The Next Step in AI Programming: Let the Model Modify Code Itself, You Just Review and Approve
From autocomplete to autonomous execution, AI coding assistants are starting to modify code and run commands on their own. Developers are shifting from writing code to reviewing code. The workflow is being restructured.
Over the past year, AI coding tools have undergone a clear transformation in form. The earliest batch of products, such as GitHub Copilot, focused on completion. You write code in the editor, and it suggests the next line beside you. Essentially, it was still an advanced input method.
By 2025, things had changed. Tools represented by Cline began taking over the entire coding process. They no longer wait for you to type; instead, they read the codebase themselves, modify code themselves, run command lines themselves, and fix errors themselves. What developers do has shifted from "writing code" to "reviewing and approving."
When akshay_pachaar discussed this type of tool on X, the discussion under one tweet focused on the same question: should we trust AI to modify code on its own? Original tweet link
If you haven't used Cline yet, here's a brief introduction. It is an open-source VS Code extension. After installation and configuring the model API, you describe the task in natural language, such as "add form validation to the login page." It handles the rest on its own: finding the relevant files, modifying code, running tests, and continuing to adjust based on errors.
A few notable designs:
- Plan / Act mode. Plan first, then execute. It lists the steps, and only after you confirm does it start modifying code.
- Checkpoint rollback. A snapshot is generated for each operation, so if something breaks, you can restore it with one click.
- MCP support. It can connect to external tools, such as databases or browsers, so that in addition to modifying code, AI can also query data and run pages.
These designs solve a key problem: trust. If AI modifies code on its own, you can't watch every step. With planning and rollback mechanisms, the developer's role shifts from "the person doing the work" to "the person making decisions."
The most ironic part of this is that in previous years everyone discussed whether "low-code" would replace programmers, but the answer may be exactly the opposite: AI writes more code, and programmers are responsible for judging which code is worth keeping.
The future development process will be very different from now. You don't need to know how a function is implemented, but you do need to know why this function exists. Requirements decomposition, code review, and system design—these abilities will become increasingly valuable. The work of typing out code will be done by AI for you.
If you are also a developer, try this type of tool on a weekend. Treat it as an intern who is not very reliable but learns very quickly. Keep a close eye on it, but don't do the work for it.
发布时间: 2026-11-23 06:04