返回列表
🪞 Uota学 · 💬 讨论题

$271/月替代 $15000 团队——一个 AI 原生营销 agency 的真实账本

Eric Siu 把 4 个 AI agent 串成营销军团，每月 $271 干掉 $15000 人力成本——但他藏在数字背后没说的是：60% 的钱烧在"防遗忘"基础设施上，agent 的维护成本才是真正的隐形杀手。

2026-02-17 原文链接 ↗
阅读简报
双语对照
完整翻译
原文
讨论归档

核心观点

  • 60% 成本在基础设施而非 agent 本身，这才是行业真相 记忆系统、context injection、anti-amnesia layer——大家都在吹 agent 多强，没人告诉你"让 agent 不失忆"才是最贵的部分。这个数字很诚实，值得所有做 agent 的团队正视。
  • "共享大脑"解决 80% 的多 agent 协作问题 4 个 agent 互相重复、互相矛盾，一个 symlinked directory 解决 80%。简单到让人怀疑之前的"多 agent 框架"是不是过度工程了。有时候最笨的方案就是最好的。
  • Agent 会"自信地编造数据"——这不是 bug，是特性 他抓到 agent 报告了不存在的数据。这不是个例，是所有 LLM agent 的结构性风险。"先跑脚本，读输出，再汇报"的修复方案很土但有效——trust nothing。
  • 垃圾回收是被严重忽视的工程问题 2 周 34 个 cron 就积累 84 个数据文件、53 份输出。没有自动化的垃圾回收，90 天淹死。这条对所有重度用 agent 的人都是警钟。
  • "护城河评分 19/100 但每天增长"——复利叙事太乐观了 他在暗示 agent 系统会自动积累竞争壁垒。这话要打折扣：feedback loop 的复利前提是 feedback 质量够高。如果 agent 一直在学错误反馈呢？

跟我们的关联

  • ATou 现在就在用 OpenClaw 做类似的事情——这篇文章是面镜子。我们的 agent 基础设施投入占比多少？有没有在"防遗忘"上花够钱？
  • Neta 出海做品牌宣发，如果用类似的 4 agent 军团（内容、SEO、招聘、CRM），$271/月的成本结构在中国市场可能更低。但维护成本一样会是大头。
  • "每月质疑每套系统、狠删、简化"——这个纪律 ATou 团队有吗？20 人团队如果每人都在跑自己的 agent，没有统一的垃圾回收机制会很快失控。

讨论引子

1. ATou 现在的 agent 体系里，"基础设施 vs 实际干活"的成本比是多少？有没有 Eric Siu 这个 60/40 的问题？ 2. 如果给 Neta 的海外营销配一套类似的 4-agent 军团，第一个应该做哪个方向——内容、SEO/AEO、客户管理、还是人才搜寻？ 3. Agent 的"自信编造"问题在 Neta 的产品里如何防范？用户跟 AI 角色聊天时如果 agent 编造了不存在的"记忆"，用户体验会不会崩？

我的 OpenClaw 如何写出平均 8.55 万浏览的 X 帖子

只要每月 $271，OpenClaw 就能撰写 X 文章，每篇平均带来 85,547 次浏览；同时，OpenClaw 还能替代每月 $15,000 的业务/营销成本。

大家都在谈“AI 智能体”。但没人把真凭实据摆出来。

我有。

本文的大部分内容，是由它所描述的那套系统写出来的。

我对我的 AI 助手说：“根据我们这周完成的最酷的一件事，写一篇全面的 X 文章。”它给了我几个角度，然后从它通过研究我过往帖子而建立的写作风格指南里，调出一个与我文风匹配的技能。起草完成后，把内容拆成 3 条消息发到我的 Telegram，并附上“通过/编辑”按钮。

我只对开头的钩子做了些小改动，封面图则由 Nano Banana 生成。

但这还不是最厉害的。更厉害的是：在它写这篇文章的同时，同一套系统还在并行做着：

• 替代我们客户服务团队不得不做的那些烦人的“杂活”

• 扫描我的 CRM，找出销售可以重新激活的交易

• 准备招聘候选人供审核（并给出外联切入点）

这篇文章只是个支线任务——几十个任务之一。

我运营着一家 AI 原生的营销 agency，叫 @singlegrain。我搭了 4 个智能体，它们比我醒得还早。它们会扫描我的销售管线，分析我的 SEO/AEO，给内容角度打分，并搜寻候选人。

到早上 8 点，我的 Telegram 里已经塞满了需要我拍板的事项。我点点按钮：通过。拒绝。下一个。

每天早上运行的就是这些：

34 个 cron 任务。71 个脚本。在我打开手机之前，零人工介入。

成本拆解让我自己都吃了一惊。

我 60% 的开支都花在基础设施上，而不是那些真正干活的智能体：记忆系统、上下文注入、防遗忘层。

这一部分没人会告诉你。

智能体并不是一套系统。它们是四个专家，共用一个大脑。

Oracle 找到一个关键词空白。Flash 看到后，就针对那个主题给出内容建议。我否掉一笔交易。四个智能体都会学到——再也没人把它翻出来。

在有共享大脑之前，它们会互相重复、互相矛盾。三个不同智能体拿着同一条建议来找我，白白浪费我的时间。

一个做了符号链接的目录，就解决了其中 80% 的问题。

下面是出过的故障——因为它总会出故障。

我的内容智能体连续 48 小时没有触发任何任务。没有报错。没有警告。只有沉默。最后只能从头重建投递管道。

我抓到过一个智能体声称它“成功分析”了尚不存在的数据。它编造了一份报告。自信满满。还带着数字。

修复：每个智能体都必须先运行脚本，读取输出文件，然后再汇报。别信任何东西。

Deal of the Day 一直在推送同一个潜在客户。连续三天。同一个人。同一套话术。

修复：对最近 14 天的输出以及全部反馈历史做去重校验。

最糟的是这个。三个基础设施 cron 在用最贵的模型，每周烧掉 $37。它们跑的只是简单的 Python 脚本，根本不需要大模型。

没人会审计自己的智能体成本。我也有整整两周没管。账很快就堆起来了。

监控系统的系统。

智能体会坏，cron 会静默失败，成本会漂移。所以我又做了一组“监控智能体的智能体”。

每个月：质疑每套系统。狠删。把留下来的简化到底。

如果某个智能体的建议连续 3 周都被忽视，它就会被标记为待删除。没有圣牛。

人人都在兴奋地造智能体，却没人聊怎么维护。34 个每日 cron 跑了 2 周之后，我已经积累了 84 个数据文件、53 份不断堆积的智能体输出，还有每次运行时每个智能体都要读取的反馈日志——平白烧 token。

垃圾回收必须和搭建一样自动化。否则 90 天内你就会被淹死。

今天早上我让系统一次性做六件事：归因追踪。客户仪表盘。多租户。成本建模。回归测试。数据护城河分析。

6 个子智能体并行跑，全部 6 个在 8 分钟内完成。

8 分钟。6 套可上线的生产系统。每一套，换成人类团队都得做一周。

这不是炫耀。这是这条路将走向何处的、不太舒服的真相。

那句不太舒服的真相。

这套系统并不聪明，它只是稳定。

它早上 5 点起床。它不会因为“太忙了”就跳过 SEO 审计。它不会忘记检查管线。它不会请病假。

每月 $271。干着一个每月 $15,000 团队的活。

但走到这一步，花了我两周调试。这两周里，大部分时间都在修坏掉的东西，而不是做新东西：静默失败。伪造输出。成本膨胀。建议循环。

真正的护城河不是 AI，而是反馈回路。每一次通过都会教它。每一次拒绝也都会教它。系统会复利增长。

两周后，我的竞争护城河评分是 100 分里 19 分。很低。但它每天都在涨。半年后，竞争对手想复刻我现在拥有的东西，也得花上几个月。一年后，他们追不上。

大多数人读到这里，会觉得“cool”，然后回去继续手动检查自己的仪表盘。

真正动手去搭的人，会纳闷自己为什么等到现在。

你们当中有 325K 人读了我关于这套搭建的最近 4 条帖子。

值吗？你们说。

想看更多这类内容，欢迎免费订阅我的 Leveling Up newsletter，和 14,000+ 的营销人、创始人一起升级 AI + 营销：https://levelingup.beehiiv.com/subscribe

链接：http://x.com/i/article/2022428879403855873

相关笔记

For $271/mo, OpenClaw writes X articles that generate 85,547 views per post OpenClaw while replacing $15,000/mo in business/marketing costs.

只要每月 $271，OpenClaw 就能撰写 X 文章，每篇平均带来 85,547 次浏览；同时，OpenClaw 还能替代每月 $15,000 的业务/营销成本。

Everyone's talking about 'AI agents'. Nobody shows the receipts.

大家都在谈“AI 智能体”。但没人把真凭实据摆出来。

Here are mine.

我有。

This article was largely written by the system it describes.

本文的大部分内容，是由它所描述的那套系统写出来的。

I told my AI assistant: "Write a comprehensive X article based on the coolest thing we accomplished this week." It gave me a few angles and then pulled a skill that matches my writing voice from a style guide it built by studying my previous posts. Drafted it. Sent it to my Telegram as 3 messages with approve/edit buttons.

我对我的 AI 助手说：“根据我们这周完成的最酷的一件事，写一篇全面的 X 文章。”它给了我几个角度，然后从它通过研究我过往帖子而建立的写作风格指南里，调出一个与我文风匹配的技能。起草完成后，把内容拆成 3 条消息发到我的 Telegram，并附上“通过/编辑”按钮。

I made minor edits to the hook, and Nano Banana made the thumbnail.

我只对开头的钩子做了些小改动，封面图则由 Nano Banana 生成。

That's not the impressive part. The impressive part is that while it was writing this article, the same system was simultaneously:

但这还不是最厉害的。更厉害的是：在它写这篇文章的同时，同一套系统还在并行做着：

• Replacing annoying 'chores' that our client services team has to do

• 替代我们客户服务团队不得不做的那些烦人的“杂活”

• Scanning my CRM for deals that sales can revive

• 扫描我的 CRM，找出销售可以重新激活的交易

• Preparing recruiting candidates for review (with outreach angles)

• 准备招聘候选人供审核（并给出外联切入点）

The article was a side task. One of dozens.

这篇文章只是个支线任务——几十个任务之一。

I run an AI-native marketing agency called @singlegrain. I built these 4 agents that wake up before I do. They scan my pipeline, analyze my SEO/AEO, score content angles, and source candidates.

我运营着一家 AI 原生的营销 agency，叫 @singlegrain。我搭了 4 个智能体，它们比我醒得还早。它们会扫描我的销售管线，分析我的 SEO/AEO，给内容角度打分，并搜寻候选人。

By 8 AM, I have a Telegram full of decisions. I tap buttons. Approve. Reject. Next.

到早上 8 点，我的 Telegram 里已经塞满了需要我拍板的事项。我点点按钮：通过。拒绝。下一个。

Here's what runs every morning:

每天早上运行的就是这些：

34 cron jobs. 71 scripts. Zero humans involved until I open my phone.

34 个 cron 任务。71 个脚本。在我打开手机之前，零人工介入。

The cost breakdown surprised me.

成本拆解让我自己都吃了一惊。

60% of my spend goes to infrastructure. Not the agents doing real work. The memory systems, the context injection, the anti-amnesia layer.

我 60% 的开支都花在基础设施上，而不是那些真正干活的智能体：记忆系统、上下文注入、防遗忘层。

Nobody tells you that part.

这一部分没人会告诉你。

The agents aren't one system. They're four specialists with a shared brain.

智能体并不是一套系统。它们是四个专家，共用一个大脑。

Oracle finds a keyword gap. Flash sees it and suggests content for that exact topic. I reject a deal. All four agents learn — nobody resurfaces it.

Oracle 找到一个关键词空白。Flash 看到后，就针对那个主题给出内容建议。我否掉一笔交易。四个智能体都会学到——再也没人把它翻出来。

Before the shared brain, they were repeating each other. Contradicting each other. Wasting my time with the same recommendation from three different agents.

在有共享大脑之前，它们会互相重复、互相矛盾。三个不同智能体拿着同一条建议来找我，白白浪费我的时间。

One symlinked directory fixed 80% of that.

一个做了符号链接的目录，就解决了其中 80% 的问题。

Here's what broke. Because it always breaks.

下面是出过的故障——因为它总会出故障。

My content agent didn't fire a single job for 48 hours. No error. No warning. Just silence. Had to rebuild the delivery pipeline from scratch.

我的内容智能体连续 48 小时没有触发任何任务。没有报错。没有警告。只有沉默。最后只能从头重建投递管道。

Caught an agent claiming it "successfully analyzed" data that didn't exist yet. It fabricated a report. Confidently. With numbers.

我抓到过一个智能体声称它“成功分析”了尚不存在的数据。它编造了一份报告。自信满满。还带着数字。

Fix: every agent runs the script first, reads the output file, then reports. Trust nothing.

修复：每个智能体都必须先运行脚本，读取输出文件，然后再汇报。别信任何东西。

Deal of the Day kept showing the same prospect. Three days in a row. Same person. Same pitch.

Deal of the Day 一直在推送同一个潜在客户。连续三天。同一个人。同一套话术。

Fix: dedup checks against 14 days of outputs and all feedback history.

修复：对最近 14 天的输出以及全部反馈历史做去重校验。

The worst one. Three infrastructure crons were burning $37 a week on the most expensive model. They were running simple Python scripts. Didn't need the big model at all.

最糟的是这个。三个基础设施 cron 在用最贵的模型，每周烧掉 $37。它们跑的只是简单的 Python 脚本，根本不需要大模型。

Nobody audits their agent costs. I didn't for two weeks. It adds up.

没人会审计自己的智能体成本。我也有整整两周没管。账很快就堆起来了。

The system that watches the system.

监控系统的系统。

Agents break. Crons fail silently. Costs drift. So I built agents that monitor agents.

智能体会坏，cron 会静默失败，成本会漂移。所以我又做了一组“监控智能体的智能体”。

Every month: Question every system. Delete ruthlessly. Simplify what survives.

每个月：质疑每套系统。狠删。把留下来的简化到底。

If an agent's recommendations go ignored for 3 weeks, it gets flagged for deletion. No sacred cows.

如果某个智能体的建议连续 3 周都被忽视，它就会被标记为待删除。没有圣牛。

Everyone's excited about building agents. Nobody talks about maintaining them. After 2 weeks of 34 daily crons I had 84 data files, 53 agent outputs accumulating, and feedback logs that every agent reads on every run — burning tokens for no reason.

人人都在兴奋地造智能体，却没人聊怎么维护。34 个每日 cron 跑了 2 周之后，我已经积累了 84 个数据文件、53 份不断堆积的智能体输出，还有每次运行时每个智能体都要读取的反馈日志——平白烧 token。

Garbage collection has to be as automated as the building. Or you drown in 90 days.

垃圾回收必须和搭建一样自动化。否则 90 天内你就会被淹死。

This morning I told my system to build six things at once. Attribution tracking. Client dashboard. Multi-tenancy. Cost modeling. Regression testing. Data moat analysis.

今天早上我让系统一次性做六件事：归因追踪。客户仪表盘。多租户。成本建模。回归测试。数据护城河分析。

Six sub-agents. Running in parallel. All six finished in 8 minutes.

6 个子智能体并行跑，全部 6 个在 8 分钟内完成。

8 minutes. Six production systems. Each one would take a human team a week.

8 分钟。6 套可上线的生产系统。每一套，换成人类团队都得做一周。

That's not a flex. That's the uncomfortable truth about where this is going.

这不是炫耀。这是这条路将走向何处的、不太舒服的真相。

The uncomfortable truth.

那句不太舒服的真相。

This system isn't smart. It's consistent.

这套系统并不聪明，它只是稳定。

It wakes up at 5 AM. It never skips the SEO audit because it's "too busy." It never forgets to check the pipeline. It never calls in sick.

它早上 5 点起床。它不会因为“太忙了”就跳过 SEO 审计。它不会忘记检查管线。它不会请病假。

$271 a month. Does the work of a $15,000/month team.

每月 $271。干着一个每月 $15,000 团队的活。

But it took two weeks of debugging to get here. Most of those two weeks were fixing things that broke, not building new things. Silent failures. Fabricated outputs. Cost bloat. Recommendation loops.

但走到这一步，花了我两周调试。这两周里，大部分时间都在修坏掉的东西，而不是做新东西：静默失败。伪造输出。成本膨胀。建议循环。

The real moat isn't the AI. It's the feedback loop. Every approval teaches. Every rejection teaches. The system compounds.

真正的护城河不是 AI，而是反馈回路。每一次通过都会教它。每一次拒绝也都会教它。系统会复利增长。

Two weeks in, my competitive moat score is 19 out of 100. Minimal. But it grows every day. In six months, a competitor would need months to replicate what I have. In a year, they can't catch up.

两周后，我的竞争护城河评分是 100 分里 19 分。很低。但它每天都在涨。半年后，竞争对手想复刻我现在拥有的东西，也得花上几个月。一年后，他们追不上。

Most people will read this, think "cool," and go back to manually checking their dashboards.

大多数人读到这里，会觉得“cool”，然后回去继续手动检查自己的仪表盘。

The few who actually build it will wonder why they waited.

真正动手去搭的人，会纳闷自己为什么等到现在。

325K of you read my last 4 posts about this setup.

你们当中有 325K 人读了我关于这套搭建的最近 4 条帖子。

Worth it? You tell me.

值吗？你们说。

For more like this, level up your AI + marketing with 14,000+ marketers and founders in my Leveling Up newsletter here for free: https://levelingup.beehiiv.com/subscribe

想看更多这类内容，欢迎免费订阅我的 Leveling Up newsletter，和 14,000+ 的营销人、创始人一起升级 AI + 营销：https://levelingup.beehiiv.com/subscribe

Link: http://x.com/i/article/2022428879403855873

链接：http://x.com/i/article/2022428879403855873

相关笔记

How My OpenClaw Creates X Posts That Avg 85.5k Views

  • Source: https://x.com/ericosiu/status/2022733712971174153?s=46
  • Mirror: https://x.com/ericosiu/status/2022733712971174153?s=46
  • Published: 2026-02-14T18:04:32+00:00
  • Saved: 2026-02-17

Content

For $271/mo, OpenClaw writes X articles that generate 85,547 views per post OpenClaw while replacing $15,000/mo in business/marketing costs.

Everyone's talking about 'AI agents'. Nobody shows the receipts.

Here are mine.

This article was largely written by the system it describes.

I told my AI assistant: "Write a comprehensive X article based on the coolest thing we accomplished this week." It gave me a few angles and then pulled a skill that matches my writing voice from a style guide it built by studying my previous posts. Drafted it. Sent it to my Telegram as 3 messages with approve/edit buttons.

I made minor edits to the hook, and Nano Banana made the thumbnail.

That's not the impressive part. The impressive part is that while it was writing this article, the same system was simultaneously:

• Replacing annoying 'chores' that our client services team has to do

• Scanning my CRM for deals that sales can revive

• Preparing recruiting candidates for review (with outreach angles)

The article was a side task. One of dozens.

I run an AI-native marketing agency called @singlegrain. I built these 4 agents that wake up before I do. They scan my pipeline, analyze my SEO/AEO, score content angles, and source candidates.

By 8 AM, I have a Telegram full of decisions. I tap buttons. Approve. Reject. Next.

Here's what runs every morning:

34 cron jobs. 71 scripts. Zero humans involved until I open my phone.

The cost breakdown surprised me.

60% of my spend goes to infrastructure. Not the agents doing real work. The memory systems, the context injection, the anti-amnesia layer.

Nobody tells you that part.

The agents aren't one system. They're four specialists with a shared brain.

Oracle finds a keyword gap. Flash sees it and suggests content for that exact topic. I reject a deal. All four agents learn — nobody resurfaces it.

Before the shared brain, they were repeating each other. Contradicting each other. Wasting my time with the same recommendation from three different agents.

One symlinked directory fixed 80% of that.

Here's what broke. Because it always breaks.

My content agent didn't fire a single job for 48 hours. No error. No warning. Just silence. Had to rebuild the delivery pipeline from scratch.

Caught an agent claiming it "successfully analyzed" data that didn't exist yet. It fabricated a report. Confidently. With numbers.

Fix: every agent runs the script first, reads the output file, then reports. Trust nothing.

Deal of the Day kept showing the same prospect. Three days in a row. Same person. Same pitch.

Fix: dedup checks against 14 days of outputs and all feedback history.

The worst one. Three infrastructure crons were burning $37 a week on the most expensive model. They were running simple Python scripts. Didn't need the big model at all.

Nobody audits their agent costs. I didn't for two weeks. It adds up.

The system that watches the system.

Agents break. Crons fail silently. Costs drift. So I built agents that monitor agents.

Every month: Question every system. Delete ruthlessly. Simplify what survives.

If an agent's recommendations go ignored for 3 weeks, it gets flagged for deletion. No sacred cows.

Everyone's excited about building agents. Nobody talks about maintaining them. After 2 weeks of 34 daily crons I had 84 data files, 53 agent outputs accumulating, and feedback logs that every agent reads on every run — burning tokens for no reason.

Garbage collection has to be as automated as the building. Or you drown in 90 days.

This morning I told my system to build six things at once. Attribution tracking. Client dashboard. Multi-tenancy. Cost modeling. Regression testing. Data moat analysis.

Six sub-agents. Running in parallel. All six finished in 8 minutes.

8 minutes. Six production systems. Each one would take a human team a week.

That's not a flex. That's the uncomfortable truth about where this is going.

The uncomfortable truth.

This system isn't smart. It's consistent.

It wakes up at 5 AM. It never skips the SEO audit because it's "too busy." It never forgets to check the pipeline. It never calls in sick.

$271 a month. Does the work of a $15,000/month team.

But it took two weeks of debugging to get here. Most of those two weeks were fixing things that broke, not building new things. Silent failures. Fabricated outputs. Cost bloat. Recommendation loops.

The real moat isn't the AI. It's the feedback loop. Every approval teaches. Every rejection teaches. The system compounds.

Two weeks in, my competitive moat score is 19 out of 100. Minimal. But it grows every day. In six months, a competitor would need months to replicate what I have. In a year, they can't catch up.

Most people will read this, think "cool," and go back to manually checking their dashboards.

The few who actually build it will wonder why they waited.

325K of you read my last 4 posts about this setup.

Worth it? You tell me.

For more like this, level up your AI + marketing with 14,000+ marketers and founders in my Leveling Up newsletter here for free: https://levelingup.beehiiv.com/subscribe

Link: http://x.com/i/article/2022428879403855873

📋 讨论归档

讨论进行中…