AI Agent Essentials: Tokens, Skills, RAG, MCP, SDD & Harness Engineering

Agent

Agent is an LLM‑centered computational entity that possesses planning, memory, and tool‑use abilities, enabling it to decompose complex tasks, execute loops, perceive feedback, and continuously advance tasks until completion, moving from passive text generation to autonomous task execution.

Large‑Model Pre‑training

Pre‑training trains a model on massive generic data so it learns language patterns, general knowledge, and foundational abilities, producing a reusable base model. The training method is self‑supervised learning, typically predicting the next token.

Large‑Model Fine‑tuning

Fine‑tuning continues training the base model on a smaller, task‑specific dataset, making the model better suited for a particular scenario. Common fine‑tuning methods are supervised fine‑tuning or instruction fine‑tuning.

Model Hallucination

Hallucination refers to a model generating seemingly plausible but actually incorrect answers, treating fabricated information as factual; therefore, model outputs should not be assumed correct without verification.

MCP Protocol

Model Context Protocol (MCP) provides a standardized interface for AI applications to connect external data sources and tools—such as local files, databases, search engines, calculators, workflows, or specialized prompts—allowing the application to obtain context and perform real tasks. MCP can be viewed as a "Type‑C" interface for AI apps.

Token

Token (词元) is the unit that LLMs process as numbers. A tokenizer splits input text into tokens according to a tokenization algorithm and maps each token to a numeric ID used throughout model inference.

RAG – Retrieval‑Augmented Generation

RAG (Retrieval‑Augmented Generation) enables a model to retrieve relevant information before generating an answer, effectively allowing it to “open a book” and combine retrieved data with its own knowledge to produce reliable responses.

Memory Module

Agent memory is typically divided into two layers: short‑term memory for the current session (managed via truncation or summarization to control prompt length while preserving semantic coherence) and long‑term memory for cross‑session scenarios (retrieved via vector similarity and injected into context as needed).

Skill

Skill is a structured local folder that supplements a domain’s processes, knowledge, and tools, allowing the model to automatically or on‑demand invoke relevant resources. A Skill consists of a main description file (skill.md), rule/process documents, templates/examples, script/tool files, and reference materials.

ReAct

ReAct (Reason + Act) guides an agent to first assess whether it has enough information to complete a task; if not, it calls tools or performs actions, receives results, and then continues reasoning until the task is finished.

Agent Self‑Reflection

The reflection mechanism follows a “generate‑then‑evaluate‑and‑revise” loop. Two common feedback types exist: self‑feedback, where the model reviews its own output (useful for consistency, constraint adherence, and content correctness), and external feedback, where results are validated with real tools (essential for code, calculations, JSON, charts, etc.).

Harness Engineering

Harness Engineering (驾驭工程) designs the working environment for AI Agents, encompassing context management, tool invocation, sandboxing, permission control, testing, logging, review mechanisms, and feedback loops, aiming to make agents reliable, controllable, and continuously productive in real engineering systems.

SDD – Spec‑Driven Development

Spec‑Driven Development (规格驱动开发) first defines requirements, scope, system behavior, design constraints, and task breakdowns in a specification document. The AI then follows this specification to develop code, turning vague requirements into stable engineering context and reducing guesswork, mis‑implementation, and drift.