When AI writes almost all code, what happens to software engineering?

Recent advances in coding models, especially Claude Opus 4.5, GPT-5.2 and Gemini 3, have made AI agents capable of completing substantial software tasks with limited supervision. Experienced engineers report a sharp change: work that once required hand-coding or teams of specialists can now be described, generated, tested and refined through conversation. Some have shipped production features from a phone, while others say agents have produced nearly all their recent code. The change is not simply better autocomplete; developers must learn to direct agents, manage their tools and context, and judge their fallible outputs. The emerging view is that software production is entering a new phase, with the amount of code written directly by people shrinking quickly.

That shift may reduce the value of rapid prototyping, memorized syntax, and expertise in a particular language or stack. But it raises the importance of broader engineering judgment: understanding systems, defining problems, reviewing results, testing thoroughly, and making sound product and technical decisions. Engineers who can lead work and connect software to user needs may become more valuable, not less. There are risks alongside the productivity gains: AI can generate more defects and complexity, exposing weak testing and engineering practices sooner, while pressure to produce more could worsen work-life balance. Product managers and engineers may also overlap more, as each can increasingly perform parts of the other’s work.