王凯 Agent 工程实操

王凯从 2024-08 开始观察 Agent 浪潮（Manus 早期），但长期警告"Agent 元年说太早"（token 太贵、商业模型未跑通）。2026-Q1 开始亲自构建 Agent 系统（Chrome MCP / Loop / 多 Agent），认知达到"Agent 是新协作流程"层级。

1. Agent 演化时间线

graph LR
    A[2023-04 SEC FastFacts<br/>Workflow 雏形] --> B[2024-Q3 Devin 演示震撼<br/>Agent 分级概念]
    B --> C[2025-01 Coding 领域出独角兽<br/>Cursor / v0 / Manus]
    C --> D[2025-03 Manus 国内爆火<br/>但本质是 token 消耗变多]
    D --> E[2025-07 ChatGPT Agent<br/>OpenAI 专门优化模型]
    E --> F[2025-Q4 MCP 生态成熟<br/>Claude 接百+服务]
    F --> G[2026-Q1 Chrome MCP 普及<br/>多 Agent 自动化体系]
    G --> H[2026-04 王凯亲建 Agent 集群<br/>多浏览器多账号多 Agent]

2. Agent 分级（参考自动驾驶）

等级	描述	当前位置
L1: 单点辅助	人提任务,AI 写,人执行	普及
L2: 多步辅助	AI 帮人拆解任务、引导	Cursor / v0 / 深度研究
L3: 半自主	人监督下 AI 完成多步任务	Devin / Claude Code Agent
L4: 完全自主	无需人介入,Agent 完成复杂任务	待 token 价再降 90%
L5: Agent 组织	多 Agent 协作完成组织级任务	OpenAI L5 设想

王凯当前定位：L3 半自主（Claude Code 跑投资策略、自动推广等）。

3. MCP 实战清单

graph TD
    A[Claude Code] --> B[金融 MCP - FMP API<br/>解决数据准确性]
    A --> C[Chrome MCP - 控制真实浏览器]
    A --> D[飞书 CLI - 文档 + bot]
    A --> E[滴答清单 MCP - 任务管理]
    A --> F[Notion / Gmail 等]
    A --> G[AgentMail - Agent 邮箱]
    A --> H[自建 MCP - tokrepo skills]

3.1 FMP MCP（金融数据准确性）

• 问题：Claude/GPT 联网搜索拿数字 100% 不可信（编造数据）
• 解决：买 FMP API → 自建 MCP → Claude 通过 MCP 拿准确数据
• 效果：投资决策质量从"猜"变成"严谨可信"
• 关键纪律：所有涉及数据的 AI 输出都要绕开联网搜索

3.2 Chrome MCP（重大升级,2026-04）

• 官方版：使用无头浏览器（没有用户数据）
• 王凯改造："session 复用 + 用户数据浏览器"
• 解决：每次链接重复授权 + 抢焦点问题
• 效果：多个 Agent 并行控制不同浏览器,互不干扰
• 王凯把此 Skill 上传 tokrepo："Chrome MCP Background Proxy"

3.3 自建 MCP 的关键

• 任何需要"准确数据"的场景都应该走 MCP,不走 AI 联网搜索
• API 月费看似贵,但精准度带来的 token ROI 极高
• 把工作流中所有"数据源"接进来,让 Claude 成为"中枢"

4. 自动化任务体系（2026-Q1 重大转折）

graph TB
    A[设定自动化任务] --> B[Agent 启动]
    B --> C[10 min 心跳检查<br/>是否停止]
    C -- 停止 --> D[Claude Code 重新启动]
    C -- 运行 --> B
    B --> E[每日报告 → 飞书 bot]
    B --> F[异常处理 → 通知]

王凯顿悟（2026-04-12）：

• 之前苦于让 Claude Code 持续执行任务,它总自作聪明停止
• 想用 /loop 配置复杂定时任务
• 顿悟：设定 10min 检查定时,如果停就让其继续即可,不需要复杂循环

自动化任务示例：

• 监控 700 只股票 + 飞书 bot 发买卖信号
• 每日跑投资标的的内在价值
• YouTube/Reddit 评论自动化推广
• 多账号多浏览器并发分发内容

5. 多 Agent / 多浏览器 / 多账号体系（2026-04）

graph TB
    A[1 台电脑] --> B[5+ 浏览器]
    B --> B1[浏览器 1 - Google 账号 A]
    B --> B2[浏览器 2 - Google 账号 B]
    B --> B3[浏览器 3 - Google 账号 C]
    B --> B4[浏览器 4 - Google 账号 D]
    B --> B5[浏览器 5 - Google 账号 E]
    B1 --> C1[Claude Code 窗口 1<br/>YouTube 推广]
    B2 --> C2[Claude Code 窗口 2<br/>Reddit 推广]
    B3 --> C3[Claude Code 窗口 3<br/>Dev.to 推广]

核心思路：

• 不在一个 Agent 内做多账号切换
• 一个浏览器一个账号体系（Google 账号互不混淆）
• 每个 Claude Code 窗口绑定一个浏览器
• 同时铺几十个平台

前置要求：

• 海外手机号（淘宝买美国/英国卡，激活国外账号体系）
• 多套 Google / Apple / Claude 账号
• 海外银行卡 / Stripe / Wise / Lemon Squeezy 等支付

6. Agent 自主决策实验（200 美金给 Polymarket）

2026-03-12 王凯实验：

• 让 Claude Code 自主决策访问类 Polymarket 加密项目
• 给 200 美金本金,完全自主跑
• 设定：让 Claude Code 看完所有 RSI AI 交易飞书文档,凝练 Skill 进行 prediction market 决策

结论（截止 2026-05）：Agent 自主决策在投资场景还不到 L4

• 提示词框定后能稳定运行
• 但策略迭代仍需要人来突破
• "我看到什么高度,Agent 才能达到什么高度"

7. Agent 在投资中的角色（最值钱场景）

graph LR
    A[市场信息] --> B[Agent 1: 拉数据<br/>FMP MCP]
    B --> C[Agent 2: 量化风险<br/>5 个变量打分]
    C --> D[Agent 3: 计算内在价值<br/>6000 字提示词]
    D --> E[人类决策<br/>当前 L2]
    E --> F[Agent 4: 执行交易<br/>挂单 / 监控]

当前定位：决策仍然人来,但 Agent 把"信息采集 + 风险量化 + 价值计算 + 执行"全包了。

8. Devin / Manus / Cursor 对比

Agent	类型	王凯评价
Devin	编程 Agent	演示震撼,实战还在 L3
Manus	通用 Agent	国内首款相对复杂,但门槛不高
Cursor	辅助编程	主力,数据分析利器
Claude Code	命令行 Agent	2026 王凯主力,改变协作流程
v0	产品设计 Agent	"万字提示词敢 PUA"
Stitch	UI 设计 Agent	"切换 Gemini 后突飞猛进"

9. Agent 工程的"自主性 vs 可控性"陷阱

王凯踩坑（"六大模型跑币圈"分析）：

• 6 个大模型给币圈 1 万美金自主决策
• 第一次问询有倾向性 → 后续输入信息变化 → 决策永远基于当前持仓
• "AI 陷入到优化细节,跳不出来说'清仓保现金'"
• 结论：纯自主决策必须有"清仓"权限,否则永远陷在细节优化中

10. AI Agent 在不同场景的适用度

场景	适用度	原因
编程	⭐⭐⭐⭐⭐	结果可验证（跑通就对）
投资决策	⭐⭐⭐⭐	结果可验证（赚钱就对）
数学题	⭐⭐⭐⭐⭐	有标准答案
内容生成	⭐⭐⭐	主观评判,易飘
市场调研	⭐⭐⭐	信息源严重影响质量
法律 / 医疗	⭐⭐	准入 + 风险高

王凯引用 Jason Wei 论文："可验证性 = AI 优势区,不可验证 = AI 弱区"

11. Agent 商业模式判断

王凯立场（2025-03-29）：

• Agent 不是创业方向（小微 team）
• 因为:
  • token 单次成本太高（10 美金/任务）
  • 大模型公司自己做（OpenAI ChatGPT Agent 用专门优化模型）
  • 生态完善前竞争是赔钱战
• 但Agent 周边机会很多：
  • 使用模板库
  • Agent 导航
  • 精选案例
  • 垂直 Agent（包一层串起来满足细分场景）

12. Loop 自动化 vs 互联网定时任务

• 之前互联网时代有 cron / 定时任务
• 但都是预写规则
• AI Agent Loop 是有判断能力的循环
• "如果某个状态出现就这么做,否则那么做"
• 王凯把人脑能想到的所有"重复 + 判断"任务都丢给 Loop

13. 王凯说

"Agent 在 token 价格再降 90% 之前不会真正爆发。"

"Agent 改变的不是流程效率,是协作模式。"

"构建稳定运行的体系,而不是直接干。"

"1 个人能 cover 100 件事,如果你会构建 Agent 体系。"

"把自动化任务都丢给 AI,我去做更高维度的事。"

---

引用

• 王凯 AI Coding 实操：从 Cursor 到 Claude Code 终端工作流 - AI Coding 基础（用 Claude Code 构建 Agent）
• 王凯 RSI AI 投资策略与"好资产 好价格 好波动"框架 - 投资策略（Agent 最值钱场景）
• 王凯 AI 时代创业方法论与商业判断框架 - 创业方法论（Agent 不是创业方向）
• 王凯 出海推广 SOP 与海外多账号矩阵实战 - 海外推广（多 Agent 多浏览器铺平台）
• newtype-ai-agent-architecture - Agent 架构演进对比

2026-06 增补：产品经理理解长任务 Agent 的最小边界

本节来自王凯 6 月 9 日与 16 日两则文字帖，均已全文阅读。帖子推荐的外链文章没有随本批提供全文，因此本节不冒充对外链实现的拆解。

读工程机制，是为了知道什么时候不能依赖它

Harness（让模型持续工作的外围执行系统）、记忆和上下文工程（决定每一步给模型看什么信息），对产品经理的价值首先是识别能力边界：它大概能做什么、会丢什么信息、为什么会卡住。作者强调，AI 越深入投资等生产力场景、系统越像黑盒，越需要形成这种半专业判断；不必为了使用工具而掌握所有底层代码。

从执行一段任务，变成围绕目标反复行动

6 月 16 日把 Claude Code 的 Loop 与 Codex 的 Goal、自动化放在同一个需求下：让 Agent 多轮推进一个目标，直到达成。相对于本页已有的定时任务经验，新增观察是协作时间跨度和任务复杂度都在变长。评估时应看跨轮是否真的保持目标和进度，而不是只比较一次对话的回答质量。

这两帖只支持上述产品层认识，没有提供停止判定、预算上限或失败恢复的具体实现；配置操作应查对应项目资料，不能根据推荐链接标题推导出一套已验证方案。

来源与关联资料

• https://wx.zsxq.com/group/48884824582158/topic/22255284111544452
• https://wx.zsxq.com/group/48884824582158/topic/45544524484285560

2026-10-17 追补：用阅读反馈改进信息源发现

来自王凯10月12日的文字使用记录。原帖链接仅定位到圈子，附件未收；未取得其评分公式、代码或推荐系统实现。

固定订阅擅长持续跟踪已知作者，也容易反复看到同一圈人的材料。王凯描述的做法，是让信息源拓展 Agent 围绕既定主题持续寻找陌生来源，再用自己每天的阅读体验检验产出。他先觉得一篇博客有价值，随后才了解作者背景；这提醒读者把文章质量与作者名气分开判断。

可把这类任务拆成“发现候选—阅读具体文章—反馈哪里有用—调整下一轮查找”四步。以下是本库应用建议：留下候选链接与发现主题，写明文章提供了什么新事实或方法，以及为何值得继续读；同时保留不相关、重复和空泛的例子。发现新域名的数量只能说明搜索活动，不能证明阅读体验改善。

何时用：固定信息流开始重复、却又没有时间逐个搜索陌生作者时。先围绕自己确实关注的主题运行，按实际读到的正文评价，再考虑是否持续跟踪。一次惊喜不代表来源长期可靠；只奖励符合旧观点的文章，也可能把反馈循环变成自我印证。原帖没有提供已验证的评分权重，本节不补造自动排名公式。

Addition: Improve source discovery through reading feedback

This addition is based on Wang Kai's written account dated October 12. The available link identifies the community rather than the individual post, and attachments were not collected. His scoring formula, code and recommendation-system implementation were not available.

Fixed subscriptions help readers follow known authors but can repeatedly surface the same circle of material. Wang describes an agent that searches for unfamiliar sources around selected topics, with the resulting feed evaluated through his daily reading. He found a blog post useful before learning about its author. The lesson is to evaluate the article's contribution separately from the author's reputation.

The task can be organized as discovering candidates, reading specific articles, reporting what was useful and adjusting the next search. As an editorial application, keep the candidate link and discovery topic, explain the new fact or method it contributes and record why further reading is worthwhile. Retain examples that were irrelevant, repetitive or vague as well. Counting new domains measures search activity, not improved reading quality.

Use this approach when an established feed becomes repetitive and manually searching for unfamiliar writers takes too much time. Begin with topics you actually care about, evaluate the full text you read and then decide whether to follow the source. One pleasant surprise does not establish lasting reliability. Rewarding only articles that agree with existing views can turn feedback into self-confirmation. The post provides no validated scoring weights; no automatic ranking formula is invented here.

2026-10-18 跨作者方法补充：逐项审计应用连接授权

本节来自Peter Yang分享的Google应用与Agent连接审计方法，作为本文对工程执行的补充，不归为王凯本人观点。它把已有的权限边界落实到逐项检查、批准和回读，不将“已经连接”当作持续使用或允许撤销的证据。

先区分登录用途与数据访问

逐个查看连接的应用名称和授权详情，区分基础登录与读取Gmail、Drive、Calendar等数据的权限。不能只按应用名字判断风险，也不能因为近期未见使用就替用户断定已经弃用。对用途不明的连接询问用户是否仍在使用；记录未知项，不把未知写成无用或安全。

建议保留一份可复核清单：应用、用户确认的用途、具体访问范围、建议保留/断开/待确认、理由，以及断开可能影响的流程。权限较宽、名称陌生或功能重复的连接值得优先核实，但这些信号本身不足以直接撤销。

展示影响后，再按明确批准逐项撤销

先展示完整建议与可能中断的功能，取得用户对具体连接的明确批准，才执行对应撤销。不要将一次审阅授权扩大成批量清理所有连接，也不要把撤销连接扩展为删除账户或取消订阅。

操作后重新查看授权状态，确认目标连接或相应访问权限是否已经移除，并记录结果。按钮点击、页面提示与最终授权状态分开保存；如果结果不明，保留待核实状态，不宣称已撤销，也不继续扩大操作范围。这里的核实限于当前可见的授权结果，不据此声称第三方已删除历史数据。

适用于个人账号连接盘点、Agent接入多个办公工具或停用旧集成前的检查。其交付物应是用户能理解并逐项批准的影响清单，以及变更后的核实记录；不是一份未经执行的风险评分。

English translation

Addition from another author: Audit connected applications one authorization at a time

This section draws on Peter Yang’s method for auditing Google application and agent connections. It supplements the engineering practices in this article and is not attributed to Wang Kai. It turns permission boundaries into individual checks, approvals and readback. An existing connection is not evidence of continued use or permission to revoke it.

Separate sign-in from data access

Inspect each connected application’s name and authorization details. Distinguish basic sign-in from permissions to read data in Gmail, Drive, Calendar and other services. Do not assess risk solely from an application’s name or assume that a connection is abandoned because recent use has not been observed. Ask the user whether connections with unclear purposes are still in use. Record unknowns without labeling them useless or safe.

Keep a reviewable list of the application, its user-confirmed purpose, specific access scope, a recommendation to keep, disconnect or clarify, the reason and any workflow that disconnection may disrupt. Broad permissions, unfamiliar names and overlapping functions warrant closer inspection, but these signals alone do not justify revocation.

Show the impact, then revoke only the specifically approved connections

Present the complete recommendations and potentially disrupted functions first. Revoke a connection only after the user explicitly approves that specific change. Permission to review connections is not permission to remove them all, and disconnecting an integration does not authorize deleting an account or canceling a subscription.

After the action, inspect the authorization state again to establish whether the target connection or corresponding access permission was removed, and record the result. Keep a button click, an interface notification and the final authorization state as separate evidence. If the result is unclear, retain a pending-verification status, do not claim revocation succeeded and do not broaden the action. This verification covers the currently observable authorization result; it does not establish that a third party deleted previously collected data.

Use this method when reviewing personal account connections, connecting an agent to several office tools or preparing to retire an old integration. The deliverable is an impact list that the user can understand and approve item by item, plus verification after each change. An unexecuted risk score is insufficient.

2026-10-22 增补：信息发现之后，再单独追溯第一手出处

发现了相关报道，不代表已经拿到能支持结论的原始证据。王凯在 DisconfirmAI 的个人实践中，区分了监控媒体和主动寻找原始信息：发现某产品发布后，再找公司的原始公告；找不到时，才按渠道权威程度选择可获得的材料。它比单纯扩充订阅列表多了一步来源追溯，但原帖没有公开评分公式或验证准确率。

可以在已有“发现—阅读—反馈”循环中增加一张出处记录：最初在哪发现、最后引用哪份原文、原文实际支持什么、哪些部分仍是转述或推断。第一手出处通常更接近发布方的原意，但官方陈述仍可能是商业宣传；一手不等于独立验证，更不意味着所有数字和效果已经成立。

找不到原文时，保留未找到、仅摘要或仅转述的状态，把替代来源和选择理由一起留下，避免在后续知识稿中丢失证据层级。这是本站建议的交付物，不是对 DisconfirmAI 内部流程实现的确认。原帖提到的信息覆盖广度和个人“70分以上”体验仅作为个人反馈，不解释成客观覆盖率或准确率。

来源：王凯关于 DisconfirmAI 信息溯源的实践记录（2026-10-22）。

October 22, 2026: Trace Original Sources After Discovery

Finding a relevant report does not mean obtaining the original evidence needed to support its conclusions. In his personal DisconfirmAI practice, Wang Kai distinguishes monitoring media from actively locating original material: after discovering a product announcement, look for the company’s announcement; if it cannot be found, choose accessible material according to source authority. This adds provenance work beyond expanding subscriptions, but the post provides no scoring formula or verified accuracy rate.

An existing discovery, reading and feedback loop can add a provenance record: where the lead was found, which original is ultimately cited, what that original actually supports and what remains a retelling or inference. A firsthand source is usually closer to the publisher’s intended statement, but an official statement may still be marketing. Firsthand does not mean independently verified or establish every claimed number and outcome.

When the original cannot be found, retain an explicit missing-original, abstract-only or secondary-account status, together with the substitute source and selection rationale. This keeps evidence levels visible in later knowledge articles. The record is this site’s proposed deliverable, not confirmation of DisconfirmAI’s internal implementation. The post’s coverage descriptions and personal score above 70 are user feedback, not objective coverage or accuracy measures.

Source: Wang Kai’s DisconfirmAI provenance practice note, October 22, 2026.

English edition

Wang Kai: Practical Agent Engineering

Historical source note: This is a full translation of the published synthesis, including its dated observations, examples and judgments. Statements about product capabilities, costs, investing and forecasts are attributed historical claims, not newly verified facts or investment advice. Later additions retain their own source and evidence limits.

Wang Kai began observing the agent wave in August 2024, including early Manus developments, but repeatedly warned that calling it “the first year of agents” was premature: tokens were expensive and business models remained unproven. In Q1 2026 he began building agent systems himself, using Chrome MCP, loops and multiple agents. His framing evolved toward agents as a new way of collaborating.

1. A Timeline of Agent Development

graph LR
 A[April 2023: SEC FastFacts<br/>Early workflow] --> B[Q3 2024: Devin demonstration<br/>Agent autonomy levels]
 B --> C[January 2025: Coding unicorns<br/>Cursor / v0 / Manus]
 C --> D[March 2025: Manus gains attention in China<br/>Higher token consumption underneath]
 D --> E[July 2025: ChatGPT Agent<br/>OpenAI optimizes a dedicated model]
 E --> F[Q4 2025: MCP ecosystem matures<br/>Claude connects to 100+ services]
 F --> G[Q1 2026: Chrome MCP adoption<br/>Multi-agent automation]
 G --> H[April 2026: Wang builds an agent cluster<br/>Multiple browsers, accounts and agents]

2. Agent Levels, by Analogy with Autonomous Driving

Level	Description	Position in the historical account
L1: Single-task assistance	A person sets the task, AI writes, and the person executes	Widespread
L2: Multi-step assistance	AI helps break down tasks and guides a person	Cursor, v0 and deep research
L3: Partial autonomy	AI completes multi-step tasks under human supervision	Devin and Claude Code Agent
L4: Full autonomy	An agent completes complex tasks without human intervention	Wang anticipated a further 90% reduction in token prices
L5: Agent organizations	Multiple agents collaborate on organization-level tasks	OpenAI's L5 vision

Wang placed his own usage at L3, including Claude Code running investment strategies and automated promotion.

3. A Practical MCP Checklist

graph TD
 A[Claude Code] --> B[Financial MCP: FMP API<br/>Data accuracy]
 A --> C[Chrome MCP: control a real browser]
 A --> D[Feishu CLI: documents and bot]
 A --> E[TickTick MCP: task management]
 A --> F[Notion / Gmail and other services]
 A --> G[AgentMail: agent email]
 A --> H[Custom MCP: tokrepo skills]

3.1 FMP MCP: Financial Data Accuracy

• Problem in Wang's account: numbers obtained through Claude/GPT web searches were, in his emphatic wording, “100% unreliable” because of fabrication. This is his characterization, not a measured universal error rate.
• His solution: pay for the FMP API, build an MCP server, and have Claude retrieve data through it.
• Reported effect: investment analysis moved from guessing toward more rigorous, trustworthy work.
• His operating rule: route data-dependent AI output away from generic web search.

3.2 Chrome MCP: An Important April 2026 Upgrade

• The version he described as official used a headless browser without user data.
• His modification combined session reuse with a browser containing the user's existing profile.
• It addressed repeated authorization and browser focus contention.
• He reported that multiple agents could control separate browsers concurrently without interfering.
• He shared the skill on tokrepo: Chrome MCP Background Proxy.

3.3 What Matters in a Custom MCP Server

Wang recommended MCP rather than general AI web search whenever accurate data is required. Although API subscriptions appear expensive, he regarded their accuracy as delivering strong returns on token spending. Connecting the workflow's data sources lets Claude act as a central coordinator. These are his engineering preferences; an API connection still requires checking data coverage and correctness.

4. An Automation System: A Q1 2026 Turning Point

graph TB
 A[Define an automated task] --> B[Start the agent]
 B --> C[Check every 10 minutes<br/>Has execution stopped?]
 C -- Stopped --> D[Restart Claude Code]
 C -- Running --> B
 B --> E[Daily report to Feishu bot]
 B --> F[Handle exceptions and notify]

Wang's April 12, 2026 realization: Claude Code kept deciding to stop when he wanted it to continue. He had tried to configure complicated scheduled tasks with /loop, then realized that a check every ten minutes could simply tell it to continue if it had stopped, avoiding a complex loop.

His examples included monitoring 700 stocks and sending trading signals through a Feishu bot, calculating investment targets' intrinsic value daily, automated promotion in YouTube/Reddit comments, and concurrent content distribution across multiple accounts and browsers. These are reported uses, not instructions or evidence of effectiveness.

5. Multiple Agents, Browsers and Accounts: April 2026

graph TB
 A[One computer] --> B[5+ browsers]
 B --> B1[Browser 1: Google account A]
 B --> B2[Browser 2: Google account B]
 B --> B3[Browser 3: Google account C]
 B --> B4[Browser 4: Google account D]
 B --> B5[Browser 5: Google account E]
 B1 --> C1[Claude Code window 1<br/>YouTube promotion]
 B2 --> C2[Claude Code window 2<br/>Reddit promotion]
 B3 --> C3[Claude Code window 3<br/>Dev.to promotion]

The core idea was to avoid switching accounts inside a single agent. Each browser had one account environment, keeping Google accounts separate, and each Claude Code window was bound to one browser. Wang described distributing work across dozens of platforms at once.

The account described prerequisites including overseas phone numbers—US/UK SIM cards purchased through Taobao to activate overseas account environments—multiple Google/Apple/Claude accounts, and overseas banking or payment services such as Stripe, Wise and Lemon Squeezy. This preserves the historical account rather than recommending procurement or circumvention practices.

6. An Autonomous Decision Experiment: $200 for a Polymarket-Type Project

On March 12, 2026, Wang described letting Claude Code independently explore a Polymarket-like crypto project, giving it $200 of capital and asking it to read all the RSI AI trading documents in Feishu, distill a skill, and make prediction-market decisions.

His conclusion as of May 2026 was that autonomous investment decisions had not reached L4. The system could run steadily within a prompt-defined scope, but humans still had to unlock improvements in strategy. In his words, the agent's ceiling depended on the height of his own understanding.

7. Agents in Investing: His Highest-Value Use Case

graph LR
 A[Market information] --> B[Agent 1: retrieve data<br/>FMP MCP]
 B --> C[Agent 2: quantify risk<br/>Score five variables]
 C --> D[Agent 3: intrinsic value<br/>A 6,000-character prompt]
 D --> E[Human decision<br/>Currently L2]
 E --> F[Agent 4: execute orders and monitor]

In this account, humans still make the decision, while agents handle information gathering, risk quantification, valuation and execution support. The separate L2 label here describes the investment decision workflow; it does not erase his broader L3 usage classification.

8. Devin, Manus, Cursor and Other Tools

Tool	Type	Wang's historical assessment
Devin	Coding agent	Impressive demonstrations; practical use still L3
Manus	General-purpose agent	A relatively complex early Chinese product, but not a high barrier to entry
Cursor	Coding assistance	A main tool and useful for data analysis
Claude Code	Command-line agent	His main tool in 2026; changed collaboration workflows
v0	Product-design agent	His colloquial comment was that it could be pushed with a 10,000-character prompt
Stitch	UI-design agent	In his experience, improved rapidly after switching to Gemini

9. The Autonomy-versus-Control Trap

In Wang's analysis of an experiment giving six models $10,000 for autonomous crypto decisions, the initial question introduced a bias. As later information changed, decisions remained anchored in current holdings. He observed that AI optimized details without stepping back to say “sell everything and keep cash.” His conclusion was that truly autonomous decisions need permission to exit positions, or the system may stay trapped in local optimization. This is a reported critique, not a validated trading rule.

10. Suitability Across Scenarios

Scenario	Rating in the source	Rationale in the source
Coding	5/5	Results can be checked: the program runs
Investment decisions	4/5	He treated profit as an observable outcome
Mathematics	5/5	Standard answers exist
Content generation	3/5	Judgment is subjective and can drift
Market research	3/5	Source quality strongly affects results
Law and medicine	2/5	Entry requirements and high risk

Wang cited work by Jason Wei to express the idea that verifiable tasks are an AI strength, and hard-to-verify tasks a weakness. The table preserves his heuristic: running code does not prove every requirement, and investment profit alone does not establish decision quality or causality.

11. Agent Business Models

On March 29, 2025, Wang argued that general agents were not an attractive startup direction for very small teams. His reasons were token costs of approximately $10 per task, competition from model companies themselves—such as OpenAI using an optimized model for ChatGPT Agent—and loss-making competition before the ecosystem matured.

He nevertheless saw opportunities around agents: template libraries, directories, curated examples, and vertical agents combining existing capabilities to serve a specific need.

12. Agent Loops Versus Traditional Scheduled Jobs

The internet era already had cron and scheduled tasks, but these followed prewritten rules. Wang saw an AI agent loop as repetition with judgment: if one state appears, take one action; otherwise take another. He assigned it the repetitive tasks involving judgment that he could identify.

13. Statements Attributed to Wang Kai

“Agents will not truly take off until token prices fall another 90%.”

“Agents change the mode of collaboration, not just process efficiency.”

“Build a system that runs reliably rather than jumping straight into execution.”

“One person can cover 100 things if they know how to build an agent system.”

“Give automated tasks to AI so I can work on higher-level things.”

These are translations of statements quoted in the published Chinese synthesis.

References and Related Notes

• Wang Kai: AI Coding — the foundations for building agents with Claude Code.
• Wang Kai: Investment Strategy — the investing use case.
• Wang Kai: Startup Methodology — his view of agents as a startup direction.
• Wang Kai: Overseas Marketing — multiple agents and browsers across platforms.
• Newtype: AI Agent Architecture — comparison of architecture evolution.

June 2026 Addition: Minimum Boundaries Product Managers Need to Understand for Long-Running Agents

This section draws on Wang Kai's text posts dated June 9 and 16, both read in full. The linked articles recommended by those posts were not provided in full with that batch, so this is not presented as an implementation analysis of those external articles.

Read Engineering Mechanisms to Know When They Cannot Be Relied On

For product managers, a harness—the execution system around a model that keeps it working—memory, and context engineering—deciding what information the model sees at each step—first help identify capability boundaries: what the system may do, what information it may lose, and why it may stall. Wang argues that as AI enters productivity settings such as investing, and systems become more opaque, this partly technical judgment becomes more important. Using a tool does not require mastering all its underlying code.

From Executing One Task Segment to Repeatedly Acting Toward a Goal

The June 16 post places Claude Code's Loop, Codex's Goal and automation under one need: let an agent advance a goal over multiple rounds until it is achieved. Beyond the scheduled-task experience already described here, the new observation concerns longer collaboration and more complex tasks. Evaluation should ask whether the goal and progress persist across rounds, rather than compare only answer quality in a single conversation.

These two posts support only those product-level observations. They do not provide concrete implementations for stopping criteria, spending limits or failure recovery. Consult the relevant project documentation for configuration; a recommended link's title does not establish a tested implementation.

Addition: Improve source discovery through reading feedback

This addition is based on Wang Kai's written account dated October 12. The available link identifies the community rather than the individual post, and attachments were not collected. His scoring formula, code and recommendation-system implementation were not available.

Fixed subscriptions help readers follow known authors but can repeatedly surface the same circle of material. Wang describes an agent that searches for unfamiliar sources around selected topics, with the resulting feed evaluated through his daily reading. He found a blog post useful before learning about its author. The lesson is to evaluate the article's contribution separately from the author's reputation.

The task can be organized as discovering candidates, reading specific articles, reporting what was useful and adjusting the next search. As an editorial application, keep the candidate link and discovery topic, explain the new fact or method it contributes and record why further reading is worthwhile. Retain examples that were irrelevant, repetitive or vague as well. Counting new domains measures search activity, not improved reading quality.

Use this approach when an established feed becomes repetitive and manually searching for unfamiliar writers takes too much time. Begin with topics you actually care about, evaluate the full text you read and then decide whether to follow the source. One pleasant surprise does not establish lasting reliability. Rewarding only articles that agree with existing views can turn feedback into self-confirmation. The post provides no validated scoring weights; no automatic ranking formula is invented here.

Addition from another author: Audit connected applications one authorization at a time

This section draws on Peter Yang’s method for auditing Google application and agent connections. It supplements the engineering practices in this article and is not attributed to Wang Kai. It turns permission boundaries into individual checks, approvals and readback. An existing connection is not evidence of continued use or permission to revoke it.

Separate sign-in from data access

Inspect each connected application’s name and authorization details. Distinguish basic sign-in from permissions to read data in Gmail, Drive, Calendar and other services. Do not assess risk solely from an application’s name or assume that a connection is abandoned because recent use has not been observed. Ask the user whether connections with unclear purposes are still in use. Record unknowns without labeling them useless or safe.

Keep a reviewable list of the application, its user-confirmed purpose, specific access scope, a recommendation to keep, disconnect or clarify, the reason and any workflow that disconnection may disrupt. Broad permissions, unfamiliar names and overlapping functions warrant closer inspection, but these signals alone do not justify revocation.

Show the impact, then revoke only the specifically approved connections

Present the complete recommendations and potentially disrupted functions first. Revoke a connection only after the user explicitly approves that specific change. Permission to review connections is not permission to remove them all, and disconnecting an integration does not authorize deleting an account or canceling a subscription.

After the action, inspect the authorization state again to establish whether the target connection or corresponding access permission was removed, and record the result. Keep a button click, an interface notification and the final authorization state as separate evidence. If the result is unclear, retain a pending-verification status, do not claim revocation succeeded and do not broaden the action. This verification covers the currently observable authorization result; it does not establish that a third party deleted previously collected data.

Use this method when reviewing personal account connections, connecting an agent to several office tools or preparing to retire an old integration. The deliverable is an impact list that the user can understand and approve item by item, plus verification after each change. An unexecuted risk score is insufficient.

October 22, 2026: Trace Original Sources After Discovery

Finding a relevant report does not mean obtaining the original evidence needed to support its conclusions. In his personal DisconfirmAI practice, Wang Kai distinguishes monitoring media from actively locating original material: after discovering a product announcement, look for the company’s announcement; if it cannot be found, choose accessible material according to source authority. This adds provenance work beyond expanding subscriptions, but the post provides no scoring formula or verified accuracy rate.

An existing discovery, reading and feedback loop can add a provenance record: where the lead was found, which original is ultimately cited, what that original actually supports and what remains a retelling or inference. A firsthand source is usually closer to the publisher’s intended statement, but an official statement may still be marketing. Firsthand does not mean independently verified or establish every claimed number and outcome.

When the original cannot be found, retain an explicit missing-original, abstract-only or secondary-account status, together with the substitute source and selection rationale. This keeps evidence levels visible in later knowledge articles. The record is this site’s proposed deliverable, not confirmation of DisconfirmAI’s internal implementation. The post’s coverage descriptions and personal score above 70 are user feedback, not objective coverage or accuracy measures.

Source: Wang Kai’s DisconfirmAI provenance practice note, October 22, 2026.

Source Links and Related Public Notes

• Reference 1
• Reference 2
• Reference 3
• Reference 4
• Reference 5
• Reference 6
• Reference 7
• Reference 8
• Reference 9