convez.ai
MCP ServerAutomaçãoIntermediário
ahnchen1983

Agentic Substrate

ahnchen1983/agentic-substrate

Arquitetura entre chat com IA e agents que executam trabalho real

Ver no GitHub
InicialPluginScore · de 0 a 100

Framework arquitetural para software nativo de LLM que descreve como LLMs computam, Skills codificam workflows e Agents orquestram tarefas. Inclui Skills instaláveis prontos para uso por profissionais de conhecimento e um modelo de cinco camadas para entender a próxima geração de software.

Instalação

$npx convez-ai ahnchen1983/agentic-substrate

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
1
Acompanhando
0
Saúde
57%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Estruturar workflows de automação com LLMs para conhecimento workers
  • Entender a arquitetura de software nativo de IA em cinco camadas
  • Converter expertise de domínio em workflows orquestrados por agents

Veredito

Framework conceitual sólido que mescla teoria com implementação prática via Skills instaláveis. Diferencial está em traduzir padrões agnósticos de arquitetura de agents em ferramentas concretas, embora ainda careça de maior maturidade em casos de produção em larga escala.

README

A tradução para o português deste README ainda está na fila.

Agentic Substrate

The Foundational Architecture of LLM-Native Software

LLM 原生軟體的底層架構原則

Not just theory - includes installable Skills for knowledge workers. 不只是理論——包含為知識工作者設計的可安裝 Skills。

GitHub Pages Course Start Here Japanese Try Skills


TL;DR - 先講重點

Agentic Substrate is the missing architecture between "chat with AI" and "AI agents that actually run work."

Agentic Substrate 是「和 AI 聊天」到「讓 AI Agent 真正執行工作」之間缺少的架構圖。

It describes the next software paradigm:

Agentic Substrate 是一套描述下一代軟體典範的框架:

LLMs compute. Tools act. Skills encode workflows. Agents orchestrate.

LLM 負責運算。工具負責行動。Skill 編碼工作流程。Agent 負責調度。

This repo gives you:

這個 repo 提供:

  • A five-layer model for understanding LLM-native software
  • A practical learning path from chat to agents
  • Installable Skills you can try immediately
  • A bridge for non-engineers to turn domain expertise into AI workflows

If you have ever thought, "AI is useful, but I keep explaining the same work process again and again," this repo is for you.

如果你曾經覺得「AI 很有用,但我一直在重複解釋同一套工作流程」,這個 repo 就是為你準備的。


Start in 3 Minutes - 三分鐘開始

If you want to...Start here
Understand the big ideaStart Here
Read the Japanese websiteJapanese Version
Read the Japanese docs日本語版 docs
Decode AI coding jargon firstAI Coding Terms
Explore the three-session courseInteractive Course
Open the detailed Day 1 courseDay 1 Interactive Page
Open the detailed Day 2 courseDay 2 Interactive Page
Open the detailed Day 3 courseDay 3 Interactive Page
Try your first SkillQuick Start
See the architectureFive-Layer Model
Learn why Skills matterSkill Anatomy
Help an AI agent understand itselfAgent Education Layer
Help an AI agent read this repollms.txt / AGENTS.md
如果你想要...從這裡開始
理解核心概念Start Here
閱讀日本語網頁版Japanese Version
閱讀日本語 docs日本語版 docs
先拆解 AI Coding 術語AI Coding Terms
查看三堂課互動講義Interactive Course
查看 Day 1 詳細互動頁Day 1 Interactive Page
查看 Day 2 詳細互動頁Day 2 Interactive Page
查看 Day 3 詳細互動頁Day 3 Interactive Page
立刻試第一個 SkillQuick Start
看完整架構Five-Layer Model
理解為什麼 Skill 重要Skill Anatomy
讓 AI agent 理解自己的架構位置Agent Education Layer
讓 AI agent 更容易讀懂本 repollms.txt / AGENTS.md

What You Can Do Here - 你可以在這裡做什麼

1. Understand the shift
   Chatbot -> Tool user -> Skill executor -> Persistent agent

2. Try the pattern
   Install a Skill and watch AI follow a real workflow

3. Build your own
   Turn your professional process into a reusable AI Skill
1. 理解轉變
   聊天機器人 -> 工具使用者 -> Skill 執行者 -> 持續性 Agent

2. 直接體驗
   安裝一個 Skill,觀察 AI 如何遵循真實工作流程

3. 建立自己的 Skill
   把你的專業流程變成可重用的 AI Skill

Concrete example: How a weekly status workflow becomes a Skill

具體範例:一個週報流程如何變成 Skill


What is This? - 這是什麼?

Agentic Substrate is a conceptual framework + practical toolkit that describes the foundational architecture of a new software paradigm - one where LLMs serve as the universal computation layer, Skills replace traditional apps, and Agents orchestrate everything.

Agentic Substrate 是一套概念框架 + 實用工具包,描述一種新軟體典範的底層架構——在這個典範中,LLM 是通用運算層、Skills 取代傳統 App、Agent 負責統籌調度一切。

Unlike frameworks built only for engineers, this is designed for everyone who works with AI - from developers to project managers, marketers, accountants, and administrators.

與只為工程師建構的框架不同,這是為所有使用 AI 工作的人設計的——從開發者到專案經理、行銷人員、會計和行政人員。


Two Reading Paths - 兩條閱讀路徑

Agentic Substrate is designed to be read in two ways:

Agentic Substrate 被設計成可以用兩種方式閱讀:

  • Human-readable path: concepts, examples, courses, and Skills that help people understand the shift from apps to agents.

  • Agent-readable path: operational protocols that help AI agents locate themselves in the stack, decide when to use tools, create Skills, validate results, and preserve reusable workflows.

  • 給人讀的路徑:用概念、案例、課程與 Skills,幫助人理解從 App 到 Agent 的轉變。

  • 給 Agent 讀的路徑:用操作協議,幫助 AI Agent 判斷自己在哪一層、何時使用工具、如何建立 Skill、如何驗證結果,以及什麼流程值得被保存。

Start the agent-readable path here: Agent Education Layer

Agent 可讀路徑從這裡開始:Agent Education Layer


Why Does This Exist? - 為什麼要做這個?

We're living through a paradigm shift:

我們正在經歷一次典範轉移:

Era 時代Logic 邏輯層Interface 介面Unit of Software 軟體單元
Desktop 桌面時代Compiled code 編譯程式碼GUI windows 圖形視窗Application 應用程式
Mobile 行動時代Runtime code 執行環境Touch screens 觸控螢幕App 應用程式
Cloud/SaaS 雲端時代APIs + microservicesBrowser 瀏覽器Service 服務
Agentic 代理時代LLM inference 推理Conversation 對話Skill 技能模組

Traditional software is deterministic - every path is pre-coded. Agentic software is probabilistic - the LLM understands intent and dynamically decides the execution path.

傳統軟體是確定性的——每條路徑都預先寫好。代理型軟體是概率性的——LLM 理解意圖,動態決定執行路徑。

This repo provides the architectural principles behind this shift, a structured learning path, and ready-to-use Skills that let you experience it immediately.

本 repo 提供這個轉變背後的架構原則、結構化的學習路徑,以及即用 Skills 讓你立即體驗。


Repo Structure - 目錄結構

Agentic-Substrate/
|
|-- README.md                          <-- You are here 你在這裡
|-- llms.txt                           [published] AI-readable project map
|-- AGENTS.md                          [published] Guidance for coding agents
|-- agent-education/                   [published] Agent-facing operational protocols
|   |-- README.md                      [published] Agent-readable entry point
|   |-- decision-patterns/             [published] Pattern selection matrix
|   |-- evolution-path/                [published] Agent self-assessment protocol
|   |-- self-model/                    [published] Five-layer self-location protocol
|   |-- skill-system/                  [published] Skill creation protocol
|   `-- tool-use/                      [published] Tool result validation protocol
|-- course.html                        [published] Three-session interactive course
|-- course-day1.html                   [published] Detailed Day 1 interactive course
|-- course-day2.html                   [published] Detailed Day 2 interactive course
|-- course-day3.html                   [published] Detailed Day 3 interactive course
|-- ja.html                            [published] Japanese website entry
|-- ja/                                [published] Japanese presentation pages
|   |-- index.html                     [published] Japanese entry at /ja
|   |-- start-here.html                [published] Japanese Start Here deck
|   |-- glossary.html                  [published] Japanese glossary deck
|   |-- five-layer-model.html          [published] Japanese five-layer model deck
|   `-- learning-path.html             [published] Japanese learning path deck
|-- QUICK-START.md                     [published] 5-min guide to your first Skill
|-- install.sh                         [published] One-command installer 一鍵安裝
|-- ROADMAP.md                         <-- Development phases 發展路線圖
|-- CONTRIBUTING.md                    <-- How to contribute 貢獻指南
|-- CHANGELOG.md                       <-- Version history 版本紀錄
|-- LICENSE                            <-- CC BY-SA 4.0 + MIT
|-- .claude-plugin/                    [published] Claude Code / Cowork plugin config
|
|-- docs/
|   |-- START-HERE.md                  [published] First reading path for humans and agents
|   |-- ja/                            [published] Japanese reading path 日本語版導覽
|   |-- course-planning-notes.md       [published] Planning notes for the three-session course
|   |-- 00-glossary/                   [published] AI coding terms mapped to Agentic Substrate
|   |-- examples/                      [published] Concrete workflow-to-Skill examples
|   |-- 01-foundations/                <-- Core Concepts 核心概念
|   |   |-- agentic-substrate.md       [published] The Core Architecture 核心架構文件
|   |   |-- skill-anatomy.md           [published] Anatomy of a Skill 技能模組解剖學
|   |   |-- markdown-as-medium.md      [published] Markdown as Native Medium 原生媒介
|   |   `-- five-layer-model.md        [published] The Five-Layer Stack 五層架構模型
|   |
|   |-- 02-architecture/               <-- Design Patterns 設計模式
|   |   |-- landscape.md               [published] Ecosystem comparison 生態比較
|   |   |-- agentic-design-patterns.md [published] Routing, Parallelization, Orchestration
|   |   |-- skill-composition.md       [published] The "a235bc1e" composition pattern
|   |   `-- memory-and-state.md        [published] How Agents remember and persist
|   |
|   |-- 03-claude-case-study/          <-- Claude as a Living Example
|   |   `-- claude-architecture.md     [published] Claude Code / Cowork / Agent SDK design
|   |
|   `-- 04-curriculum/                 <-- Learning Path 學習課綱
|       |-- you-already-know.md        [published] "You Already Know" - for non-engineers
|       |-- learning-path.md           [published] Structured curriculum overview
|       |-- level-1-conversation.md    [published] Level 1: Getting value from chat
|       |-- level-2-tools.md           [published] Level 2: From thinking to doing
|       |-- level-3-skills.md          [published] Level 3: Your expertise, reusable
|       `-- level-4-agents.md          [published] Level 4: AI as a colleague
|
|-- skills/                            <-- Installable Skills 可安裝技能
|   |-- skill-self-discovery/          [published] Interactive: discover your existing skills
|   |-- conversation-to-skill/         [published] Turn chat chaos into structured Skills
|   |-- skill-builder/                 [published] Guided Skill creation for any role
|   `-- examples/                      [published] Ready-to-use Skills for knowledge work
|       |-- meeting-notes-to-actions/  [published] Transform: notes -> action items
|       |-- document-reviewer/         [published] Process: multi-pass document review
|       |-- research-and-summarize/    [published] Process: structured research
|       |-- project-status-tracker/    [published] Orchestration: unified status view
|       |-- accounting-reconciler/     [published] Transform: financial reconciliation
|       `-- content-pipeline/          [published] Process: idea -> published content
|
`-- index.html                         [published] Interactive Architecture Visualization

[published] = Ready public entry point


Core Thesis - 核心論點

The future of software is not "AI features added to existing apps." It's a new computational substrate where LLMs are the logic layer, Skills are the functional units, and Agents are the orchestrators.

軟體的未來不是「在現有 App 上加 AI 功能」。 而是一種新的運算基底——LLM 是邏輯層、Skill 是功能單元、Agent 是調度者。


Start Here - 從這裡開始

Just want to try it? -> Quick Start - 5 Minutes to Your First Skill

Interactive visualization -> Explore the Five-Layer Model (deploy to GitHub Pages for the best experience)

Three-session course -> Interactive Course

New to this? Start with -> The Agentic Substrate: Core Architecture

Want to understand Skills? Read -> Anatomy of a Skill

Why Markdown? Read -> Markdown as the Native Medium

Not technical? Start here -> You Already Know How to Build a Skill

How does this compare to other frameworks? See -> Landscape Analysis

What's the development plan? Check -> Roadmap


Origin - 起源

This framework emerged from hands-on experience building AI-native workflows for proposal writing, project management, and team coordination - using Claude Code, Cowork, and custom Skills. It represents the accumulated insight from building 40+ real-world projects where AI agents didn't just assist, but operated as collaborative team members.

這個框架源自於實際動手建構 AI 原生工作流程的經驗——為提案撰寫、專案管理、團隊協作而建,使用 Claude Code、Cowork 和自定義 Skills。它代表了在 40 多個真實專案中累積的洞見,在這些專案中,AI Agent 不只是輔助工具,而是作為協作團隊成員運作。


Authors 作者

  • Ahn Chen - Framework architect, practitioner with 40+ AI-native projects 框架架構師
  • Claude (Opus 4.6) - Co-author, research synthesis, and writing 共同作者、研究整合與撰寫

Built by practitioners, for practitioners. 由實踐者為實踐者而建。


License 授權

Content is licensed under CC BY-SA 4.0. Code is licensed under MIT.

內容以 CC BY-SA 4.0 授權。程式碼以 MIT 授權。

Ver o original no GitHub

Bom para

Relacionados