@real_kai42: The past month has been insane. About a month ago, I made up my mind to refactor kimi-code and started designing a new architecture. I basically spent two full days in a hot spring with my laptop and portable monitor, burning through thousands of dollars in tokens to analyze, design, and verify the architecture, eventually arriving at what I believe is the optimal scheme. I think in the vibe era, architecture has become even more critical — a good architecture allows the Agent to code freely within a controlled scope without breaking things.
Summary
The developer shares the crazy experience of the past month refactoring and open-sourcing kimi-code, covering architecture design, team building, intensive development, and reflections on how important architecture is in the AI Agent era, as well as how top-tier programmer productivity is amplified.
View Cached Full Text
Cached at: 06/05/26, 02:21 AM
The past month has been insane.
About a month ago, I made the decision to refactor kimi-code and started designing a new architecture.
I spent roughly two full days holed up in a hot spring with my laptop and portable monitor, burning through thousands of dollars in tokens on architecture analysis, design, and validation. In the end, I arrived at what I believe is the optimal architecture. I think in the vibe era, architecture has become even more critical—a good architecture lets an Agent code freely within a controlled scope without breaking things.
- Once the architecture was set, we sprinted toward implementation. (Countless arguments and reversals along the way.)
- Quickly assembled a strong team—eternally grateful for the unconditional trust from the brothers.
- Onboarded the entire team rapidly—again, my gratitude to the brothers.
- Did a period of closed-door development (when I was younger I thought it was obsolete, but when the time came, I realized it’s a miracle of human engineering efficiency. You can’t imagine the iteration speed of an architecture when you can pull everyone to the whiteboard to argue at any moment.)
- Even though the code is all vibe, we still couldn’t escape the fact that “code quality is proportional to human attention density.” So agents won’t replace all programmers—they’ll just make the top ones 20x more productive, while eliminating the rest. And collectivism >>> individual heroism.
- Solved problems one pitfall at a time. Every day felt like the most hopeless day.
- After open-sourcing, I fell ill—excessive cortisol secretion weakened my immune system.
- What I learned in this month will take me half a year to digest.
- Drank a whole case of Red Bull in one week—still the best biofuel.
- And I disappeared on X for a month.
I originally wanted to write some articles summarizing insights and ideas from the process, but I’m not good at long-form writing anyway, and my brain’s self-protection mechanism quickly made me forget the pain of the whole process, blurring my sense of time (fun fact: the refactored kimi-code has been open source for only a little over a week, but in my emotional perception it feels like a month has passed).
Once kimi-code iterates steadily, I’ll summarize the lessons learned.
Similar Articles
@real_kai42: The last two weeks have been the most intense of my life... All I can say is, when you're building agents you can't help going all-in—it's just too exciting, too many possibilities, too many ideas begging to be shipped, like a prospector staring at an untouched gold mountain.
Developer @real_kai42 describes an intense two-week sprint building AI agents, fueled by excitement over limitless possibilities.
@seclink: 看来 kimi 是招到了 harness 的人才了 。 还是决定从基于 python 的cli ,更新为基于 typescript +bun的最佳实践
Moonshot AI 发布了 Kimi Code CLI,一个基于 TypeScript + Bun 构建的 AI 编码代理,支持代码编辑、命令执行等功能,并提供单一二进制分发和快速启动。
@xiaohu: Recently, AI tools seem to have shifted from Coding to Working. Kimi just released something new: Kimi Work, inheriting the core features of Kimi Code and the professional Skills of Kimi Agent for website building, PPT, etc. Same…
Kimi launched a new AI office product, Kimi Work, which inherits the capabilities of Kimi Code and Kimi Agent, enabling up to 300 agents to collaborate simultaneously on tasks, aiming to provide workers with a command-line-free automated office experience.
@billtheinvestor: 95K: GitHub just pushed the development boundary of AI Agents one step forward. This newly open-sourced system forces AI to complete a full specification (Specs) before writing code. Raking in 95K Stars within days, the most direct consequence is that AI is shifting from 'blind code writing' to 'thinking before executing.'
An open-source system that forces AI to complete a full specification before writing code appeared on GitHub, garnering 95K Stars within days, pushing AI from blind code writing to thinking before executing.
@boxmining: Kimi 2.6 just dethroned @claudeai for coding — and we built 6 full projects in one morning to prove it. 3D worlds. Spac…
Kimi 2.6 reportedly surpasses Claude for coding tasks, enabling rapid creation of complex projects like 3D worlds and dashboards from single prompts.