From LLM to Agent Skill: Understanding AI's Underlying Logic

The Concept Chain

AI has a logical structure that builds from simple to complex. Understanding this chain — LLM → Token → Context → Prompt → Tool → MCP → Agent → Agent Skill — reveals how modern AI systems actually work.

A Bilibili creator (Mark’s Tech Workshop) published a 32-minute video explaining these concepts from a bottom-up engineering perspective. It has since garnered 1.13 million views and 54,000 likes, making it one of the most-watched AI tutorials in Chinese.

1. LLM: The Foundation

What is LLM?

Large Language Model — a Transformer-based architecture (2017, Google) that works as a “text continuation game.” The core principle: predict the next most probable word based on probability, outputting one token at a time.

Key Insight: “LLMs are essentially a text continuation game” — they predict the next most probable word in a continuous loop until completion.

2. Token: The Basic Unit

A token is the smallest text unit processed by LLM. Importantly, tokens are not equal to words.

Tokenization Process:

• Encoding: Text → Tokens → Token IDs
• Decoding: Token IDs → Tokens → Text
• Algorithms: BPE, WordPiece, SentencePiece

Cost factor: Input tokens are typically more expensive than output tokens.

3. Context & Context Window

Context is the conversation history and memory that maintains continuity. The Context Window is the maximum tokens a model can process.

Warning: Even with large context windows, models suffer from “lost in the middle” phenomenon — they pay less attention to information in the middle of long contexts.

4. Prompt Engineering

Advanced Technique: Chain-of-Thought

Chain-of-thought prompting significantly improves model performance by breaking complex problems into step-by-step reasoning. Particularly effective for mathematical, logical, and planning tasks.

5. Tools & Function Calling

Tools are external capabilities that extend LLM beyond text generation — APIs, databases, computation engines.

Function Calling Process:

1. User query received
2. Context assembly (tools + messages)
3. LLM decides if tool needed
4. Tool execution by developer code
5. Results returned to LLM
6. Final response generated

Tool Definition Structure:

6. MCP: Model Context Protocol

MCP is an open standard for connecting AI applications to external systems — think of it as “USB-C for AI applications.” Created by Anthropic, open-sourced November 2024.

MCP Architecture:

• Protocol: JSON-RPC 2.0 based
• Components: Host, Client, Server
• Host: Container/coordinator
• Client: 1:1 server connections
• Server: Provides capabilities

MCP Capabilities:

• Resources: Data access (files, databases)
• Tools: Functions to execute
• Prompts: Reusable templates
• Sampling: LLM inference requests
• Roots: Filesystem boundaries

MCP Adoption (2026):

• SDK Downloads: 97M+/month (from 100K in 1 year)
• Public Servers: 13,230+
• Company Servers: Grew 232% in 6 months
• Supported Tools: Claude, ChatGPT, VS Code, Cursor, 30+ more

7. AI Agent: Autonomous Systems

AI Agent = LLM + Planning + Memory + Tools

AI Agents are not just LLMs — they are complete systems that use LLMs as reasoning engines with supporting components to reason, plan, and execute autonomously.

Agent Components:

• LLM Brain: Cognitive core for reasoning
• Planning Logic: Problem decomposition
• Memory: Short-term (context) + Long-term (vector DB)
• Tools: External capabilities
• Observation: Environment perception

Agent Loop (See-Think-Act-Learn):

1. See: Observe environment (emails, sensors, web)
2. Think: Decide actions using LLM reasoning
3. Act: Execute actions (send emails, update CRMs)
4. Learn: Log outcomes for improvement
5. Repeat: Continuous operation cycle

Agent Types:

8. Agent Skill: Extensible Capabilities

Agent Skills are lightweight, open format for extending agent capabilities. Created by Anthropic, now an open standard supported by 30+ AI tools.

Skill Structure:

• SKILL.md: Required metadata + instructions
• scripts/: Optional executable code
• references/: Optional documentation
• assets/: Optional templates, resources

Progressive Disclosure Pattern:

• Discovery (~100 tokens): Load name and description at startup
• Activation (< 5000 tokens): Load full SKILL.md when task matches
• Execution (as needed): Load additional resources

Security Warning: Security researchers have found 341 malicious agent skills. Always verify skill sources before installation.

Real-World Applications

Framework Comparison

Key Takeaways

1. AI Agents ≠ LLMs — Agents are complete systems with planning, memory, and tools
2. MCP is the Standard — Open protocol connecting AI to external systems
3. Tools are Essential — Function calling transforms LLMs into interactive agents
4. Skills Enable Extensibility — Lightweight format for adding capabilities
5. Progressive Disclosure — Load skills on-demand to optimize token usage
6. Security Matters — Verify skill sources to prevent malicious code execution
7. Enterprise Adoption Growing — 57% of organizations have agents in production

What AI concept surprised you the most when you first understood it?

從 LLM 到 Agent Skill：理解 AI 底層邏輯

概念鏈

AI 有一個從簡單到複雜的邏輯結構。理解這條鏈——LLM → Token → Context → Prompt → Tool → MCP → Agent → Agent Skill——揭示了現代 AI 系統實際如何運作。

一位B站創作者（馬克的技術工作坊）發布了一支32分鐘影片，從底層工程角度解釋這些概念。至今已獲得113萬次觀看和5.4萬個讚，成為中文最受歡迎的AI教程之一。

1. LLM：基礎

什麼是 LLM？

大語言模型——基於 Transformer 架構（2017年，Google），作為「文字接龍遊戲」運作。核心原理：根據機率預測下一個最可能的詞，一次輸出一個 Token。

關鍵洞察： 「大模型本質上就是一個文字接龍遊戲」——它們在持續循環中預測下一個最可能的詞，直到完成。

2. Token：基本單位

Token 是 LLM 處理的最小文字單位。重要的是，Token 不等於詞。

標記化過程：

• 編碼： 文字 → Token → Token ID
• 解碼： Token ID → Token → 文字
• 演算法： BPE、WordPiece、SentencePiece

成本因素： 輸入 Token 通常比輸出 Token 昂貴。

3. Context 與 Context Window

Context 是維持對話連續性的對話歷史和記憶。Context Window 是模型可處理的最大 Token 數。

警告： 即使有大型 Context Window，模型仍面臨「中間遺失」現象——它們對長上下文中間部分的關注較少。

4. Prompt 工程

進階技術：思維鏈

思維鏈提示通過將複雜問題分解為逐步推理，顯著提升模型效能。在數學、邏輯和規劃任務中特別有效。

5. Tool 與 Function Calling

Tool 是擴展 LLM 超越文字生成的外部能力——API、資料庫、計算引擎。

Function Calling 過程：

1. 接收用戶查詢
2. 上下文組裝（Tool + 訊息）
3. LLM 決定是否需要 Tool
4. 開發者代碼執行 Tool
5. 結果返回給 LLM
6. 生成最終回應

6. MCP：模型上下文協議

MCP 是連接 AI 應用程式與外部系統的開放標準——可以將其視為「AI 應用程式的 USB-C」。由 Anthropic 創建，2024年11月開源。

MCP 架構：

• 協議： 基於 JSON-RPC 2.0
• 組件： Host、Client、Server
• Host： 容器/協調器
• Client： 1:1 伺服器連接
• Server： 提供能力

MCP 採用（2026年）：

• SDK 下載量：每月 9700 萬+（1年內從 10 萬增長）
• 公開伺服器：13,230+
• 公司伺服器：6個月內增長 232%
• 支援工具：Claude、ChatGPT、VS Code、Cursor 等 30+ 個

7. AI Agent：自主系統

AI Agent = LLM + 規劃 + 記憶 + Tool

AI Agent 不只是 LLM——它們是完整的系統，使用 LLM 作為推理引擎，並配備支援組件以進行推理、規劃和自主執行。

Agent 組件：

• LLM 大腦：推理的認知核心
• 規劃邏輯：問題分解
• 記憶：短期（上下文）+ 長期（向量資料庫）
• Tool：外部能力
• 觀察：環境感知

Agent 循環（觀察-思考-行動-學習）：

1. 觀察： 感知環境（郵件、感測器、網路）
2. 思考： 使用 LLM 推理決定行動
3. 行動： 執行行動（發送郵件、更新 CRM）
4. 學習： 記錄結果以改進
5. 重複： 持續運作循環

8. Agent Skill：可擴展能力

Agent Skill 是擴展 Agent 能力的輕量級開放格式。由 Anthropic 創建，現為 30+ 個 AI 工具支援的開放標準。

Skill 結構：

• SKILL.md：必需的中繼資料 + 指令
• scripts/：可選的可執行代碼
• references/：可選的文件
• assets/：可選的範本、資源

漸進式披露模式：

• 發現（~100 Token）：啟動時載入名稱和描述
• 啟動（< 5000 Token）：任務匹配時載入完整 SKILL.md
• 執行（按需）：載入額外資源

安全警告： 安全研究人員已發現 341 個惡意 Agent Skill。安裝前務必驗證 Skill 來源。

實際應用

框架比較

關鍵要點

1. AI Agent ≠ LLM — Agent 是包含規劃、記憶和 Tool 的完整系統
2. MCP 是標準 — 連接 AI 與外部系統的開放協議
3. Tool 至關重要 — Function Calling 將 LLM 轉化為互動式 Agent
4. Skill 實現可擴展性 — 添加能力的輕量級格式
5. 漸進式披露 — 按需載入 Skill 以優化 Token 使用
6. 安全很重要 — 驗證 Skill 來源以防止惡意代碼執行
7. 企業採用增長 — 57% 的組織已有 Agent 投入生產

哪個 AI 概念在你第一次理解時最讓你驚訝？

References:

• “Attention Is All You Need” — Vaswani et al. (2017)
• Anthropic MCP Specification (2024)
• Bilibili Video: Mark’s Tech Workshop — “從 LLM 到 Agent Skill” (2026)
• LangChain, CrewAI, AutoGen, LlamaIndex Documentation
• OpenAI Function Calling API Reference