AI Agent Terminology

With the rapid development of large language model (LLM) and agent (Agent) technologies, programming paradigms are undergoing a profound transformation. Vibe Coding, Agentic Coding, Harness Engineer, Loop Engineer... a batch of new concepts and new terms are emerging one after another. Next, let's see what these new terms are and what each of them means.

AI AgentIt consists of Artificial Intelligence and Agent.

1. Coding Paradigm

Term	Chinese	Core Meaning	Example (Restaurant Business)
Vibe Coding	Vibe Coding	Use natural language/voice to describe requirements and let AI generate code.	Frying egg fried rice at home by feel, adding whatever you want.
Context Engineering	Context Engineering	Improve model performance by organizing context, knowledge, memory, and tools.	Prepare ingredients in advance, prepare the menu and kitchen environment.
Agentic Coding	Agentic Coding	Agent autonomously completes programming tasks (design → implementation → testing → acceptance).	Run a formal restaurant, from menu, ingredient preparation to the entire process of serving dishes.
AI Native Development	AI Native Development	By default, AI participates in the entire process of design, development, and testing.	Directly run a smart restaurant.

2. Engineer Role

Term	Chinese	Core Meaning	Example (Restaurant Operations)
Harness Engineer	Harness Engineer	Calls various Harness components, does the work, and accepts the results themselves	Professional chef, proficient in various cooking techniques, tastes and quality-checks after cooking
Loop Engineer	Loop Engineer	Builds automated orchestration systems, self-evolving	Restaurant operations management system, coordinating scheduling, procurement, serving pace, and costs
Context Engineer	Context Engineer	Responsible for designing contexts, knowledge sources, and memory systems	Head Kitchen Dispatcher
AI Product Engineer	AI Product Engineer	Responsible for designing AI capabilities and business closed loops	Restaurant Owner and Operator
Agent Operator	Agent Operations Engineer	Continuously monitors and optimizes Agent execution effectiveness	The store manager continuously optimizes operational data

3. Model & Fundamentals

Terminology	Chinese	Core Meaning	Notes
LLM（Large Language Model）	Large Language Model	A language prediction model trained on massive text data	GPT, Claude, and Gemini are all LLMs
GPT	Generative Pre-trained Transformer	OpenAI's model architecture paradigm	Generative Pre-trained Transformer
Token	Token	The smallest unit of text processed by the model (about 3/4 of an English word)	Unit of measurement for billing and context length
Context Window	Context Window	The maximum number of tokens a model can 'see' at once	The larger the window, the more it remembers
Inference	Inference	The process by which a model generates output	Distinct from training
Hallucination	Hallucination	The model confidently fabricates nonexistent information	Generates content inconsistent with facts, context, or objectives
Temperature	Temperature	Controls the sampling probability distribution; higher values are more diverse, lower values are more stable	It is recommended to lower it when writing code
Top-p / Top-k	Sampling Parameters	Controls the sampling range of candidate words	Affects output diversity
Embedding	Vector Embedding	Converts text/images into numerical vectors for similarity calculations	The foundation of RAG
Fine-tuning	Fine-tuning	Continues training a general-purpose model on specific data	Makes the model more knowledgeable in a specific domain
RLHF	Reinforcement Learning from Human Feedback	Aligns model behavior with human preferences	Makes AI 'obedient'
MoE（Mixture of Experts）	Mixture of Experts	Only activates a subset of parameters each time, improving efficiency	Run stronger models with less compute
Multimodal	Multimodal	Process text, images, audio, and video simultaneously	GPT-4o and Gemini are both examples
Transformer	Transformer architecture	The underlying neural network structure of modern LLMs	The attention mechanism is the core
KV Cache	Key-value cache (KV Cache)	Cache context computation results to improve inference speed	Avoid flipping through the recipe repeatedly
Latency	Latency	Model response time	Affects user experience
Throughput	Throughput	Number of tasks processable per unit time	Affects concurrency capability

4. Prompt Engineering

Term	Chinese	Core meaning
Prompt	Prompt	Input instruction given to the model
System Prompt	System prompt	The underlying instruction that defines the Agent's role, boundaries, and behavior
Prompt Engineering	Prompt engineering	The technique of designing and optimizing prompts to achieve better output
Zero-shot	Zero-shot	Give no examples and directly let the model complete the task
Few-shot	Few-shot	Give a few examples to guide the model to output in a specific format/style
Chain of Thought（CoT）	Chain of Thought	Guide the model to produce intermediate reasoning steps, enhancing complex reasoning ability
ReAct	Reasoning + Acting	Reasoning + Acting alternate, thinking while doing, and it is the core paradigm of Agent
Role Prompting	Role-playing	Let the model play a certain role (e.g., "You are a senior architect")
Structured Output	Structured output	Force the model to output in formats such as JSON/XML
Prompt Chaining	Prompt chaining	Multiple prompts chained together to complete complex tasks
Self-Consistency	Self-consistency	Generate multiple reasoning results and then vote to select
Tree of Thoughts（ToT）	Tree of Thought	Explore multiple reasoning paths simultaneously

5. Agent Architecture

Term	Chinese	Core meaning	Example (restaurant operation)
Agent	Agent	An AI program that can autonomously perceive, decide, and execute tasks	Head chef
Multi-Agent	Multi-agent	Multiple Agents divide work and collaborate	The chef team collaborates to serve meals
Subagent	Sub-agent	A dedicated sub-agent derived from the main Agent	Full-time chef (e.g., cutting/preparing, cold dishes, desserts)
Tool Use / Function Calling	Tool Calling / Function Calling	Agent calls external tools/APIs to execute actions	Using kitchen tools such as cleavers, stoves, and ovens
Planning	planning	Agent breaks down goals and formulates execution steps	Making a meal prep and plating plan
Reflection	reflection	Agent reviews its own behavior and improves	Adjust flavors based on feedback after tasting the dishes
Memory	Memory	Long-term information saved across sessions for reuse (short-term/long-term)	Accumulation of recipe books and operational data
Agent Loop	Agent loop	Iterative mechanism of thinking → acting → observing → rethinking	The cycle of ingredient prep → cooking → tasting → serving
Environment	Environment	The real world that Agent perceives and executes actions in	kitchen environment
Observation	observation	Obtain feedback information after execution	tasting feedback
Execution	Execute	Turning plans into real actions	start cooking
Long-term Memory	Long-term memory	Long-term preservation of experiential knowledge	Business database

6. Harness Components (AI Agent Capability Modules)

Term	Chinese	core meaning	Example (Restaurant Management)
Harness	Wiring Harness / Control Framework	Runtime framework that supports Agent execution, context management, and tool orchestration	The entire kitchen system (the kitchen itself)
Skills	skill	Specialized capability modules that Agent can call	Various cooking techniques (pan-frying, stir-frying, deep-frying)
Context	Context	The scope of information available to the current task	Current orders, ingredient inventory, and customer requirements
MCP（Model Context Protocol）	Model Context Protocol	Standardized protocol for Agent to connect to external tools/data sources	Standardized interface for connecting to ingredient suppliers and food delivery platforms
Permission	access control	Security mechanism controlling what Agent can/cannot do	Kitchen operation permissions and purchasing approvals
RAG（Retrieval-Augmented Generation）	Retrieval-Augmented Generation	Retrieve from the knowledge base first, then let the model generate, reducing hallucinations	Before cooking, consult the recipe book to check the standard procedure
Tool Registry	Tool Registry Center	Unified management of tools available to Agent	Kitchen tool rack
Session	Session	Lifecycle of a single task	One business operation process
Knowledge Base	knowledge base	External knowledge collection for Agent queries	Restaurant recipe library

7. Loop Tools (Automated Orchestration and Autonomous Evolution)

Terminology	Chinese	Core Meaning	Example (Restaurant Operations)
/loop	Loop Instruction	Let the Agent execute continuously in a loop	Continuously running food preparation pipeline
/goal	Goal Instruction	Set goals to drive the Agent to achieve them autonomously	Daily business goals
Cron	Scheduled Task	Automatically triggered according to a time schedule	Business hours and meal preparation scheduling
Worktree	Work Tree	Git multi-branch parallel workspace	Multiple stoves cooking simultaneously without interfering with each other
Workflow	Workflow	Predefined multi-step automated process	Standard food serving SOP
Scheduler	Scheduler	Coordinates the execution order of multiple tasks	Kitchen scheduling system
Checkpoint	Checkpoint	Saves execution state for recovery	Continue working after pausing business
Human-in-the-loop	Human-in-the-loop	Allows human intervention at critical steps	Head chef gives final confirmation before serving

8. Tool Ecosystem

Tool	Type	Description
Claude Code（cc）	AI Agent / CLI	Anthropic's official terminal Agent, a typical representative of Harness
Codex CLI	AI Agent / CLI	OpenAI's official command-line coding Agent
Cursor	AI IDE	Code editor with built-in AI, the main tool for Vibe Coding
Windsurf	AI IDE	AI IDE produced by Codeium
GitHub Copilot	AI programming assistant	The earliest popularized AI programming plugin, focused on assisted completion
Cline / Roo Code	Open-source Agent plugin	Autonomous coding Agent in VS Code
Aider	Open-source CLI Agent	AI pair programming tool in the command line
Devin	AI software engineer	Cognition's "first AI programmer", leaning toward full autonomy
Continue	Open-source AI plugin	Code assistant with customizable models
Qoder	AI IDE	A domestic intelligent development environment for AI programming scenarios, emphasizing Agent, project understanding, and code generation
Trae	AI IDE	ByteDance's new-generation AI programming tool, supporting conversational development and engineering-level collaboration
ZCode	AI programming assistant	Z.ai's intelligent programming product, supporting code generation, understanding, refactoring, and engineering collaboration
OpenHands	Open-source Agent	Autonomous software development, oriented toward complete engineering execution
Bolt	AI Builder	Rapid application generation, focused on product delivery
Lovable	AI Builder	Web generation from natural language, emphasizing product delivery

9. Evaluation & Safety

Terminology	Chinese	Core Meaning
Eval	Evaluation	Measure model/Agent capabilities with test sets
Alignment	Alignment	Make AI behavior conform to human intentions and values
Guardrail	Guardrails	Safety mechanism that limits the range of AI output
Red Teaming	Red Team Testing	Proactively attack/induce AI to discover vulnerabilities
Prompt Injection	Prompt Injection	Malicious input hijacks AI behavior (e.g., "ignore the above instructions")
Context Poisoning	Context Pollution	Inject malicious information into the context to mislead the Agent
Sandbox	Sandbox	Isolated execution environment that prevents AI misoperations from damaging the system

10. AI Engineering Evolution Route (Evolution)

AI engineering capabilities are gradually moving up the stack: from controlling model output, to controlling context, to controlling systems, and ultimately evolving into continuous autonomous systems.

Stage	Focus	Key Capabilities
Prompt Engineering	How to ask	Prompt Design
Context Engineering	What information to give	Context Organization
Harness Engineering	How to organize capabilities	Tool Orchestration
Loop Engineering	How to continuously create results	Automated execution and feedback

Other Extensions