This document provides an introductory overview of Agentic Coding, presenting the current landscape of the field.
So, what is this article about?
The goal is to:
Summarize core concepts of Agentic Coding
Explain practical ways to work with and apply Agentic Coding in real projects
Clarify several non-obvious or confusing topics that many people often wonder about
Main sections:
Conceptual foundations (things you must understand to practice effectively)
The current chaos of Agentic AI tools
Why there are so many Kits and Skills
Achieving 5×–10× coding performance
(These are required to understand and practice Agentic Coding effectively)
The following concepts are organized from beginner level upward to make them easier to approach.
Most modern Agentic Coding tools are built around these ideas. The community has produced a large amount of documentation and tutorials—easy to find—but understanding what these concepts actually mean is the key.
Command /kəˈmɑːnd/ → lệnh, chỉ thị
When you install an AI Agent IDE such as Cursor, Antigravity, or Claude, and type something like “create a new file”, that is a command.
This is the most basic interaction model:
One task → one instruction → AI executes it
Plan /plæn/ → kế hoạch
One drawback of simple command-based interaction is that you only see what the AI did after it finishes.
In Plan mode, the AI first explains:
What it intends to do
How it will do it
You can review, discuss, and modify the plan. Once the plan looks good, execution begins.
Example: To store a light/dark theme setting, options include localStorage, sessionStorage, cookies, or backend storage. Without a plan, the AI may choose an approach that seems reasonable to it—but not to you.
Rule /ruːl/ → quy tắc
When coding with AI, you often repeat the same constraints.
For example:
Your team agrees on CamelCase naming
But code style conventions suggest snake_case
You keep reminding the AI: “Use CamelCase”
Instead, define this once as a rule, and the AI will consistently follow it.
Examples:
“Always use Tailwind CSS for styling”
“Never install new libraries without approval”
Skill /skɪl/ → kỹ năng
If you keep repeating the same instructions across similar tasks, you may want to package them into a reusable skill.
Rules define what is allowed or forbidden. Skills define how to solve a specific type of problem, in your preferred style.
Example: Instead of repeatedly saying:
“Write REST APIs”
“Include try/catch”
“Log errors and retry if needed”
You define a skill called API_Standard, and simply invoke it.
Protocol /ˈproʊtəkɒl/ → giao thức
If you have ever had to describe UI designs to an AI step by step, MCP removes that friction.
MCP allows AI to:
Connect directly to Figma and read designs
Fetch up-to-date documents automatically
Instead of guessing, the AI can see and read external sources.
Sub-agent /sʌb ˈeɪdʒənt/ → tác nhân phụ
AI agents have limited context windows (e.g., 200k tokens). Large tasks with many sub-tasks can degrade performance over long conversations.
Sub-agents act like specialized assistants, each responsible for a specific domain or task.
Hook /hʊk/ → móc nối, điểm chặn
If certain actions should always happen after the AI finishes a task—without requiring reasoning—you can use hooks.
Examples:
Automatically run Prettier after code generation
Run security scans after AI changes code
Workflow /ˈwɜːrkfloʊ/ → quy trình làm việc
Many people apply Agentic Coding by mimicking real-world software development:
Requirement → Design → Planning → Coding → Review → Testing → Deployment → Monitoring
Workflow connects all concepts into a logical sequence, defining when and how each concept should be used.
With good prompts + solid workflows, output quality improves dramatically.
Kit /kɪt/ → bộ công cụ
You may understand Rules, Skills, Plans—but still struggle to write effective prompts.
That’s where kits come in.
Instead of:
Writing React rules from scratch
Configuring MCP for PostgreSQL manually
Others have already bundled best practices into standardized kits.
Concepts (Rules, Skills, MCP, …): Tools
Workflow: How to use tools effectively
Kit: A standardized bundle of tools + workflow
There are many Agentic AI tools today: Cursor, Antigravity, Claude Code, OpenCode, Augment, etc.
This abundance can feel overwhelming.
However, discussions with practitioners reveal:
For ~90% of tasks and users, these tools are not fundamentally different.
If you are a typical developer working on company tasks, most tools will work just fine.
CLI (Command Line Interface) /ˌsiː el ˈaɪ/ → giao diện dòng lệnh GUI (Graphical User Interface) /ˌdʒiː juː ˈaɪ/ → giao diện đồ họa
GUI tools: Friendly UI, easy onboarding, strong UI preview
CLI tools: Terminal-first, highly customizable
Choose based on preference.
FOMO /ˈfoʊmoʊ/ → sợ bị bỏ lỡ
Common reasons:
Strong trend-setting (skills, workflows, kits)
Perceived higher intelligence
Better performance on long, heavy coding tasks
High code output with relatively low cost
For typical users: differences remain marginal.
A practical trick:
Use Claude Code as a proxy
Connect it to cheaper third-party model subscriptions
Switch accounts when quotas are exhausted
This setup is ideal for:
Solo builders
Freelancers
Experts with frequent new projects
Cursor: Smooth UX, editor-first AI integration
Claude Code: Fast, terminal-based, strong agent behavior
Antigravity: Thought-flow and context control
OpenCode: Open-source and transparent
Augment: Optimized for large codebases
Don’t over-FOMO for most use cases
Special needs require special setups
Every day, new Kits and Skills appear:
AI Dev Kits
Claude Kits
Workflow frameworks
Specialized UI/UX or React skills
These packages are powerful because:
They reduce setup overhead
They are tested and refined by their creators
If it feels overwhelming, remember:
Kits are meant to increase productivity, not create confusion.
Pick a few that match your workflow and update occasionally.
Performance /pərˈfɔːrməns/ → hiệu năng
Claims of “5×–10× performance” should be compared against:
No AI usage
Or very basic command-only AI usage
Agentic Coding shines when concepts, workflows, and kits work together—not when AI is used as a simple autocomplete.