Most days, "writing code" goes like this:
- I describe what I want
- an agent writes it
- then I verify it
I type more English than code. I didn't have a number for that until I read Sonar's 2027 State of Code Developer Survey. I work at Sonar, which ran the survey.
First, who answered?
The survey reached 1,026 developers and engineering managers who use AI coding tools at work, and it ran in September 2026. 91% are senior ICs, managers, or directors and above, and about three in four respondents manage other developers. The full report has every chart and the method notes.
More reading than writing
People were asked how their working time splits between creating code and reviewing it, before AI tools and since. 18% say they spent more time reviewing than creating before. 53% say they do now.
The work moved: the typing went to the agent, and the judgment stayed with the dev, along with the blame if I approve something I shouldn't have.
Fewer people are sure they truly check everything
The statement was "I always check my AI-generated or assisted code before committing it." This year, 39% completely agree. Last year, 48% did.
Put that next to the first chart and it gets uncomfortable. More of the week goes to review, and fewer people will say, without hedging, that all of it gets checked.
The catch: that drop is only in the strongest answer.
Count the people who somewhat agree and you get 72% this year against 75% last year, which is barely a change. So I wouldn't tell anyone that developers stopped checking AI code.
What slipped is the word "completely." And I believe it, because an agent can write more in an afternoon than I can read carefully in a day, and I'd rather admit that than pretend.
It changes what I want from my setup. The lines I don't read still need a check. It can't be me, and it shouldn't be the agent that wrote them.
Do you read every line your agent writes before you commit? What do you skip?
Reviewing code is not the same as knowing it
The third chart is about people, and it's the one I keep thinking about. 58% are very or extremely concerned that over-reliance on AI erodes their team's understanding of the codebase. Last year that was 53%. And 53% feel the same way about their own coding abilities eroding, up from 48%.
The limit here is simple. It measures worry.
But the worry makes sense to me. I know a piece of code when I've written it, broken it, and fixed it at least once, and approving a diff somebody else wrote, human or agent, doesn't get me there. I can approve something on Monday and be lost in it by Friday.
What part of your codebase could you no longer explain without asking the AI?
Where this leaves us
Review takes up more of the week, fewer people are sure they check every change, and more of them worry about losing touch with their own code.
And I'm probably not going back to typing every line. 84% of people in this survey say AI helps them deliver features and fixes faster. But I won't merge a function I can't explain without scrolling back, and a green build doesn't settle that for me.
The 77-page 2027 State of Code Developer Survey has the rest of the numbers, and you can filter the key findings by industry, company size, or region.



Top comments (0)