The Engineer Who Stopped Writing Code

Boris Cherny has spent his career writing code. Now he doesn't write any of it. This isn't a retirement story—Cherny created Claude Code at Anthropic, and over the past year, he's watched it become capable enough that he's handed over the keyboard entirely.

The transition happened faster than anyone expected, including him. When Claude Code launched a year ago, it handled maybe 10 percent of his work. By May, with the Sonnet 4 and Opus 4 models, that number had crept to 30 percent. Then Opus 4.5 arrived in November, and something shifted. "It just suddenly jumped for me from like 50% to 100%," Cherny told The Vergecast. "And that was actually very sudden, but it also just felt very natural."

What changed? Two things, according to Cherny. First, the model started testing its own code—opening browsers, clicking around, adjusting elements that were off by a few pixels. Second, the code itself got good enough that he stopped needing to open a text editor to fix it. "I don't have to fiddle with it by hand. And that was actually kind of nice because that means I can move on to the next thing."

This is the inflection point people have been predicting for years, but it arrived quietly, during the holidays, when people went home bored and discovered Claude Code actually worked. No press release made this moment happen. The technology simply crossed a threshold.

Cherny's metaphor for the shift: "You used to play the violin and now you're conducting the orchestra." It's an appealing image, but it raises the question of what happens when the orchestra doesn't need a conductor either. For now, that's not the concern. The harder problem, Cherny says, is constantly recalibrating expectations. Things that didn't work with one model suddenly work with the next. "All the stuff that I would have thought didn't work, I just assume it'll work at some point."

Who This Is Actually For

Anthropic built Claude Code as a developer tool, naturally enough. It lives in the terminal. It speaks the language of engineers. But from the beginning, non-developers started using it anyway, jumping through technical hoops to access something that clearly wasn't designed for them.

Cherny noticed this first with Anthropic's own data scientists, who were running multiple instances of Claude Code in parallel, generating charts in the terminal. Then the sales team started using it. "I think now like half of our sales team uses Claude Code every week," he said. Companies like Ramp report product managers and data scientists using it alongside engineers.

This created a design problem. Do you teach everyone to use the terminal, or do you rebuild the interface? Anthropic chose both. They created Cowork, which runs the same underlying agent but adds safeguards—virtual machines, deletion protection, things that would annoy engineers but protect less technical users from foot-gunning themselves.

Meanwhile, Claude Code itself remains aggressively customizable. "Engineers love to customize everything," Cherny noted. "There's no two engineers that have the same setup." The tool has hundreds of configuration options. And because Claude Code is Claude Code, you can ask it to reconfigure itself. Don't like the color scheme? Tell it to change the theme. It will write the code to modify its own appearance.

This recursive capability—the tool improving itself at your direction—suggests where this technology might go. Not replacement of human judgment, but automation of everything that doesn't require it.

What Gets Automated First

Cherny maintains a simple heuristic: anything he does by hand becomes a candidate for automation with each new model release. This is busywork elimination, not creative work replacement. The journalist using Claude Code to consolidate notes from ten different apps into Obsidian isn't having the tool write his articles. He's having it handle the tedious migration work that would otherwise consume hours.

This distinction matters. The anxiety around AI coding tools often centers on whether they'll replace programmers. The more immediate reality is that they're replacing the parts of programming that programmers have always wished they didn't have to do. Testing. Configuration. Migration. The work that's necessary but not interesting.

Anthropic's broader strategy reveals itself here. For them, coding capability isn't just a product feature—it's the path to safe AGI. "The model is software," Cherny explained, "and the way that it interacts with the world is through tools and through other software that it writes." By keeping that interaction mediated through code, they maintain a layer of interpretability and control.

Whether that strategy works remains to be seen. But the tactical reality is clear: a year into Claude Code's existence, it has found product-market fit in a way few AI tools have. Spotify, Shopify, Netflix, Nvidia, Snowflake—the adoption list reads like a who's who of companies that make software.

The Conducting Problem

Cherny's violin-to-conductor metaphor only works if conducting remains a human job. Right now, that seems safe. The tool still requires direction. It still makes mistakes that need catching. But the trajectory is obvious. Each model release narrows the gap between "tool that assists" and "tool that executes independently."

Engineers, Cherny argues, are used to this. Tech stacks change constantly. New frameworks arrive. New languages become popular. Adaptation is part of the job description. But there's a difference between learning a new framework and watching your entire role transform from creator to orchestrator.

The question isn't whether this change is happening—it demonstrably is. The question is how quickly and how completely. A year ago, Cherny was writing 90 percent of his code. Today, he writes none of it. That's not a gradual transition. That's a phase change.

And he's the person who built the tool. If the creator of Claude Code no longer writes code, what does that signal for everyone else?

Bob Reynolds is Senior Technology Correspondent at Buzzrag