停止只把 AI 用作自動補全：我的智能體編碼之旅
Stop Using AI Just for Autocomplete: My Journey to Agentic Coding
作者從只把 AI 當作自動補全，轉向智能體編碼，並建立開源 agentic-coding-template，讓智能體參與架構思考、日常開發流程及工具操作。作者曾以相關工作流程把一個有 20 年歷史、含 2-million-line PHP 程式碼的應用程式遷移至 Nuxt，並指出重複任務可改用本機腳本，避免消耗模型 token。
I see a lot of developers using AI tools like GitHub Copilot, ChatGPT, or Claude every day. But for many, AI is still just a glorified autocomplete or a faster StackOverflow. We are missing out on massive potential.
As the Zen of AI coding states: "Agents are not just for coding."
Moving from passive prompting to Agentic Coding changes how we approach software architecture. The AI becomes a true pair-programming partner, handling the boilerplate so we can focus on problem-solving.
Where Agentic Workflows Make the Biggest Difference:
The Central Context Hub: The AI isn't just a chatbot; it's where I structure my thoughts and tasks before writing a single line of code.
Killing Routine Tasks: The agent analyzes my diffs, writes meaningful commit messages, and helps with deployment preparations.
Docs That Don't Lie: Inline documentation and work-step logs happen almost automatically as part of the flow.
Expanding via MCP: This is the game-changer. Using the Model Context Protocol (MCP), the agent controls Playwright for E2E testing, reads logs, and interacts directly with the file system.
Building the Agentic Coding Template
To make this actionable, I created an open-source baseline: the agentic-coding-template.
I recently used these exact workflows to migrate a 20-year-old, 2-million-line PHP application to Nuxt in a remarkably short time. Here is what my Git history taught me along the way:
Code & Docs: Learning to provide precise context so the AI writes code that actually fits the existing architecture and documents it instantly.
Tooling & Git Flows: Moving the AI out of the editor. Having an agent that understands diffs and writes commits keeps the project history clean.
MCP Ecosystem: Watching the agent run a Playwright test, see the error, and fix its own code autonomously was a major "aha" moment.
When NOT to Use AI: Pragmatism matters. Instead of burning tokens on repetitive tasks, I now have the AI write robust local scripts that run for free and fast multiple times.
If you want to stop writing boilerplate and get back to software architecture, check out the template.
Link to Repo: https://github.com/wrufeger/agentic-coding-template
How are you integrating AI into your daily workflows?
來源：DEV Community · dev.to