AI Coding Tools Can Build Apps, But Still Need You

Imagine asking an AI to build an entire website for you, not just write a few lines of code. That's what "agentic IDEs" promise: programming tools that act like autonomous helpers, taking natural-language instructions and creating complete applications. Researchers from the University of Milano-Bicocca put three of these tools — GitHub's Copilot, Cursor, and Windsurf — to the test. Their job: build five full web apps from scratch, with no human coding in between.

Here's the good news: the AI tools were surprisingly skilled at the boring, repetitive parts of software development. They nailed standard features like CRUD operations (creating, reading, updating, and deleting data) and authentication (login and signup systems). These are the bread-and-butter tasks that make up a huge chunk of everyday programming. For those, the tools were "high maturity," meaning they produced working, reliable code most of the time.

The bad news? When the researchers asked for something out of the ordinary, things fell apart. One test involved a "task queue architecture" — a way to manage background jobs, like processing videos or sending emails in batches. The AI tools made significantly more errors here. This suggests that AI assistants learn patterns from the mass of existing code they're trained on, so they're great at common things and less reliable for creative or unusual solutions.

The bigger picture: these tools are not a replacement for human developers. Instead, they're shifting the developer's role from writing every line of code to acting more like a manager — giving instructions in plain English, checking the AI's work, and fixing mistakes. That's a real change in the job, but not the end of it. For the rest of us, it means software may get cheaper and faster to build, but we'll still need humans to catch the errors and make sure the final product actually works.