Kimi K2.6 代理蓝图:一人团队的 8 万美元月收入公式
本文介绍使用 Kimi K2.6 搭建单人 AI 代理公司的完整方法。Kimi K2.6 采用 MoE 架构,总参数 1 万亿,激活 32B,SWE-Bench 得分 65.8,内置工具调用。其 Agent Swarm 可并行运行 300 个子代理,单次运行产出 100+ 文件。作者提供了一套操作路径:技术栈包括 Kimi API、CLI、Swarm、MCP 服务器、n8n;服务线包括获客系统、知识库、客服自动化等;客户获取通过监控招聘信息并自动生成个性化方案;成本模型显示月开销 $500,月利润可达 $72k-75k。整体偏向营销风格,营收数据未经验证。
Kimi K2.6 can replace an entire dev team. here's the exact blueprint
By @Asteri_eth · 2026-05-29T08:33:09.000Z

300 agents. $500/mo overhead. one person. $80k/mo. the math is real.
Most agencies have 10-15 people.
A PM, several devs, a designer, a copywriter, QA.
The client pays $15k-50k per project most of that goes to salaries of people doing repetitive work that a model can now handle for the cost of API tokens.
Kimi K2.6 just made that entire model obsolete.
Kimi K2.6 可以取代整个开发团队。以下是确切蓝图
作者 @Asteri_eth · 2026-05-29T08:33:09.000Z

300 个代理,每月 $500 开销,一个人,月入 $8 万。数学是真实的。
大多数代理机构有 10-15 人。
一个项目经理,几个开发人员,一个设计师,一个文案,一个 QA。
客户为每个项目支付 $15k-50k,其中大部分用于支付那些做重复性工作的人的工资,而这些工作现在一个模型仅用 API token 成本就能处理。
Kimi K2.6 让整个模式过时了。
WHAT IS KIMI K2.6
Most people building AI agencies use ChatGPT or Claude.
Those are excellent models but Kimi K2.6 was built from the ground up for agentic workflows.
It doesn't answer questions. it executes real multi-step tasks in real environments without stopping to ask for permission at every step.
Architecture: Mixture-of-Experts
Total parameters: 1 trillion
Activated per token: 32 billion
Context window: 128,000 tokens
SWE-Bench Verified: 65.8
LiveCodeBench v6: 53.7
Tool calling: native, built into architecture

SWE-Bench 65.8 it solves nearly 2/3 of real GitHub engineering problems without human input.
Every hour the model spends solving a problem is an hour you're not paying a developer at $80.
什么是 Kimi K2.6
大多数构建 AI 代理的人使用 ChatGPT 或 Claude。
那些是优秀的模型,但 Kimi K2.6 从底层为代理工作流而构建。
它不回答问题。它在真实环境中执行真正的多步骤任务,不需要每一步都停下来请求许可。
架构: 混合专家
总参数: 1 万亿
每 token 激活: 320 亿
上下文窗口: 128,000 tokens
SWE-Bench 验证:65.8
LiveCodeBench v6:53.7
工具调用: 原生,内置于架构

SWE-Bench 65.8:它可以在没有人工输入的情况下解决近三分之二的真实 GitHub 工程问题。
模型花在解决问题上的每一个小时,都是你不需要支付 $80 给开发人员的一小时。
THE AGENT SWARM
Most people are still using AI as a chat window.
Moonshot AI shipped something different.
Kimi K2.6 Agent Swarm: 300 parallel sub-agents, 4,000 steps per run.
K2.5 K2.6
————————————————————————————————————————
parallel agents 100 300
steps per run 1,500 4,000
output chat real files
One run delivers:
100+ files 100,000-word literature review 20,000-row dataset
Search, analysis, coding, writing, visual generation all running simultaneously.
Not a chatbot. a system.
Instead of sequential work (research → analysis → writing → code), you launch Swarm and everything runs in parallel.
Weeks of work → hours.

Agent Swarm(代理群)
大多数人仍在使用 AI 作为聊天窗口。
Moonshot AI 推出了不同的东西。
Kimi K2.6 Agent Swarm:300 个并行子代理,每次运行 4,000 步。
K2.5 K2.6
————————————————————————————————————————
并行代理 100 300
每次运行步数 1,500 4,000
输出 聊天 真实文件
一次运行交付:
100+ 文件、10 万字文献综述、2 万行数据集
搜索、分析、编码、写作、视觉生成同时运行。
不是聊天机器人。是一个系统。
代替顺序工作(研究 → 分析 → 写作 → 代码),你启动 Swarm,一切并行运行。
数周的工作 → 几小时。

AGENCY MODEL
The path to $80k/mo isn't landing one massive client.
It's building a repeatable delivery system and stacking retainers.
month 1-2: $8k-10k/mo - 2 clients, learning + templates
month 3-4: $15k-20k/mo - first retainer ($5k/mo)
month 5-6: $25k-35k/mo - mostly retainers, 70% automated
month 7-9: $45k-60k/mo - Swarm deployed, acquisition automated
month 10-12: $70k-80k/mo - 1 person, Kimi doing 80% of execution
[ OVERHEAD ]: $500-1,500/mo throughout [ TEAM SIZE ]: 1 person [ NET PROFIT AT $80k ]: $72k-75k
The key shift happens at month 4-5.
One-off projects become monthly retainers.
Kimi builds. you handle strategy and client relationships. recurring revenue compounds.
代理模型
达到 $80k/月的路径不是获得一个大客户。
而是建立一个可重复的交付系统并叠加 retainer(月度服务费)。
第1-2月: $8k-10k/月 - 2个客户,学习+模板
第3-4月: $15k-20k/月 - 第一个 retainer ($5k/月)
第5-6月: $25k-35k/月 - 主要是 retainer,70% 自动化
第7-9月: $45k-60k/月 - Swarm 部署,获客自动化
第10-12月: $70k-80k/月 - 1个人,Kimi 完成 80% 执行
[ 开销 ]: 全程 $500-1,500/月 [ 团队规模 ]:1人 [ $80k 时净利润 ]:$72k-75k
关键转变发生在第4-5个月。
一次性项目变成月度 retainer。
Kimi 构建。你处理策略和客户关系。经常性收入复利增长。
WHAT TO SELL
The biggest mistake: offering "AI solutions."
Nobody knows how to evaluate that.
Businesses pay for specific outcomes they can't achieve themselves and don't want to spend 6 months hiring someone to deliver.
Five services that convert best:
1/ Automated lead gen systems - $5k-10k
scrapes, qualifies, reaches out automatically every business with a sales team needs this almost none have it
2/ Internal knowledge bases - $8k-15k
indexes all company documentation employees get instant answers companies with 50+ people pay without hesitation
3/ Customer support automation - $5k-12k upfront + retainer
handles 70-80% of tickets without humans e-commerce is the easiest sell
4/ Data analysis pipelines - $3k-8k 5/ competitor monitoring systems $3k-8k
卖什么
最大的错误:提供“AI 解决方案”。
没人知道如何评估它。
企业为那些他们自己无法实现、又不想花 6 个月招聘人员来交付的具体结果买单。
五个转化最好的服务:
1/ 自动化潜在客户生成系统 - $5k-10k
自动抓取、筛选、联系客户;每个有销售团队的企业都需要,但几乎没有企业拥有
2/ 内部知识库 - $8k-15k
索引所有公司文档,员工可即时获得答案;拥有 50 人以上的企业会毫不犹豫地付费
3/ 客户支持自动化 - $5k-12k 前期费用 + retainer
处理 70-80% 的工单而不需要人工;电子商务是最容易销售的对象
4/ 数据分析管道 - $3k-8k 5/ 竞争对手监控系统 - $3k-8k
THE TECHNICAL STACK
Kimi K2.6 API - core reasoning + code generation
Kimi CLI - terminal agent for local dev
Kimi Swarm - parallel execution at scale
MCP servers - 14,000+ tools for external systems
GitHub MCP - repos + PR management
Postgres MCP - database operations
Slack MCP - client communication automation
n8n - workflow orchestration
Kimi CLI is the piece most people miss.
Point it at a codebase. describe what needs to be built. it figures out the architecture, writes the code, runs tests, reports what it did.
No setup ceremony. just output.
技术栈
Kimi K2.6 API - 核心推理 + 代码生成
Kimi CLI - 本地开发的终端代理
Kimi Swarm - 大规模并行执行
MCP 服务器 - 14,000+ 工具用于外部系统
GitHub MCP - 仓库 + PR 管理
Postgres MCP - 数据库操作
Slack MCP - 客户通信自动化
n8n - 工作流编排
Kimi CLI 是大多数人忽略的部分。
将它指向一个代码库。描述需要构建什么。它会找出架构、编写代码、运行测试、报告它做了什么。
无需设置仪式。只有输出。
SKILL INJECTION
Instead of retraining the model, you give it a markdown file at runtime.
It becomes a specialist for the duration of the task.
/skills/xlsx/SKILL.md → Excel specialist /skills/docx/SKILL.md → Word document expert /skills/webapp/SKILL.md → web app builder
Build a custom skill file per client vertical:
healthcare → HIPAA compliance fintech → financial regulations e-commerce → Shopify architecture
This library is the real competitive moat.
A competitor can't copy it in a week it's built from months of actual client work.
技能注入
无需重新训练模型,你可以在运行时给它一个 markdown 文件。
它在任务期间变成专家。
/skills/xlsx/SKILL.md → Excel 专家 /skills/docx/SKILL.md → Word 文档专家 /skills/webapp/SKILL.md → Web 应用构建者
为每个客户垂直领域构建自定义技能文件:
医疗保健 → HIPAA 合规 金融科技 → 金融法规 电子商务 → Shopify 架构
这个库是真正的竞争护城河。
竞争对手无法在一周内复制它——它来自数月实际客户工作的积累。
CLIENT ACQUISITION
Target identification: set an agent to monitor job listings daily.
Companies posting for "data analyst", "automation engineer", "Python developer" those are companies with a problem they're trying to hire their way out of.
Those are your prospects.
- WHAT THE AGENT DOES
reads their website, LinkedIn, recent news generates personalized outreach explains exactly what problem they have references how it was solved for similar companies
A detailed technical proposal that takes a consultant half a day to write takes Kimi K2.6 about 4 minutes.
One question closes more clients than anything else on discovery calls:
"what's the most repetitive thing your team does that feels like it shouldn't require a human?"
The answer is the first project.
客户获取
目标识别:设置一个代理每天监控招聘信息。
发布“数据分析师”、“自动化工程师”、“Python 开发者”等职位的公司,是那些正试图通过招聘解决问题的公司。
这些就是你的潜在客户。
- 代理做什么
读取他们的网站、LinkedIn、最新新闻 生成个性化外联 准确解释他们有什么问题 引用类似公司是如何解决的
一份详细的、需要顾问花半天时间撰写的技术提案,Kimi K2.6 只需大约 4 分钟。
在发现电话中,有一个问题比其他任何问题都能赢得更多客户:
“你的团队做的最重复、感觉不该需要人工的事情是什么?”
答案就是第一个项目。
THE MATH
Traditional agency delivering a $10k project:
developers (60h × $80): $4,800
project manager: $1,200
design + QA: $1,000
——————————————————————————————————
profit: $3,000 (30%)
AI agency delivering the same $10k project:
Kimi K2.6 API costs: $150-300
time (strategy + review): $600
tools + infrastructure: $100
——————————————————————————————————
profit: $9,000 (90%)
$72k monthly profit. one person. $500 in overhead.
A traditional agency owner nets $15k-20k from the same revenue after salaries, office, overhead.
The bottleneck in an AI agency is never technical execution.
Kimi handles that.
The bottleneck is understanding what the client actually needs, setting expectations correctly, making sure the output solves the real problem.
That part is still human and it's worth $70k-80k/mo when the execution layer costs $500.
This is probably just the beginning.
Sources:
github.com/MoonshotAI/Kimi-K2 github.com/MoonshotAI/kimi-cli arxiv.org/pdf/2507.20534
If this was useful follow @Asteri_eth
More on AI tools, agencies, and what's actually worth building
数学计算
传统代理交付一个 $10k 项目:
开发人员(60h × $80): $4,800
项目经理: $1,200
设计 + QA: $1,000
——————————————————————————————————
利润: $3,000(30%)
AI 代理交付同一个 $10k 项目:
Kimi K2.6 API 成本: $150-300
时间(策略 + 审查): $600
工具 + 基础设施: $100
——————————————————————————————————
利润: $9,000(90%)
月利润 $72k。一个人。$500 开销。
传统代理机构老板在扣除薪水、办公室、开销后,从相同收入中净赚 $15k-20k。
AI 代理的瓶颈从来不是技术执行。
Kimi 处理了它。
瓶颈在于理解客户实际需求、正确设定期望、确保输出解决真正的问题。
那部分仍然是人类的,并且当执行层只花 $500 时,它值 $70k-80k/月。
这很可能只是开始。
来源:
github.com/MoonshotAI/Kimi-K2 github.com/MoonshotAI/kimi-cli arxiv.org/pdf/2507.20534
如果本文有用,请关注 @Asteri_eth
更多关于 AI 工具、代理以及真正值得构建的东西