When Anthropic CEO Dario Amodei predicted last March that artificial intelligence would write nearly all software code within a year, the tech world scoffed. It sounded like a typical executive overpromise, detached from the gritty reality of software engineering. But nine months later, the laughter has died down.
Boris Cherny, the developer spearheading Anthropic’s Claude Code, revealed a statistic last week that fundamentally changes the profession. Over the last 30 days, across 259 pull requests, every single line he contributed was generated by the AI he helps build.
“100% of my contributions to Claude Code were written by Claude Code.”
— Boris Cherny, Anthropic developer
This goes beyond internal marketing. Theo Browne, a technology commentator and CEO of T3 Tools, found himself building entire applications—including an image generation platform and a mobile app—without ever opening a traditional code editor. For two decades, productivity meant typing syntax into a terminal. Now, it happens in a browser tab.
The crisis of purpose
The transition has triggered a crisis of purpose for veterans. Browne described his CTO finishing days where he shipped massive features and resolved critical support tickets, yet felt he accomplished nothing. If you aren’t fighting with syntax, it doesn’t feel like work.
“You suddenly stop feeling productive.”
— Theo Browne, CEO of T3 Tools
The job has mutated from craftsmanship to orchestration. Early assumptions that AI would merely prop up junior developers were wrong. Industry data suggests the opposite. Senior engineers are adopting these tools aggressively, using them to bridge a time gap rather than a skill gap. They aren’t asking the AI how to write a loop; they are demanding it build entire subsystems.
Volume explosion, quality concerns
The result is an explosion in volume. Browne recounted generating a 12,000-line project in a single afternoon—a feat that statistically takes an average developer months to achieve manually.
Context
This aligns with Jevons paradox: as the cost of code plummets to zero, the demand for it doesn’t stabilize—it skyrockets. Companies that previously couldn’t afford custom software are suddenly entering the market.
But volume is not value. A recent report from CodeRabbit indicates that AI-generated code introduces 1.7 times more problems than human code. When the AI fails, it does so spectacularly, creating high-issue outliers that require intense scrutiny. The bottleneck has shifted from writing logic to reviewing an avalanche of automated output.
1.7×
More problems in AI-generated code vs human code
We have stopped writing software and started managing the entities that do.
The junior developer problem
This leaves the industry with an existential question. If senior engineers are becoming architects of AI agents, there is no clear path for a novice to build the intuition required to review that code. Browne admits he doesn’t know how a junior developer survives this shift. The only hope may lie in the tools themselves, with environments like Cursor adding “teacher modes” to explain the logic they just generated.
For now, the editor isn’t quite dead, but for the top engineers in the field, it is collecting dust.