AI wrote 80% of her code. She's terrified.

Amy Surrett graduated with a Bachelor's in software development in 2022, just months before ChatGPT launched. Now, just three years later, she estimates AI tools write 80-90% of her code. A year ago, that figure was 5-10%. The shift happened that fast. In January, she booted up Anthropic's Claude Code to build a payment feature for a client. Coding it by hand would have taken two or three days. Claude did it in just over an hour. "It felt like the point of no return," she told Business Insider. "This industry is not going to be the same. My job is not going to be the same." She is not alone. Across the software engineering world, a reckoning is underway. In the space of a few weeks late last year, OpenAI, Anthropic, and Google all released new AI models that drastically improved their coding tools. Almost overnight, AI was suddenly good at complex tasks - the kind that used to take humans years to master. Andrej Karpathy, a former founding research scientist at OpenAI who recently joined Anthropic, wrote in a February X post that it was "hard to communicate how much programming has changed due to AI in the last 2 months." Before December, he said, coding agents "basically didn't work." Suddenly, there was takeoff. At tech giants like Google, AI is now writing as much as 75% of the company's code. The implications are staggering. Software engineering, a role that employs tens of millions globally, is undergoing a full-blown transformation. AI is sparking layoff fears, spawning a new lexicon of terms like "tokenmaxxing," and driving hundreds of billions of dollars of AI investments. For engineers, the emotional impact is complex. They're more productive than ever. They're inspired. And some of them are terrified. "It's a double-edged sword, because in some ways I'm getting more done but also doing less work, so it feels less productive," said Surrett. Kent Dodds, who left his job at PayPal in 2019 to run a business teaching software engineers, experienced his own existential crisis in January. He set up an agent with Cursor, an AI-assisted coding tool, to build a tool that would let his students download his videos to watch offline while encrypting them to prevent unauthorized sharing. The agent "nailed it on first try," he said. Potentially weeks of work had evaporated in a single morning. "That was my first existential crisis," he told Business Insider. It wouldn't be his last. Over the past several months, these models have continued to improve, which Dodds said means they are "behaving a lot more like a regular software developer." "I don't know what the ceiling is, or how fast we're going to hit it, but we certainly aren't anywhere close to it just yet," he said. The role of the software engineer is quickly becoming that of an agent supervisor. At the AI Engineer Europe conference in London in April, Ryan Lopopolo, a member of OpenAI's technical staff, said onstage: "In the last six months, we have seen coding agents take over the world." Overseeing agents used to be "more of a synchronous process," Alex Ponomarev, the founder of Volt, a boutique software development agency, told Business Insider. "You'd have Claude Code running, it does something, then it stops, you have to tell it what to do next." That's no longer the case, he said. That doesn't always mean less work. Some engineers have expressed frustration about having to clean up AI-generated code or fix clunky vibe-coded apps built by tech-novice coworkers. With more time spent managing agents and less time in the code, some developers' schedules look different, as they restructure tasks around creating specs for AI and take breaks while their token limits reset. "It's basically not even worth my time to be manually writing code when I can have something like Claude doing it for me," Danial Qureshi, a software developer in Toronto, previously told Business Insider. The catch, of course, is that this extraordinary new power is nobody's secret weapon. If you have it, so does everyone else, and it's getting harder to compete against the bots. In response, engineers are doubling down on their distinctly human qualities. Over the last year, Dodds said the number of questions he receives from students has seen a "drastic decline." He chalks it up to coding agents' ability to answer those queries much faster. Where Dodds not long ago would teach engineers how to code, today he has a new curriculum for what he calls "product engineering" - a focus on what to build, not how to build it. The value of humans, as Dodds sees it, is judgment. What problems should be solved? What are the benefits and drawbacks? What would be truly useful to the user? "I'm teaching the last skill that the last software engineer needs to have," he said. For executives, the strategic stakes are clear. The companies that figure out how to harness AI coding tools while retaining top human talent will have a massive advantage. Those that don't risk being left behind. The question is no longer whether AI will change software engineering. It already has. The question is how leaders will adapt.