The Code Writing Revolution: Why Software Engineers Are Letting AI Do All the Work

Software engineering just hit an inflection point that nobody saw coming this fast. Engineers at the world's leading AI companies have stopped writing code. Not most of their code, not 80% of their code, but all of it. Every single line now comes from the same AI models they're building.

Boris Cherny leads Claude Code development at Anthropic, and he recently revealed he hasn't manually written a single line of code in over two months. When AI researcher Andrej Karpathy asked about the percentage of code written by AI at these companies, Cherny's answer was blunt: 100%. He shipped 22 pull requests one day and 27 the next, each one entirely generated by Claude. A researcher at OpenAI echoed the same reality, admitting they don't write code anymore either. Their take? Programming was always tedious grunt work, and they're relieved it's over.

What This Means for the Industry

The transformation goes deeper than productivity metrics. These AI tools are fundamentally changing what it means to be a software engineer. Anthropic now hires primarily generalists instead of specialists because traditional programming expertise matters less when AI handles implementation. Engineers focus on directing what gets built rather than building it themselves, becoming editors and architects instead of coders.

The numbers tell the story. While Anthropic reports 70 to 90% of their code comes from AI company-wide, other tech giants lag behind with Microsoft and Salesforce sitting around 30%. A recent study found roughly 29% of Python code on GitHub in the U.S. is now AI-generated. That gap won't last long, though. Cherny predicts most companies will reach similar adoption rates within months because the technology works too well to ignore.

The Entry-Level Problem Nobody Wants to Discuss

Here's where things get uncomfortable. Entry-level software engineering positions are disappearing as AI-generated code increases, and while whether these trends are directly linked remains unclear, the correlation is impossible to ignore. Traditional career ladders in software development always relied on junior engineers grinding through repetitive tasks to build skills, but that training ground is vanishing. If AI handles the tedious work that taught fundamentals, how do new engineers develop expertise?

Tech companies frame this shift as democratization, claiming anyone can now build software by prompting AI in plain English with no technical background required. But democratization and job displacement often look identical from different angles. The promise of accessibility doesn't address what happens to the traditional pathways into the profession.

Why Engineers Love It (Even If It Threatens Their Field)

Despite the existential questions, engineers using these tools report unprecedented joy in their work. Cherny describes his current role with genuine enthusiasm, explaining that all the tedious aspects of coding disappear and he spends his time thinking creatively about what to build next instead of wrestling with syntax and debugging. Claude even handles his project management tasks, automatically messaging team members when they haven't updated shared documents.

This liberation from grunt work explains why adoption spreads so quickly once engineers try these tools. The experience of focusing purely on creative and strategic decisions while AI handles execution is transformative. One engineer built Cowork, Anthropic's file management agent for non-coders, in roughly ten days using Claude Code. That kind of velocity was unthinkable even two years ago.

The Limitations Still Matter

AI-generated code isn't perfect. Models make conceptual errors, overcomplicate solutions, and leave unused code scattered through projects, which means engineers still need to review, edit, and guide the process. But here's the key insight: those limitations are temporary, and code quality improves with each model iteration. The engineers who've gone all-in on AI generation aren't concerned about current imperfections because the trajectory is clear.

What Happens Next

We're watching an industry transform in real time, and the question isn't whether AI will write most code but how quickly other companies catch up to the leaders. Anthropic's CEO Dario Amodei predicts AI will handle most or all software engineering work from start to finish within six to twelve months. That timeline felt absurd six months ago, but today it looks conservative.

The shift extends beyond coding. Cherny expects similar statistics for all computer work in the coming months because if AI can write code at this level, it can handle other structured digital tasks just as effectively. The software industry built the modern world, and now it's rebuilding itself using the tools it created. Whether that rebuilding creates more opportunities than it destroys depends on how we navigate the transition.

For now, the engineers at the frontier aren't writing code anymore. They're teaching AI to do it for them, and by all accounts, they've never been happier with their jobs. The rest of us are just watching to see if that happiness spreads or if we're witnessing the beginning of something more complicated.