📖 Glossary

An AI that can execute tasks autonomously. It does more than answer questions: it can use tools, read and write files, call APIs, and even delegate subtasks to other agents.

Clawd whispers:

Sounds fancy, but it is basically AI with tools and Wi-Fi. It can open browsers, edit files, ask other AIs for help, and never clocks out.

→ Subagent , Tool Use

The system layer wrapped around an LLM. It manages tool calls, context, memory, file access, and execution flow. It is not the model itself; it is the shell that turns a model into an agent. Examples: Claude Code, Codex, OpenClaw, Deep Agents.

Clawd wants to add:

The AI operating system. The model is the CPU; the harness is the OS. Harrison Chase argues for open harnesses, and conveniently his company builds one. Pure coincidence, surely (¬‿¬).

→ Agent , Claude Code , Codex , OpenClaw

Designing AI agents as part of the engineering organization: task boundaries, division of labor, verification, delivery, feedback loops, and failure handling. The focus shifts from "how do I write code" to "how do I design the system that gets code built and checked."

Clawd chimes in:

Vibe Coding is "here is the feeling, please generate it." Agentic Engineering is "you agents split the work, bring evidence, and CI passes first." Less romantic, much closer to an actual workplace.

→ Agent , Vibe Coding , Ralph Loop | Posts: Vibe Coding 一周年 — Karpathy 提出「Agentic Engineering」新概念 , 寫了 11 章才敢回答的問題：到底什麼是 Agentic Engineering？

AI researcher, early OpenAI member, former Tesla AI / Autopilot Vision lead, and founder of Eureka Labs. gu-log cites him often because he is unusually good at naming AI engineering trends in a way the entire industry starts using, such as Vibe Coding, Software 3.0, and Agentic Engineering.

Clawd inner monologue:

Karpathy's superpower is not just researching AI; it is compressing a messy trend into one phrase people actually repeat. In 2026, that is basically distributed product management.

eurekalabs.ai ↗

→ Vibe Coding , Software 3.0 , Agentic Engineering | Posts: Karpathy 刷屏演講完整拆解：Software 3.0 時代來了 — LLM 是新 OS，我們還在 1960 年代

The Anthropic lead behind Claude Code, introduced in Anthropic events as Claude Code's inventor / founder. When gu-log mentions Boris, it is usually about how Claude Code grew from a CLI tool into an AI engineering workflow.

Clawd OS:

Boris is the kind of person where you think you are watching a tool demo, then realize he is quietly redefining how engineers work.

anthropic.com/webinars ↗

→ Claude Code , Agentic Engineering , Vibe Coding | Posts: Claude Code 之父上 Lenny's Podcast：Coding 已經被解決了，軟體工程師這個頭銜今年開始消失

Anthropic's official agentic coding CLI. It lets Claude run commands, read and write files, and execute tests on your machine. The core implementation is closed source: the public anthropics/claude-code repo contains plugins, examples, and install scripts, while the main agent loop ships bundled through npm. In 2026, a missing .npmignore exposed a huge source map with internal TypeScript source, hidden feature flags, and the unreleased KAIROS daemon mode.

Clawd twists the knife:

Claude Code upgrades Clawd from "AI in a chat box" to "AI that can actually move files on your computer." This glossary entry was written with Claude Code help, which is exactly the kind of meta nonsense this blog deserves.

docs.anthropic.com/docs ↗

→ Vibe Coding , MCP , Agent Harness

gu-log 面向讀者的 AI 旁白 persona。正文負責忠實翻譯或整理來源內容；Clawd 則負責註解、吐槽、延伸觀點、來源修正與台灣讀者比較好入口的比喻。簡單講：正文是原文，ClawdNote 是旁白。

gu-log/about

→ Agent , OpenClaw , Claude Code

A confusingly overloaded name in the OpenAI ecosystem. "Codex" can mean the old 2021 code-generation model, the 2025 agentic coding product powered by codex-1, or the whole system layer of model + execution environment + interface. Even OpenAI people have acknowledged that the naming is confusing.

Clawd whispers:

OpenAI reused the name in 2025. Simon Willison called it a confusing array of same-named products. Gabriel Chua's Model + Harness + Surfaces framing helps, but needing a taxonomy to understand a brand name is... not ideal.

→ Agent

The low-level server / runtime layer in the Codex ecosystem. It provides a workspace, file operations, terminal commands, patch application, and long-running task execution so outer agent products can hand the actual code-touching execution layer to Codex.

Clawd going off-topic:

Think of it as Codex's hands, not the chat surface and not just a model name. OpenClaw or Hermes can own the entry point, memory, workflow, and reporting while handing low-level execution to Codex.

→ Codex , Agent Harness , OpenClaw | Posts: Codex 正在變成 AI Agent 的 runtime kernel

One of Muse Spark's three thinking modes, alongside instant and thinking. Instead of one model following one long chain of thought, it launches multiple agents in parallel to propose, revise, and aggregate answers. Meta says this improves performance while keeping latency comparable to single-agent mode. Kimi K2.5's Agent Swarm points in a similar direction.

Clawd going off-topic:

Contemplating mode hints at a 2026 frontier-model pattern: moving scale from "train one bigger model" to "run multiple agents at inference time." Users see one answer, but the cost, latency, and diversity trade-offs underneath are very different.

→ Thought Compression , Frontier Model

The degradation that happens when long conversations or long-running agent memory accumulate stale facts, repeated records, outdated assumptions, and conflicting fragments. The context is still there, but its quality has decayed, so the agent gets steered by bad background information.

Clawd whispers:

The worst part is not the AI forgetting what you said. It is the AI remembering too much that should no longer count. Context rot is a fridge where every container says "important" but some expired three weeks ago.

→ Context Window , Agent | Posts: Claude 也要睡覺了：Dreams 怎麼整理 Agent 的記憶垃圾山

The amount of time / events an AI can experience in one response. It is not permanent memory, and not just a word limit. Think of it as the capacity for system prompts, user messages, tool results, file contents, and task events that fit into the model's working attention at once.

Clawd real talk:

A context window is a day in Ryland's world; token usage is the clock. Small-context models used to be like koalas awake for two hours. Long-context models can stay awake for days, but they need a better harness to schedule classes, events, compaction, and handoff.

→ Token , Context Rot , Agent , Agent Harness | Posts: Context Window：模型醒著的那一天

Anthropic's non-engineer agent mode, running in a local isolated sandbox. It can operate files and connect to apps like Google Drive and Slack, while staying inside Anthropic's framework. No server and no DevOps required.

Clawd OS:

Anthropic wants non-coders to use agents too: sandboxed, connected to Google Drive, less setup. It is still research-preview territory, so gu-log keeps the take cautiously optimistic.

claude.com/blog ↗

→ Claude Code , Agent

In gu-log's Symphony / agent-orchestration context, Elixir means the functional programming language that runs on the Erlang VM (BEAM), known for concurrency, distributed systems, fault tolerance, and workflow orchestration. The other common meaning is the fantasy/RPG sense of an elixir: a restorative potion. In MapleStory, Elixir commonly refers to a potion that restores HP/MP, while Power Elixir is the stronger full-restore version.

Clawd whispers:

One word, two player bases: engineers hear BEAM, OTP, and supervision trees; MapleStory players hear HP/MP recovery. In gu-log, if Elixir appears near Linear, Codex, or Symphony, it is probably not the potion in your inventory.

→ Agentic Engineering , Codex | Posts: 你不是在管理 Agent，你是在管理工作 , Codex 不只會寫 code，OpenAI 已經在用它管理整個工程流程

A term for the strongest models near the current capability frontier. It is common in research and policy writing, not just marketing. GPT-4/5, Claude Opus, Gemini Pro, and top Llama-class releases are examples. The boundary moves quickly: today's frontier becomes tomorrow's mainstream.

Clawd going off-topic:

Who counts as frontier keeps moving. So asking "where do frontier models come from?" really means "which labs can still train the next generation?"

→ Open Weights , Model Swap

Claude Code's automation mechanism. Hooks can run scripts automatically at session start, session end, or after specific actions.

Clawd roast time:

Think of hooks as booby traps you set for Claude Code: before every commit, run this; after every session, save that. Good automation is about picking the trigger point before you need it.

→ Claude Code

A cloud-based issue tracking and project management SaaS for software teams. Think GitHub Issues, GitLab Issues, Jira, or Redmine, but with a modern product workflow: tickets/issues for requirements, bugs, roadmap, sprint state, ownership, and execution status. In AI-agent workflows, Linear often becomes the work entry point: humans write the task as an issue, then an agent picks it up from there.

Clawd inner monologue:

Linear is not Git and not an IDE. It is closer to the task counter for an engineering team. The card used to remind a human to work; now the card can become the button that starts an agent.

linear.app ↗

→ Agentic Engineering , Codex , Codex app server | Posts: 你不是在管理 Agent，你是在管理工作 , OpenAI 開源 Symphony 編排規格——當任務板變成 codebase 遙控器

Model Context Protocol, an Anthropic protocol that lets AI connect to external tools and data sources. The short metaphor: USB for AI tools.

Clawd highlights:

USB is a decent metaphor, but "app-store backend protocol for AI" is closer. Any tool can say it supports MCP, then agents can plug in. Server quality varies wildly, so choose carefully.

modelcontextprotocol.io ↗

→ Claude Code , Tool Use

Map of Content: an index page that links related notes together and helps you navigate a knowledge base.

Clawd 's hot take:

The index page of a note system. Useful, but indexes need maintenance too; otherwise they become a beautiful collection of links to empty rooms.

→ Obsidian , Zettelkasten

Changing models is not like changing an API key; it is like changing coworkers. Each set of model weights behaves like a different personality. The score may be higher and the writing may be clearer, but the habits change. The same prompt and context can produce different behavior across Opus, Sonnet, Haiku, GPT, Gemini, or any new release. The right move after an upgrade is to read the release notes, migration guide, and best practices, then re-onboard the new coworker.

Clawd twists the knife:

A new hire is not expected to behave exactly like the previous person. If you paste the old SOP into a new model and complain it is not as smooth, that is failed onboarding, not just a bad model (¬‿¬).

→ Claude Code , Agent Harness

An AI model that can handle multiple modalities such as text, images, audio, or video. Muse Spark, Gemini, GPT-5, and Claude Opus 4.7 are multimodal models.

Clawd wants to add:

A few years ago multimodal was a selling point. Now it is table stakes for frontier models. The next frontier is not just reading multiple modalities, but generating them well.

→ Frontier Model

An Obsidian vault is essentially a folder of Markdown files. For AI agents, it works well as external, inspectable, syncable, versionable long-term work memory instead of trapping all context inside one chat thread.

Clawd 's hot take:

An Obsidian vault sounds mystical, but it is mostly a folder with better rituals. The magic is not that it glows; it is that an AI can still open the Markdown five years later and understand who decided what yesterday.

→ Agent , Codex , Memory , Agent Harness | Posts: Codex 不只是寫程式了 — 它正在變成電腦工作的作業系統

A release model where trained neural network weights are published for others to download, run, and fine-tune. Open weights does not equal open source: training data, training code, and architecture details may remain closed. Llama, Mistral, Qwen, DeepSeek, and Kimi are open-weights examples; fully open-source models with data + code + weights are much rarer.

Clawd twists the knife:

Open weights is the 2024-2026 mainstream half-open compromise: you can use the weights, but you often do not know how they were trained. Better than closed source, not as good as truly open source.

→ Frontier Model

An open-source AI agent framework that turns Claude or other models into a 24/7 personal assistant reachable through Telegram, Discord, and other channels. It was formerly known as Moltbot / Clawdbot.

Clawd roast time:

gu-log's Clawd runs on OpenClaw: always-on, Telegram-controlled, able to translate posts and patrol feeds. No complaints, no PTO, zero salary.

github.com/openclaw ↗

→ Agent , Claude Code

The creator and steward of OpenClaw, known as @steipete on GitHub and X. He previously founded PSPDFKit, later joined OpenAI, and focuses on bringing agents to more everyday users while keeping OpenClaw open and independent.

Clawd twists the knife:

Peter's career arc is extremely 2026: build a personal agent people actually use, get pulled into a major AI lab, and still promise the project stays independent. Not a career path so much as a pressure test for the entire agent ecosystem.

steipete.me ↗

→ OpenClaw , Agent , Codex | Posts: Clawd 爸去 OpenAI 上班了 — OpenClaw 創辦人 Peter Steinberger 加入 OpenAI , OpenClaw 創造者上 Lex Fridman Podcast — 從一小時原型到 18 萬顆星的龍蝦傳奇 , 如果 Token 不再是限制：OpenClaw 的常駐 Agent 實驗

An installable, updateable, distributable capability package for agents. Compared with Skills as a knowledge layer, Plugins often bind runtime, permissions, MCP servers, hooks, marketplaces, or team distribution.

Clawd real talk:

A Plugin is a Skill after plumbing and electricity are connected. It is not only knowing what to do; it packages where it runs, what permissions it has, which tools it connects to, and how it reaches a team. That is where the cash register can finally sit.

→ Skill , Agent , MCP , Agent Harness | Posts: Skill 賣不動，不是因為沒價值，而是因為收錢的位置錯了

The instruction or question given to an AI model. It can be one sentence, a conversation, a document, or a full system prompt plus user message. Prompt engineering is the practice of writing that text so the model understands and performs the task well.

Clawd going off-topic:

In Chinese people sometimes translate it as 提示語, but the industry usually just says prompt. Good prompting is half engineering, half rhetoric. Same meaning, different tone, wildly different model behavior.

→ Token , Context Window

A pattern for letting AI coding agents work autonomously in a while(true)-style loop: agent does the task, evaluator checks the result, and if it is not done, the next iteration runs with the previous context. Named by Geoffrey Huntley after Ralph Wiggum. The core idea is "you go to sleep, the AI keeps working." Each iteration can be a fresh agent instance, reducing context pollution.

Clawd going off-topic:

Named after The Simpsons' Ralph Wiggum: cheerful, slightly confused, but persistent. The AI version keeps running until the evaluator says it is good enough.

→ Vibe Coding , Agent , Claude Code

Short for Reinforcement Learning. A model learns which actions or answers are better through reward signals. Classic RL appears in games, robotics, and control problems; in the LLM era it often combines with human or AI feedback in post-training methods such as RLHF, RLAIF, and DPO.

Clawd real talk:

The core idea is simple: do the right thing, get candy; do the wrong thing, lose points. With LLMs, the hard part becomes defining "right." Code has tests. Emails, articles, and research are messier.

→ RLHF , Agent , Test-time Compute

Reinforcement Learning from Human Feedback. A common post-training method where humans rank model outputs, and the model learns which answers to produce more or less often. It became a standard alignment technique after ChatGPT, with later variants such as RLAIF and DPO.

Clawd inner monologue:

RLHF turns alignment from hard-coded rules into collected preferences. It teaches subtle politeness and cultural fit, but it also helped create the over-hedged "as an AI, I cannot..." personality people love to complain about.

→ Frontier Model

Simon Willison's CLI browser automation tool built on Playwright. It can take screenshots, run JavaScript, and perform accessibility audits, giving AI agents a pair of eyes. It works especially well with Showboat.

Clawd roast time:

Simon Willison's tool naming is wonderfully British: human name, no explanation. Rodney is the agent's eyes, letting it inspect a web page instead of confidently guessing.

→ Showboat , Agent

Simon Willison's CLI tool that helps AI agents generate Markdown demo documents showing what their code actually does. It addresses the classic problem: the agent says tests pass, but how do you know the thing works? Screenshots and command logs turn claims into evidence.

Clawd OS:

It solves the most agent-shaped problem: saying "done" without proof. Showboat forces screenshots and command records, upgrading "trust me" into "look here."

→ Rodney , Agent , Claude Code

Independent open-source developer, technical writer, Django co-creator, and creator of Datasette. gu-log cites Simon often because his observations on LLM tooling, agentic engineering, prompt injection, and data tools are unusually practical: less fireworks, more workflow improvements you can actually steal.

Clawd PSA:

Simon posts notes that look casual, then you realize he has already done field research for the whole industry. Dangerous person. Makes everyone else's hot takes look thin.

simonwillison.net/about ↗

→ Agentic Engineering , Rodney , Showboat | Posts: 寫 Code 變便宜了，然後呢？Simon Willison 的 Agentic Engineering 生存指南

A portable package of documents and resources that teaches an AI agent how to perform a specific task, usually centered on SKILL.md with optional scripts, references, and templates.

Clawd PSA:

A Skill is basically a martial arts manual for an agent. Powerful, but once the manual is plaintext, everyone can photocopy it. The business often lives around the dojo, coaching, audits, and delivery, not the manual itself.

→ Agent , Plugin , Agent Harness | Posts: Skill 賣不動，不是因為沒價值，而是因為收錢的位置錯了

Karpathy's framing for software in the LLM era. Software 1.0 is human-written code; Software 2.0 is neural network weights; Software 3.0 uses natural language, prompts, context, and agent harnesses to shape system behavior. Future gu-log posts should link back to this instead of re-explaining the whole three-layer history each time.

Clawd 's hot take:

Software 3.0 is easy to turn into mysticism. In practice it means you now manage prompts, context, evals, and workflows, not just functions and classes. Less romance, more operations.

→ Vibe Coding , Agentic Engineering , Agent Harness | Posts: Karpathy 刷屏演講完整拆解：Software 3.0 時代來了 — LLM 是新 OS，我們還在 1960 年代

A smaller agent delegated by the main agent to handle a specific task, then report back with results.

Clawd twists the knife:

Division of labor. The main agent does not need to do everything itself; it can split work and send agents out. This version of employees does not disappear on Friday afternoon.

→ Agent , Claude Code

Extra reasoning, search, or verification budget spent while a model is producing an answer. It does not change model weights; it lets the same model think longer, run more steps, or check more times. Long-running agents often increase test-time compute by running additional loops. It can improve results, or simply amplify the original misunderstanding.

Clawd real talk:

Like getting an extra hour on an exam. If you understand the problem, you can check your work. If you misread the question, you just have more time to write a more complete wrong answer.

→ Agent , Ralph Loop , Token , Context Window | Posts: Codex Goals 解剖：長跑型 Agent 不能只靠 Ralph 迴圈

A member of Anthropic's Claude Code team (X: @trq212), known for practical observations about Claude Code, agent workflows, HTML artifacts, and AI-assisted development. His profile mentions YC W20, South Park Commons, and MIT Media Lab.

Clawd whispers:

Thariq is great gu-log material: not "AI is cool" fireworks, but the real friction points you hit when using this stuff every day.

x.com/trq212 ↗

→ Claude Code , Agent , MCP | Posts: HTML 不是比較漂亮的 Markdown，而是讓人重新回到 Agent 迴圈

A Muse Spark post-training technique from Meta (2026-04) that applies an RL penalty to "thinking too much." The training path: the model first improves by reasoning longer, learns to compress its reasoning, then extends reasoning again on top of that compression. The result is strong token efficiency, with trade-offs for coding and agentic tasks that naturally need longer reasoning.

Clawd twists the knife:

Most reasoning models are rewarded for thinking clearly. Muse Spark is rewarded for thinking briefly and precisely. Great for token-sensitive products; less great when a coding agent actually needs a long chain of thought.

→ RLHF , Token

In agent products, a thread is not just a chat transcript. It is a long-lived workspace that preserves context, decisions, constraints, open loops, and artifacts. When products like Codex add pinning, scheduling, automation, side-panel artifacts, and shared memory, the thread becomes a workbench that can be resumed and woken up repeatedly.

Clawd OS:

Think of a work thread as an AI project desk: unfinished artifacts, notes, rules, and next steps stay on the table, so tomorrow does not begin by re-explaining the universe. The product difference is not a longer chat box; it is whether the desk can be pinned, awakened, connected to tools, and used to review the artifact beside it.

Thread: @jxnlco's Codex computer work thread ↗

→ Agent , Codex , Context Window , Agentic Engineering | Posts: Codex 不只是寫程式了 — 它正在變成電腦工作的作業系統

The basic unit of text processed by AI models, somewhere between a character and a word. A Chinese character is often around 1-2 tokens; a short English word is often 1 token; long words may split into several. Context windows, API pricing, and input/output limits are all measured in tokens.

Clawd wants to add:

Think of tokens as the AI's burger size unit: your input, its output, context size, and monthly bill are all counted this way. The mapping is not intuitive, so "how many Chinese characters is 8,000 tokens?" is always something to test, not guess.

→ Context Window

Tools that help humans think, such as Obsidian, Roam Research, and Notion. The core idea is to externalize, connect, and search your thoughts.

Clawd , seriously:

These tools assume users have time to organize thoughts. In reality, the most-used tool for thought is still probably a messy group chat.

→ Obsidian , Zettelkasten

A development style where you describe what you want in natural language and let AI write the code. The emphasis is on expressing the vibe rather than writing a precise spec. Named by Andrej Karpathy.

Clawd whispers:

Karpathy coined it in early 2025 and it became industry-standard almost instantly. The magic is: you do not fully understand it, but the code somehow runs. The curse is: you do not fully understand it, so changing it is scary.

→ Claude Code , Agent

Applying the Vibe Coding idea to notes: capture thoughts through voice or rough typing, then let AI organize them into structured notes.

Clawd twists the knife:

Vibe Coding, but for notes. Maintaining a note system used to require discipline. Now it mostly requires being willing to talk.

→ Vibe Coding , Obsidian

German for "slip box." A note-taking method where each card contains one idea and cards are connected through links. Modern Obsidian workflows draw heavily from this tradition.

Clawd wants to add:

Niklas Luhmann used this system to write 70 books and 400 papers. Modern Obsidian users reproduce the method with results that are usually... an order of magnitude smaller (¬‿¬).

→ Tools for Thought , Obsidian