Exploring the Evolving Role and Core Value of Software Engineers in the AI Era

On November 5, 2026 at 09:06, Juejin published "What Are People Doing in AI Coding?", a piece asking what remains of a developer's job when AI can read code, change features, write tests, and even keep attempting fixes based on error messages. The article frames a concrete question: if code is mostly written by AI, what is left for developers to do.

Instead of surveying tools or benchmarks, the article walks through one realistic development process and argues that between stating a requirement and generating code there are many points where a human still has to judge. The faster code is produced, the more these decisions shape the final result. A cancel-order scenario illustrates the kind of rules that need settling: which orders can be cancelled, whether a paid order can be cancelled, whether a cancellation triggers a refund, when inventory is released, how repeated clicks are handled, and which state an order should stay in when an external system call fails. The article says AI can help list these questions and can guess answers from existing code, but a guess still needs confirmation, especially where business rules and trade-offs are involved.

Two easily missed cases get particular attention. Users often need only to submit a cancellation request, with the actual cancellation remaining subject to review; if AI is asked to implement instant cancellation from the start, the more complete the code looks, the larger the rework may turn out to be. And when a project already has a unified pagination entry point, leaving that context out can lead AI to write a second pagination implementation that runs on its own yet raises maintenance costs once placed inside the project. The article also notes that when documentation describes one rule and the code implements another, someone must confirm which one reflects the current business, and that this pushes developers to write down the experience previously kept in their heads so it becomes reference material colleagues and AI can rely on.

On AI's tendency to keep going after a task is done — tidying nearby code after fixing one problem, extracting a shared component after adding an endpoint — the article takes a measured view: each change may have its own justification, but every added modification expands the surface that must be understood, reviewed, and verified. Citing a case where the task was merely fixing decimal-place display for an amount but AI also altered things on its own initiative, the piece concludes that developers should focus on the critical path and the business boundary, and positions their work as clarifying which requirements are worth building, supplying project context, finding trustworthy sources, resolving conflicts between materials, and verifying results independently.