Software development is where the impact of AI showed up earliest and runs deepest. Tools that were completing single lines two years ago now drive whole tasks: opening files, editing them, running tests, and repeating until they pass.
What actually changed
Teams that adopted this properly report a large share of their code being generated, mostly in the repetitive parts. Boilerplate, tests, migrations, documentation, porting between languages. That work was always tedious and is now largely handled.
What did not
What has not moved to the machine is deciding. Which architecture, which trade-off, what can be simplified and what cannot. An agent can write a hundred variants of a solution, but it cannot know which one fits your organisation’s constraints, because those constraints are written down nowhere.
The same holds for security and for the core logic of a system. The practice that has settled in serious teams is that AI writes a large share of the code while an experienced engineer reviews whatever is critical.
The new skill
The job has shifted from writing to reviewing. That sounds like relief, but it is not. Reviewing someone else’s code is harder than writing your own, and reviewing code that looks convincingly correct is harder still.
One concrete consequence is that junior developers have a problem. People used to learn by writing dull code until it became easy. If a machine takes that part, the path from beginner to someone who can judge quality becomes unclear. This is an open question nobody has answered well yet.
What to do if you lead a team
Introduce a rule that the author understands every line they submit, regardless of who wrote it. That single rule solves most quality problems.
Measure what matters – production defects, time to delivery – and not lines of code. Line count is now a useless signal.
For solo developers
The advantage is disproportionately larger for small outfits than for big ones. One person with good tooling now ships what recently took three. If you work alone, this is the biggest single productivity jump in a decade and it would be a waste not to use it.