than proceeding only through an in-memory state machine.
This offers concrete benefits. After interruption, the system need not rely entirely on conversational recollection. Completed steps, mistakes, and necessary changes can be visible. Model capability and operating methods improve together; a smarter foundation model is not the sole explanation.
His statement that repair is no longer a problem remains a personal assessment. Files can be wrong, and the system may not recognize an error. An editorial requirement is versioning, outcome checks, and recovery paths. First test detection, then correct repair. Persistence creates a condition for recovery without proving long-task reliability.
3. Proactivity has levels; managing AI requires a task loop
Listen: 23:28–31:12. Sending a report at eight every morning differs from noticing an unrequested problem and proposing action. Qu explicitly distinguishes a preset trigger from exploration based on understanding a business. The latter includes proposing a solution and setting up a specialist agent to execute it.
Wang calls this AI managing AI: generating requirements, configuring, and scheduling alongside execution. Several models chatting do not automatically create it. The system needs connections to business information, documents, and tools, followed by adaptation. He acknowledges that installation does not yet mean immediate understanding of everything.
Experience transfers conditionally. Skills may preserve methods, but an agent optimized around one company’s repository and processes may be unhelpful elsewhere. Reusable methods and private context differ. An editorial boundary is to let discovery generate suggestions first, while messaging, data modification, and publication require corresponding authority. Proactivity is not authorization, and learning across agents does not imply permission to export internal material.
4. An internal pipeline reveals both opportunity and acceptance responsibility
Listen: 31:34–45:36. Wang describes connecting Linear requirements, GitHub collaboration, and Slack discussion to AI. A task triggers implementation, testing, and a pull request; a person checks and merges. Rather than humans distributing every request and directing the AI, the system absorbs more intermediate work.
He reports building the workflow in a week, increasing his personal efficiency tenfold over the previous year, and producing a new product in two weeks. These are internal experiences without comparable samples. Pull requests, commits, and code lines do not directly measure productivity. His earlier comparison of 1,600 commits with a small team’s annual work likewise does not establish equivalent functionality or quality.
He compares a harness to a saddle: a layer making powerful capabilities controllable and collaborative in a particular team, with both infrastructure and end-user properties. Backend feedback can improve through state tools and read-only database access. Observing state does not require arbitrary write access; the conversation explicitly discusses read-only permission.
The new product targets founders, broadly skilled product managers, and builders, seeking to reduce terminal switching. Moving people from direction to final approval still requires evidence that requirements were met, tests were meaningful, and regressions were avoided. An editorial caution is that a merge button is not a review process. Approval becomes a responsibility point only with sufficient evidence.
5. Large individual budgets come from work value, not token rankings
Listen: 45:57–51:45. Wang suggests individuals and small teams could spend at enterprise levels if AI performs work that previously required employees. One thousand customers paying US$100,000 annually yields US$100 million per year. This is hypothetical arithmetic, not achieved ARR or proof that all individuals can pay that amount.
The question is which spending is replaced and what value is created. Replacing part of an engineering budget needs visible savings or revenue. High consumption can also mean retries and inefficiency. A roughly US$20,000 monthly team bill and seven people matching an earlier thirty-to-fifty-person team are contemporary self-reports.
Wang advocates room for employees to explore, alongside strongly worded views on removing those who cannot adapt. This article does not make that a universal management method. An editorial approach evaluates budget with task quality, completed work, review effort, and user benefit. Spending alone is not ability. Exploration may defer short-term returns while retaining scope, review dates, and loss limits.
6. Shorter plans revise problem selection; they do not make accuracy irrelevant
Listen: 41:49–42:35; 48:24–55:15. Wang acknowledges that Sheet0’s earlier goal of complete explainability and accuracy exceeded what models could deliver. Users would say they wanted it, yet absolute accuracy was not everyone’s immediate barrier when other value was available. The observed bottleneck now is engineers’ attention across terminals, motivating one further step along existing demand.
This directly revises the previous reliability thesis. It should not become merely “better technology brings optimism,” or a claim that accuracy no longer matters. Different consequences require different standards; financial, medical, and other high-risk delivery cannot borrow general efficiency experience as its quality test. This limitation is editorial.
Wang asks the team whether, returning to March 2025, they would build Genspark. The question challenges attachment to differentiation. Forecasts encourage commitments; mistaken commitments make turning harder. Following clear signals can preserve flexibility, but a team still needs its own user value. A competitor’s financing is not PMF.
He also worries that employment losses and falling demand could reinforce each other, motivating shorter planning. This is an unverified macro concern, not an established mechanism or individual career advice. A shorter cycle should produce faster evidence and reversible adjustments without removing responsibility for results.
Exercise: make automatic development into one reviewable delivery
Choose a small, reversible requirement and define acceptance. Record six stages: intake, plan, implementation, testing, pull request, and merge. Preserve status, outputs, and unresolved questions at each. Deliberately interrupt once to see whether recovery repeats actions or loses constraints.
Ask another person to judge from the pull request alone what changed, why it meets the requirement, what was tested, and what risks remain. If a dozen windows are still needed to reconstruct events, the management layer has not reduced attention costs. Measure configuration, completion, and review time separately rather than treating generated volume as total benefit. This is an editorial exercise, not Sheet0’s internal procedure.
Sources and listening guide
Prepared from the complete transcript and official introduction and chapters, without replaying audio, inspecting product code, or validating efficiency and business figures. Music and subtitle recognition after 55:20 are excluded from interview claims. No exact name is invented for the new product.
- 01:29: Capability released through OpenClaw.
- 05:29: Changed views on vertical agents and SaaS.
- 19:52: Files and persistent task state.
- 23:24: Scheduled tasks versus proactive discovery.
- 31:28: The internal task pipeline.
- 34:43: Harnesses and collaborative control.
- 51:48: The earlier goal versus the present bottleneck.
Related themes: agent architecture, persistent state, and AI organizations. Compare Hidecloud’s early agent thinking and Albert’s success-probability framework to examine capability, resources, and product choices together.