Raymond's User Manual

A Full Shift to AI Agent Collaboration in 2026

6 min read

This was a change in working style that began in early 2026. Raymond Hou (Raymond) shifted his use of AI from “a chat and outsourcing tool used by switching among graphical interfaces” to “AI Agent collaboration centered on the terminal and sharing one local memory across devices.” The defining event was building his personal AI counterpart, “Kairos” (publicly nicknamed “Raymond the Little Mong”), from scratch in February 2026 and bringing it into the team collaboration environment. Raymond frames the transition as evolving from “using AI tools” to “raising an AI counterpart.”

Ending: Treating AI as a Chatbot and Switching Among Interfaces

Before this transition, Raymond, like most people, mainly used AI for “chat and Q&A”: type a question in a web page or desktop app and get an answer from models such as ChatGPT or Gemini, then copy and paste code or plans, run them, and paste errors back into the chat. Raymond later compared this to “hiring a consultant to draw a construction plan while you still carry all the bricks to build the house.” At the same time, tools that could “take action,” such as Claude Cowork (a desktop assistant) and graphical IDEs such as Antigravity, lowered the barrier but tied users to their own windows and accounts. Switch to another chat window and it forgets; tools and memory are scattered across interfaces. This was the state the transition would leave behind: AI was useful but fragmented, passive, and insufficiently familiar with the user.

Drift: Building an AI Counterpart from Scratch in February 2026

The central part of the transition happened during a February 2026 trip with family in Henan, China, when Raymond could work only during commutes and in spare moments. Open-source Agent frameworks packaged for direct deployment (OpenClaw, which Raymond jokingly called “raising a lobster”) were popular in the community. They came with all permissions enabled and worked immediately after installation. Raymond deliberately chose the opposite path: using Claude Code in the terminal to build his ideal assistant from scratch, layer by layer. He wanted to understand “what the system is actually doing,” rather than adopt a machine someone else had assembled and not know how to fix it when it broke.

Key actions during this transition period were concentrated in February:

  • February 8: Raymond asked AI to name itself and chose “Kairos” from a list of candidates—the Greek word for “the right moment”—then built the core memory mechanism first.1
  • The first week (February 8–14): He completed a three-layer memory system (refined summaries, knowledge relationships, and daily logs). Its core design principle was “AI doesn’t need to remember every detail, but it needs to know where to find it.” At the same time, he deployed a Discord bot on a Mac mini at home, shifting interaction from “I go find the AI” to “the AI comes to me.”
  • February 13: Kairos went live in the team Discord as Raymond’s counterpart, available for teammates to use at any time.2
  • The second and third weeks: He gradually delegated everyday tasks to the Agent, including a Gmail assistant, household expense tracking, and Home Assistant smart-home maintenance. The system also began conducting its own review and suggesting improvements on the day its weekly quota refreshed.

The tools remained mixed during the process. Raymond used the open-source tool HappyCoder with tmux for a time to control the system remotely from his phone. Then, when Claude officially released its remote-control feature on February 25, he archived all his HappyCoder settings in about ten minutes and switched to the official option; a tutorial article he had not yet published became obsolete as a result.3 This kind of choice—“wait a little and let things settle”—recurred during the transition.

Emergence: A Collaboration Style Centered on the Terminal

A sign that the transition had stabilized was Raymond’s explicit choice in later public tutorials to “teach the terminal directly,” rather than starting with a friendlier IDE such as Antigravity. His judgment was that everyone would eventually move toward the terminal, so there was no need to take an extra detour. Supporting this choice was the beginner’s mindset he drew from learning touch typing: “be willing to let yourself get worse temporarily.” Giving up a familiar workflow is necessary to reach a higher ceiling.

The emerging collaboration style has several features. First, turn intuition into executable rules. Raymond wrote Kairos’s email-reading logic, response tone, prioritization, and boundaries into a knowledge base as Skills and Rules. He calls the process “onboarding a new hire, except the new hire is yourself.” Second, tools have different roles rather than replacing one another. He publicly clarified that “now that there is an AI Agent, you can throw out Notion and n8n” is a misunderstanding: Notion manages collaboration and public documents, n8n runs fixed workflows that require no judgment, and the AI Agent handles tasks that require reasoning. They operate in parallel, and what they share underneath is an understanding of one’s own workflow.4 Third, personalization comes first. He repeatedly emphasizes that with someone else’s Skills and framework, “the key part comes after you copy it: change it.” Adjust the process to your own rhythm, and someone else’s knowledge becomes your own.

By mid-2026, the collaboration had grown from a single machine into a cross-device system: a Mac mini ran automation 24/7, a MacBook could continue the work seamlessly when he was out, and a phone could connect remotely. At the same time, Raymond made a practical correction. In later videos, he acknowledged that desktop versions (Claude Code and Codex desktop) are lower-friction entry points for beginners, while the terminal remains for advanced users who want to extend their capabilities. He also proposed a principle of portability and avoiding lock-in to a single company: keep the memory system local and let Codex and Claude Code read the same set of rules as a single source of truth (SSOT).

Significance of the Transition

This transition matters because it upgraded “using AI” into “running an AI counterpart that can grow,” giving Raymond’s accumulated documents, writing, and workflows a direct purpose for the first time. Without those materials, the Agent would not know how to help him. It also echoes his longstanding Super Individual path: free a company of one from as many tedious administrative tasks as possible, so accumulated knowledge can be called on repeatedly. Raymond’s footnote to this transition captures its direction most clearly: “The biggest by-product of training AI is understanding yourself better.”

Sources

Footnotes

  1. Raymond Hou, “I Just Asked My AI Agent to Name Itself,” public Facebook post, 2026-02-08. View original ↩

  2. Raymond Hou, “My AI Agent,” public Facebook post, 2026-02-13. View original ↩

  3. Raymond Hou, “AI Counterpart Assistant Claude,” public Facebook post, 2026-02-25. View original ↩

  4. Raymond Hou, “Now That We Have AI Agents, Can We Throw Out Notion and n8n?” public Facebook post, 2026-03-23. View original ↩

Graph View

Backlinks