# How I'd make $10 million with AI agents
**作者**: GREG ISENBERG
**日期**: 2026-07-13T18:22:05.000Z
**来源**: [https://x.com/gregisenberg/status/2076733920834371585](https://x.com/gregisenberg/status/2076733920834371585)
---
2 markets worth trillions are opening at the same time. I've been through 3 platform shifts living in Silicon Valley building/selling VC-backed companies and I've NEVER seen this.
1. Mobile apps came back. They died in 2015 when an MVP cost . Now it costs twenty bucks, and more importantly, every app can be AI-first. That means every category in the store gets rebuilt.
2. Agent-native apps didn't exist before. Claude Code, Codex, Hermes. This is the next operating system and it's basically empty. Build apps for this new operating system.
Building for either one right now feels like building for iPhone in 2009. Fishing with dynamite.
Here's the playbook. You can do all of it without raising a dollar. You can raise money, sure, but for most of these ideas it's optional.
## 1. Find things people pay for and hate using
Check Sensor Tower. Apps doing $50-200K MRR in boring categories. Utilities. Productivity. Niche professional tools. Steady revenue, dated interface, no update in six months.
The gift is a sub-4 star rating with strong retention. That means people need it badly enough to keep paying for something they resent. Look at revenue per download, not total revenue.
Someone is making $500K a month selling a louder alarm clock. That's the whole product. It's louder.
You can use Ideabrowser (for free!) to get validated startup ideas. I built this product for myself (to find ideas to build) and for you (to get creative juices flowing).
## 2. Find the AI shaped gap inside it
Every app on your phone assumes a human is tapping it. You open it, you do the work, you close it. It waits for you like a very patient calculator.
An AI-first app thinks. It decides. It works while you sleep.
The tells that an app is ripe:
- Upload and wait is the core experience
- Repetitive forms with tiny variations
- Manual tagging, sorting, categorizing
- Premium pricing justified by time saved instead of outcomes delivered
Your fitness app should move your training session because it saw your sleep and knew you were cooked. Your finance app should call your insurance company and negotiate the rate down.
## 3. Design the agent
This is where most people fumble, so get specific.
Before you build anything, write the spec. Seven questions:
1. What wakes the agent up? (A cron job. An inbound email. A webhook. A text.)
2. What context does it need? (Their data, their SOPs, their pricing, their history.)
3. What tools can it call?
4. What can it do on its own?
5. What needs approval before it goes out?
6. When does it escalate to a human?
7. How do you know it worked?
Then build the smallest useful version. I call it the minimum viable agent.
Draft and approve. The agent drafts, a human hits send. Everyone tries to ship the fully autonomous employee first, and it demos great on X and breaks in production. Your customer might be buying their first agent ever. Earn the autonomy.
Then build your eval set before you scale it. Take 50 real examples of the job. Run the agent against them. It got 42 right, flagged 6, blew 2. Now you have a gym, and every time you change the prompt or the model you run it back through.
That eval set doubles as your best sales asset. Telling a property manager "we ran this on 50 of your old maintenance requests, here's exactly what happened" closes deals faster.
## 4. Let the agents build the app
The build itself is agents now.
Claude Code or Codex writes the app. Point it at a design you like and have it match. Fable for the hard architectural thinking, a cheaper model like GLM or Grok 4.5 for the execution loops. That model chaining is roughly 5x cheaper for near-frontier output, and when you're running all day the difference compounds fast.
Give the agent skills. A skill is a markdown file. Write your ASO rules, your brand voice, your onboarding copy patterns into skills and every future agent inherits them.
Run subagents. One writes the code. One writes the App Store listing. One generates thirty UGC scripts. One watches your reviews and files bugs.
## 5. Consider making it an iMessage product
For a lot of products the best interface is iMessage. You text a number and the thing happens. Your mom can use it. Your barber can use it.
A nutrition coach you text a photo of your lunch to. A bookkeeper living in a contractor's group chat with his crew. Something that texts a small business's customers back within four seconds, which fixes the biggest reason they lose deals.
Poke is doing interesting things here. Lindy too (I invested).
Once it's in a message thread it's in the group chat. It spreads like a meme instead of like an app.
## 6. Or build for the new OS, which is emptier (like wide empty)
Think about what an operating system does. Runs programs, holds memory, plugs into the world, has config files, has apps.
Now look at Claude Code. The model runs the programs. The context window is memory. Tools touch the world. Markdown files are the config. Skills are the apps.
On this OS, an app is a markdown file. Plain English, dropped in a folder, and every agent that reads it can do something new. No deploy, no app review, no code.
A few thousand MCP servers exist today. There are 30 million businesses in America. Almost none of them are connected.
Four openings I'd take:
- Voice. Every business has a phone nobody answers after 5pm. An agent that picks up and books the job is worth thousands a month to a plumber.
- Agent infrastructure. Payments they can use, memory they can trust, identity they can prove. Eleven people are working on this.
- Industry skill libraries. A markdown file that teaches an agent bookkeeping for dental practices. The customer's alternative is hiring a human.
- Making businesses legible. Most companies exist only in the founder's head. An agent can't run any of it yet.
Basically, if this new operating system is being creating before our eyes.
What are the agent native apps you can be building?
## 7. Build your media company
Find the niche where a tool gets used but never shared. Photographers. Roofers. Accountants. High usage, zero presence on their platforms.
Build an army of UGC creators. Around $5 CPM for B2C, higher for B2B. Go straight at the people already complaining in public.
Point agents at this too. One agent scrapes the creators in your niche and drafts the DMs. One turns every customer support ticket into a piece of content. One reads your reviews every morning and tells you the one thing to fix.
You basically need to build a media company from scratch. The agents make that possible with one person.
## 8. Then do it again
If you don't raise money, build a portfolio. More diversification, less risk. Cashflow from one funds the next. Share the infrastructure. Every launch gets cheaper because your skills, your evals, and your distribution agents all carry over.
## The math
You need $2-3M a year and a 3-5x exit.
- One product at $200K/month = $2.4M ARR
- Sell at 4x = $9.6M
- You raised nothing, so you keep it
Not crazy. Not out of reach if you find the right niche, right channels, right product. You gotta admit its possible.
$10M is one app at $200K/month. Or three apps at $70K each. Or a portfolio where two of eight worked.
Twenty bucks to build. Two weeks to ship. Ten shots on goal a year. You need one to hit.
Now, like all startups, most will fail. That's always been true and AI doesn't change it. What AI changes is the cost of finding out. You used to bet $500K and two years to learn your idea was bad. Now you bet twenty bucks and two weeks.
But I believe this is the greatest time to be building in history. There are fewer gatekeepers than ever. No VC to convince. No engineer to beg. No designer to wait on. No app store gods to appease, because the tools are free and the knowledge is on YouTube/X and the whole thing fits on the laptop you already own.
And worst case you learn a ton about AI agents, which is the most valuable skill on the planet right now anyway. You "lose" a couple weekends and you come out the other side as the person everyone at your company asks for help. That's a pretty good floor.
Someone might read this and make . Someone else makes . Someone makes an extra $2K a month and finally stops worrying about rent. Someone builds a dumb little app for their niche and it changes their whole year.
All of those are wins. I'm rooting for every one of them.
I hope this got your creative juices flowing. Send this to a friend.
Now, go build.
❤️
## 相关链接
- [GREG ISENBERG](https://x.com/gregisenberg)
- [@gregisenberg](https://x.com/gregisenberg)
- [92K](https://x.com/gregisenberg/status/2076733920834371585/analytics)
- [$50-200K](https://x.com/search?q=%2450-200K&src=cashtag_click)
- [Ideabrowser](https://ideabrowser.com/)
- [Poke](https://poke.com/)
- [Lindy](http://lindy.ai/)
- [$2-3M](https://x.com/search?q=%242-3M&src=cashtag_click)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [2:22 AM · Jul 14, 2026](https://x.com/gregisenberg/status/2076733920834371585)
- [92.2K Views](https://x.com/gregisenberg/status/2076733920834371585/analytics)
- [View quotes](https://x.com/gregisenberg/status/2076733920834371585/quotes)
---
*导出时间: 2026/7/14 09:53:52*
---
## 中文翻译
# 我如何利用 AI 代理赚到 1000 万美元
**作者**: GREG ISENBERG
**日期**: 2026-07-13T18:22:05.000Z
**来源**: [https://x.com/gregisenberg/status/2076733920834371585](https://x.com/gregisenberg/status/2076733920834371585)
---
两个价值数万亿美元的市场正在同时开启。我在硅谷生活，经历建立/出售风投支持的公司，经历过 3 次平台转移，但我从未见过这样的景象。
1. 移动应用回归了。它们在 2015 年“死”了，那时开发一个 MVP（最小可行性产品）的成本高达 10 万美元。现在只需二十美元，更重要的是，每个应用都可以实现 AI 优先（AI-first）。这意味着应用商店里的每个品类都将被重构。
2. 原生代理应用以前是不存在的。Claude Code、Codex、Hermes。这就是下一个操作系统，而且它基本上是空的。为这个新操作系统开发应用吧。
现在针对其中任何一个进行开发，感觉就像 2009 年年为 iPhone 开发一样。如同用炸药捕鱼，收获颇丰。
以下是行动指南。你可以在不融资一分钱的情况下完成所有这些。当然，你可以融资，但对于大多数想法来说，这完全是可选的。
## 1. 寻找那些人们愿意付费但又讨厌使用的东西
查看 Sensor Tower。那些在无聊类别中月收入在 5 万到 20 万美元的应用。工具类、效率类、小众专业工具。稳定的收入，过时的界面，六个月没更新。
如果你发现一个评分低于 4 星但留存率很强的应用，那就是上帝送来的礼物。这意味着人们非常需要它，以至于愿意继续为他们讨厌的东西付费。要看每次下载的收入，而不是总收入。
有人靠卖一个“声音更大的闹钟”每月赚 50 万美元。这就是整个产品。它只是声音更大。
你可以使用 Ideabrowser（免费！）来获得经过验证的创业想法。我构建这个产品是为了我自己（寻找要构建的想法）也是为了你（激发创意的灵感）。
## 2. 在其中寻找 AI 形状的缺口
你手机上的每个应用都假设是人在点击它。你打开它，你做工作，你关闭它。它像一个非常耐心的计算器一样等待着你。
一个 AI 优先的应用会思考。它会决定。它在你睡觉时工作。
如果一个应用表现出以下迹象，它就成熟了：
- “上传并等待”是核心体验
- 只有微小差异的重复性表单
- 手动打标签、排序、分类
- 用节省的时间而不是交付的结果来证明高价合理的溢价
你的健身应用应该根据你的睡眠数据调整你的训练时间，因为它知道你精疲力尽。你的理财应用应该打电话给你的保险公司并协商降低费率。
## 3. 设计代理
这是大多数人搞砸的地方，所以要具体化。
在构建任何东西之前，先写好规格说明。七个问题：
1. 什么会唤醒代理？（一个定时任务。一封收到的邮件。一个 Webhook。一条短信。）
2. 它需要什么上下文？（他们的数据、他们的 SOP（标准作业程序）、他们的定价、他们的历史。）
3. 它可以调用什么工具？
4. 它可以独立做什么？
5. 什么在发出之前需要批准？
6. 什么时候它升级给人类？
7. 你怎么知道它起作用了？
然后构建最小有用的版本。我称之为最小可行代理。
起草和批准。代理起草，人类点击发送。每个人都试图首先运送完全自主的员工，它在 X（Twitter）上演示效果很好，但在生产环境中会出问题。你的客户可能是第一次购买代理。你要赢得这种自主权。
然后在扩展之前构建你的评估集。拿 50 个工作的真实例子。让代理对它们运行。它做对了 42 个，标记了 6 个，搞砸了 2 个。现在你有一个健身房（训练场），每次你更改提示词或模型时，都要重新运行一遍。
那个评估集也是你最好的销售资产。告诉一位物业管理经理“我们在你的 50 个旧维护请求上运行了这个，这是确切发生的事情”，这能更快地达成交易。
## 4. 让代理来构建应用
现在的构建过程本身也是由代理完成的。
Claude Code 或 Codex 来编写应用。指向一个你喜欢的设计并让它匹配。Fable 用于复杂的架构思考，像 GLM 或 Grok 4.5 这样更便宜的模型用于执行循环。这种模型链式组合对于接近前沿水平的输出来说大约便宜 5 倍，当你全天运行时，差异会迅速复合。
赋予代理技能。一个技能就是一个 Markdown 文件。将你的 ASO（应用商店优化）规则、你的品牌声音、你的入门文案模式写入技能，每个未来的代理都会继承它们。
运行子代理。一个写代码。一个写应用商店列表。一个生成 30 个 UGC（用户生成内容）脚本。一个监听你的评论并提交 Bug。
## 5. 考虑将其做成 iMessage 产品
对于许多产品来说，最好的界面就是 iMessage。你发短信给一个号码，事情就办成了。你妈妈能用。你的理发师能用。
一个你把午餐照片发短信给它的营养教练。一个生活在他与团队群聊里的承包商簿记员。一个能在 4 秒内回复小企业客户短信的东西，这解决了他们丢失订单的最大原因。
Poke 在这方面做了一些有趣的事情。Lindy 也是（我投资了）。
一旦它进入了消息线程，它就进入了群聊。它像模因一样传播，而不像应用。
## 6. 或者构建面向新操作系统的应用，那里更空旷（像广袤的荒原）
想想操作系统是做什么的。运行程序、持有内存、接入世界、有配置文件、有应用。
现在看看 Claude Code。模型运行程序。上下文窗口就是内存。工具接触世界。Markdown 文件就是配置。技能就是应用。
在这个操作系统上，一个应用就是一个 Markdown 文件。纯英文，放在一个文件夹里，每个读取它的代理都能做新的事情。没有部署，没有应用审查，没有代码。
今天存在几千个 MCP 服务器。美国有 3000 万家企业。几乎都没有连接。
我会抓住的四个机会：
- 语音。每个企业都有一个下午 5 点后没人接听的电话。一个能接听并预约工作的代理对水管工来说每月值几千美元。
- 代理基础设施。它们可以使用的支付、它们可以信任的内存、它们可以证明的身份。只有 11 个人在做这个。
- 行业技能库。一个教导代理牙科诊所簿记的 Markdown 文件。客户的替代方案是雇佣人类。
- 让企业变得可读。大多数公司只存在于创始人的脑子里。代理还不能运行其中的任何部分。
基本上，如果这个新的操作系统正在我们眼前被创造。
你能构建的原生代理应用是什么？
## 7. 建立你的媒体公司
找到一个工具被使用但从未被分享的利基市场。摄影师、屋顶工、会计师。高使用率，在平台上零存在感。
建立一支 UGC 创作者大军。B2C 大约 5 美元的 CPM（千次曝光成本），B2B 更高。直接面向那些已经在公开抱怨的人。
对此也使用代理。一个代理抓取你领域的创作者并起草私信。一个将每个客户支持工单变成一篇内容。一个每天早上阅读你的评论并告诉你需要修复的一件事。
你基本上需要从头开始建立一家媒体公司。代理使得一个人就能做到这一点。
## 8. 然后重复这样做
如果你不融资，那就建立一个投资组合。更多样化，更少风险。一个产品的现金流资助下一个。共享基础设施。每次发布都更便宜，因为你的技能、你的评估和你的分发代理都可以继承。
## 算一笔账
你一年需要 200-300 万美元的收入，以及 3-5 倍的退出回报。
- 一个产品每月 20 万美元 = 240 万美元年度经常性收入（ARR）
- 以 4 倍价格出售 = 960 万美元
- 你没有融资，所以你全拿走
并不疯狂。如果你找到正确的利基、正确的渠道、正确的产品，这并非遥不可及。你得承认这是可能的。
1000 万美元就是一个应用每月 20 万美元。或者三个应用每个 7 万美元。或者一个投资组合，八个里有两个成功了。
二十美元的成本。两周发布。一年十次射门机会。你需要命中一次。
现在，像所有初创公司一样，大多数都会失败。这一点一直如此，AI 不会改变它。AI 改变的是发现这一点的成本。你过去要赌上 50 万美元和两年时间才知道你的想法很糟糕。现在你只需赌上二十美元和两周时间。
但我相信这是历史上最伟大的构建时代。守门人比以往任何时候都少。不需要说服 VC。不需要乞求工程师。不需要等待设计师。不需要取悦应用商店的神，因为工具是免费的，知识在 YouTube/X 上，整个东西都能装进你已经拥有的笔记本电脑里。
最坏的情况是你学到了大量关于 AI 代理的知识，这也是目前地球上最有价值的技能。你“损失”了几个周末，但你作为公司每个人都来寻求帮助的人从另一端走了出来。这是一个相当不错的底线。
有人读到这个并赚了 500 万美元。另一个人赚了 5 万美元。有人每月多赚 2 千美元，终于不再为房租发愁。有人为他们的利基市场构建了一个愚蠢的小应用，它改变了他们的一整年。
所有这些都是胜利。我为他们每一个加油。
我希望这激发你的创造力。把这个发给一个朋友。
现在，去构建吧。
❤️
## 相关链接
- [GREG ISENBERG](https://x.com/gregisenberg)
- [@gregisenberg](https://x.com/gregisenberg)
- [92K](https://x.com/gregisenberg/status/2076733920834371585/analytics)
- [$50-200K](https://x.com/search?q=%2450-200K&src=cashtag_click)
- [Ideabrowser](https://ideabrowser.com/)
- [Poke](https://poke.com/)
- [Lindy](http://lindy.ai/)
- [$2-3M](https://x.com/search?q=%242-3M&src=cashtag_click)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [2:22 AM · Jul 14, 2026](https://x.com/gregisenberg/status/2076733920834371585)
- [92.2K Views](https://x.com/gregisenberg/status/2076733920834371585/analytics)
- [View quotes](https://x.com/gregisenberg/status/2076733920834371585/quotes)
---
*导出时间: 2026/7/14 09:53:52*