📖 AI Agent Glossary (English)¶

80+ essential terms explained in plain English.

1. Core Concepts¶

Term	Plain-English Explanation
Agent	An AI that can think AND act on its own — like a "digital employee"
LLM (Large Language Model)	The "brain" of an Agent — understands and generates text. GPT, Claude, DeepSeek
Tool	An Agent's "hands" — lets it check weather, search web, run calculations
Memory	An Agent's "notebook" — remembers what you said. Short-term & long-term
Prompt	The instruction you give to an AI — "what to do"
Prompt Engineering	The skill of writing better prompts to get better results
Context	Everything the Agent can currently "see" — like the conversation history
Token	The basic unit of text AI counts/bills by. ~1 Chinese char ≈ 1-2 tokens
Inference	The process of the AI "thinking" and producing a result
Context Window	How much text the model can "see" at once (e.g., 128K tokens ≈ 100k words)

Term	Plain-English Explanation
ReAct	The "think → act → observe → think again" loop (Reasoning + Acting)
Tool Calling	The model says "I want to use tool X" and actually calls it
Function Calling	Another name for tool calling — the model outputs which function + args
Chain-of-Thought (CoT)	Making AI "think step by step" for more accurate answers
RAG (Retrieval-Augmented Generation)	Search a knowledge base first, then answer — makes AI answer from YOUR documents
Fine-tuning	Extra training on your data to make the model an "expert" in your domain
Multi-Agent	Multiple agents working together — one researches, one analyzes, one writes
Orchestration	Directing multiple agents: who does what, in what order
Workflow	A fixed sequence of steps — like an assembly line
Human-in-the-Loop (HITL)	A human confirms at critical steps — for safety-critical scenarios

Term	Plain-English Explanation
LangGraph	Framework that uses "graphs" to orchestrate complex agent flows
CrewAI	Multi-agent framework where agents play different "roles" like a team
AutoGen	Microsoft's framework — agents solve problems by "chatting" with each other
MAF (Microsoft Agent Framework)	AutoGen's next generation (2025)
LlamaIndex	The go-to framework for RAG / knowledge-base Q&A
Dify	Drag-and-drop platform — build AI apps visually, no code needed
Ollama	Run LLMs on YOUR computer — free, private, offline
MCP (Model Context Protocol)	Standard protocol for connecting tools — write once, use everywhere
SDK	Software Development Kit — official ready-made libraries
API	A remote service interface — the paid channel to use big models
LangSmith	A "dashboard" to debug and monitor agents — see cost, latency, traces

Term	Plain-English Explanation
Embedding	Converting text into "number coordinates" so computers can measure similarity
Vector Database	A database specialized for storing embeddings (ChromaDB, FAISS, Pinecone)
Vector Store	Same as vector database — a warehouse for vectors
Semantic Search	Search by MEANING, not just keywords. "It's cold today" matches "temperature dropped"
Chunking	Splitting long documents into smaller pieces before storing
Index	A pre-built "table of contents" for fast lookup
Checkpoint	Saved agent state mid-run — recover after a crash
Persistent Memory	Memory that survives restarts (stored in DB / vector store)
Short-term Memory	Remembered within a conversation, forgotten after

Term	Plain-English Explanation
Model	The trained "brain" itself (GPT-4o, Claude 3.5, etc.)
Multimodal	Can process text + images + audio + video together
Token Limit	Maximum tokens the model handles per request
Latency	Time from request to response — lower is better
Throughput	How many requests processed per unit time
Deployment	Putting your code/model on a server to serve users
On-premise	Running on YOUR own hardware — data never leaves

Term	Plain-English Explanation
Hallucination	The AI confidently making things up
Accuracy	Percentage of correct answers
Precision	Of the things it said "yes" to, how many were really correct
Recall	Of all the things it SHOULD have found, how many it did find
Evaluation (Eval)	Testing/scoring the AI's answers with a test set
Observability	Being able to see what's happening inside the system
Tracing	Recording the full path of a single request — find where it's slow
Logging	Recording what happened during runtime
Cost Optimization	Spending less money — smaller models, caching
Guardrails	Filters preventing harmful/inappropriate output
Streaming	Output appearing word-by-word — the "typing effect"

Abbrev	Full Form	Meaning
AI	Artificial Intelligence	人工智能
LLM	Large Language Model	大语言模型
NLP	Natural Language Processing	自然语言处理
RAG	Retrieval-Augmented Generation	检索增强生成
MCP	Model Context Protocol	模型上下文协议
API	Application Programming Interface	应用程序接口
SDK	Software Development Kit	软件开发工具包
CoT	Chain-of-Thought	思维链
HITL	Human-in-the-Loop	人在回路
ML	Machine Learning	机器学习
AGI	Artificial General Intelligence	通用人工智能
JSON	JavaScript Object Notation	轻量数据格式

Agent = thinking + acting + remembering

Tool = a skill plugin you give to the Agent

Prompt = your remote control for the Agent

Come back to this table whenever you meet an unfamiliar term!