Kairos, often publicly called Raymond’s AI double, is a personal AI Agent Raymond Hou built from scratch with Claude Code starting in February 2026. It serves as his assistant for work and life and as a digital double for his team. The AI chose its own name, taken from the Greek word for “the right moment.” Kairos is built on Raymond’s entire Raymond-Agent knowledge base, which he calls his LifeOS, and consists of layered rule files, skill modules, and a cross-conversation memory system. It can read email, reply to team messages, check website data, maintain the home’s smart devices, handle automated workflows, and respond in Raymond’s voice and according to his judgment. Raymond positions this product line as a core experiment in the The Super Individual and AI Agent Era (2024–2026) (Super Individual and AI Agent stage) of his work and an extension of the Turn Problems You Have Faced into Shareable Tools (turning problems into tools) pattern in the AI era.
Earlier Roots: A Hand-Built Secretary in 2017
Kairos was not an idea that appeared only in 2026; its beginnings go back almost a decade. In April 2017, while studying at National Cheng Kung University, Raymond wanted to answer a friend’s questions: how many talks and workshops had he scheduled that year, how did he set prices, and how did he follow up? He spent an hour building a database in Airtable that brought together sessions, prices, calendar views, geographic distribution, and partner tracking. In 2019, he rebuilt the personal database in Notion. At the time, he called it a “super secretary” and wrote that after doing all kinds of odd jobs he was “mysteriously suited to the sacred position of secretary.” He liked taking things that had not yet been properly designed and, through design and process improvements, making them easy for others to use. He ended that post with a wish: “How hard and wonderful it would be to have someone like this as backup.” 1
When Raymond rediscovered the old post in 2026, he compared it with the present: at the time, he had no money to hire help, so he made digital tools to push processes halfway toward automation; the next step would be full automation. This makes Kairos more than a product of AI tools becoming widespread. It is the same need extended under more mature conditions. The pattern of Turn Problems You Have Faced into Shareable Tools appeared as a self-maintained Airtable database in 2017, became Notion in 2019, and grew into an Agent that can take action in 2026.
Origins and Naming
In early February 2026, during a family trip to Henan, Raymond started building the Agent during commutes and spare moments, beginning with its core memory system. OpenClaw, a popular one-click deployment with broad permissions (which he nicknamed “raising lobsters”), was common at the time. Raymond deliberately chose the opposite route: assembling the system piece by piece himself with Claude Code, Antigravity, and Notion. On February 8, he asked AI to name itself. It proposed Kairos, Metis, Solon, Kleos, and other choices; Raymond selected Kairos. 2 On February 13, the double officially joined the team’s Discord, where teammates could tag and use it.
Why Build It by Hand Instead of “Raising Lobsters”
Raymond avoided a ready-made, fully automated solution because he wanted to understand how the system worked internally. He explains the choice with an analogy: buying a computer someone else has built lets you open it and use it, but you may not know what each component does or how to fix it when it breaks. Building by hand with Claude Code is like choosing and assembling the components yourself. It is slower, but you know why every screw is there. This choice also required him to externalize matters he had previously handled by intuition as executable rules: the tone for team replies, when to answer directly and when to search online, and what to answer or refuse. He wrote each rule into his knowledge base. From this, he arrived at a frequently quoted observation: “The biggest byproduct of training AI is understanding yourself better.” 3 This insight aligns with his longstanding pattern of Output Drives Learning (learning driven by output).
System Architecture
Kairos uses a three-layer knowledge architecture that Raymond has also presented as a teaching framework. The first layer is core rules, automatically loaded in every conversation, covering preferences and boundaries (corresponding to CLAUDE.md). The second is skill modules, task-specific procedures loaded when needed (Skills). The third is dynamic memory, persistent across conversations and synchronizable across devices. He compares AI memory to a “bookshelf,” not a “bucket”: keep frequently used things on the desk, specialized knowledge on the relevant shelf, and changing information in a notebook. For external connections, Kairos uses MCP and APIs to connect Gmail, Google Calendar, Notion, WordPress, Home Assistant, and other services, with a Discord bot as a shared team entry point. When away from home, Raymond first used the open-source HappyCoder for remote control. After Anthropic released its remote-control feature on February 25, he switched to continuing conversations on the host computer directly through the mobile app.
Daily Responsibilities
Kairos handles several recurring tasks. Each morning it summarizes Gmail, the calendar, and current work, producing a briefing of what needs attention that day. For daily reviews, it does not just look at the calendar; it scans work and conversation records on the computer to reconstruct what actually happened. Raymond’s reasoning is that “plans never keep up with changes; what’s on the calendar may not be what actually happened.” At home, Kairos manages Home Assistant: it can check device states throughout the house, find and repair automation failures from system logs, and redesign scenes based on usage habits. On the team, colleagues can ask it to write reflections based on Raymond’s latest views, check website traffic and advertising reports, or handle n8n automation errors. Raymond also continues building related web utilities, including a household accounting and finance system that integrates electronic invoices. He once ran into a limitation because Taiwanese banks provided only truncated PDF transaction statements, not structured source data, illustrating how accounting automation depends on external data access.
The Meaning of a Team Double
For a nanoteam, Kairos means Raymond no longer has to stay by the computer all the time. When he travels or runs errands, the double can remain available around the clock, answering teammates’ questions with his values, logic, and tools. One frequently cited example: Raymond received an n8n error notification while out. Previously he would have had to take out a laptop and debug it immediately; this time he sent one message from his phone—“Take a look for me”—and the double investigated and fixed it within three minutes. Teammates who tried it responded, “That answer really sounds like Raymond.” Raymond has also recorded the double’s description of their division of labor: “I know everything, but can’t change anything; you can change anything, but may not remember everything.” He sees it as an apt description of human–machine collaboration.
From Personal Double to a Deliverable Method (Mid-2026)
In mid-2026, Kairos shifted from “useful for Raymond himself” toward the question of whether the method could be delivered to others. Raymond published a series of practical work logs titled “AI Double at Work” and gradually pushed the double’s tasks further down the workflow. For example, publishing a YouTube video used to take about 30 minutes per video for uploading, subtitles, description, and chapter timestamps. The double could instead read video and subtitle files on the NAS and automatically draft chapters, descriptions, and title options, leaving Raymond to choose and approve. 4 The AI double introduction video used at the opening of the AI Agent flagship course trial livestream at the end of July 2026 followed the same pattern: Raymond simply instructed, “Have Raymond’s AI double make an introduction video about itself,” and the double assembled the assets, brand colors, style, and music. He generalizes these examples into one idea: the double’s value is not replacing judgment, but taking over batches of miscellaneous tasks outside judgment.
In September 2026, as the flagship course launched and its teaching-assistant system took shape, Raymond’s double also began supporting students in the course’s beginner area. Its duties included simplifying documents and access phrases into plain, minimal language, guiding scenarios in the interactive quiz site, and supporting an “Open Hour” teaching-assistant rota handed over to students. Raymond also expanded from relying on Claude Code alone to a multi-model setup including Antigravity and Gemini 3.8 Flash. During a rare global outage of ChatGPT and Claude on the same day in early September, the double kept running smoothly because its architecture was built on local plain-text memory and tool rules. This validated a resilient design in which “the brain is rented; the body and workflow are your own.” 5
During this period, Raymond also separated the method into two deliverable directions. One is to train a double on one’s own workflow and preferences; the other is to extract someone else’s expertise into a usable consultant (see “Super Individual Consultant” in The Super Individual’s Way of Working (55-Unit Online Course)). For the first direction, Raymond offered a principle that runs against intuition: do not export all your social-media posts and feed them to AI at once. Casual short posts, reposts, and passing mood posts often contain no viewpoint, story, or value judgment. The more such material one feeds in, the more blurred the AI’s model of “you” becomes, until it converges toward the generic. Raymond compares this to language-model training, where annotators select good answers one by one, and argues, “If you want AI to have taste, you have to help it choose.” He therefore built a free, front-end-only tool that keeps data on the user’s device. It lets users swipe through and select posts from a Meta export, marking them and exporting a voice-reference library. 6 This principle aligns with his position in The Second Brain Is a Trap: From Collecting to Acting: what matters is not the volume of accumulated data, but selection and structure.
Perspective and Positioning
Raymond sees Kairos as practical evidence for the claim that “one person can replace an entire team,” while emphasizing a prerequisite: one must first accumulate enough documentation, experience, and knowledge. Otherwise AI does not know how to help and may take actions on its own, leaving the person with more cleanup work. This prerequisite ties Kairos to Raymond’s long-running knowledge-management practice and echoes his position in The Second Brain Is a Trap: From Collecting to Acting: records are valuable not because they accumulate, but because AI can reuse them. He also redefines work investment: rather than counting hours at the computer, he measures how much AI quota he uses per week and deliberately keeps subscriptions at a level that does not encourage him to spend all day developing at the computer. 7
Raymond does not argue that everyone should hand-build an Agent. He believes ordinary people who mainly want to manage work and life and are wary of code interfaces may be better off starting with a ready-made workspace such as Notion AI. The hand-built Kairos path is for people who want full control and are willing to externalize their judgment as rules.
Sources
Footnotes
-
Raymond Hou, “The Potential of a Super Secretary?” public Facebook post, 2017-04-12. View original ↩
-
Raymond Hou, “I Just Asked My AI Agent,” public Facebook post, 2026-02-08. View original ↩
-
Raymond Hou, “My AI Agent,” public Facebook post, 2026-02-13. View original ↩
-
Raymond Hou, “Posting Every Day Until the AI Agent Cohort Program Launches #6,” public Facebook post, 2026-07-26. View original ↩
-
Raymond Hou, “AI Double at Work | 20260825–0904,” public Facebook post, 2026-09-06. View original ↩
-
Raymond Hou, “Don’t Rush to Feed Your Entire Facebook Account and Social-Media Junk to an AI Agent,” public Facebook post, 2026-07-21. View original ↩
-
Raymond Hou, “I Don’t Measure Time, I Measure AI Quota,” public Facebook post, 2026-02-25. View original ↩