Code Overload: How the Generative Boom Turned Software Engineering into a Ghost Town

The open-plan floors that were once bustling with junior engineers exchanging Stack Overflow links are now quieter than they were a year ago. This is evident in almost every mid-size tech company in San Francisco. Desks are more vacant. Pull-request chatter used to fill Slack channels, but at two in the morning, AI-generated commits merge in bulk. Over the course of roughly eighteen months, something peculiar has occurred. The field of software engineering, which was once thought to be the safest of the century, is currently undergoing its most confusing transition since the shift to cloud computing.

The change was not gradual. It arrived quickly following the November release of updated versions of Codex and Claude Code by Anthropic and OpenAI. Coding agents that had previously been helpful but constrained suddenly became truly capable. The tools were quietly adopted by engineers who had previously opposed them. Working with the security startup StackHawk, one financial services company saw a tenfold increase in monthly code production, going from 25,000 to 250,000. A million lines of unreviewed code were in the backlog. When you try to picture someone reading that number, it ceases to be abstract.

Silicon Valley wasn’t truly prepared for the paradox that has emerged. Businesses are drowning in generated code, but the people who wrote it are being let go in waves. In recent months, Pinterest, Block, Atlassian, and Meta have all eliminated engineering positions, citing AI effectiveness in internal memos. According to Andrew Bosworth of Meta, projects that previously required hundreds of engineers can now be completed by tens. These are the kinds of lines that end up in a memo and spread throughout an entire industry in a matter of days.

However, the code is not self-reviewing. The majority of the excitement was overshadowed by that. The more output AI generates, the more senior engineers are required to identify security holes, subtle bugs, and the kind of architectural drift that only becomes apparent three quarters later when the cloud bill starts to rise. There aren’t enough application security engineers on the planet to evaluate what American companies alone are producing, according to Costanoa Ventures adviser Joe Sullivan. If they could find them, the big companies he counsels would each hire five to ten more. This might end up being the most sought-after position in technology for the remainder of the decade.

Observing the change from the outside has an almost melancholic quality. Operators of boot camps that were unable to enroll students quickly three years ago are now having trouble placing graduates. The rungs that early-career engineers used to climb—entry-level coding jobs—are disappearing. According to a Reuters article, former boot camp participants who retrained for tech are now vying for fewer junior positions. The creator of Ruby on Rails and longtime opponent of AI in programming, David Heinemeier Hansson, recently changed his mind. He acknowledged that his initial resistance stemmed from the tools’ inadequacy after trying the new models. That is no longer the case.

What happens when this code eventually breaks, however, is the deeper issue. A peculiar new type of technical debt—debt without authorship—is produced by AI-generated systems. Since no human ever took a particular shortcut, nobody can recall why. In InfoWorld, David Linthicum referred to this trend as the “AI coding hangover,” and businesses that hurried to replace developers are already experiencing it. The cost of cloud computing is increasing more quickly than income. Diagnosing outages is more difficult. In the hopes that no one will notice, some businesses are covertly rehiring engineers under different titles.

It’s difficult to avoid the impression that the industry is about to enter a phase that will shape the work of the next ten years, not just in the tech industry but in all fields. It’s highly likely that accountants, analysts, designers, and writers will experience the same fate as programmers. It appears that engineers who focus on judgment, architecture, and review—tasks that AI can help with but cannot fully own—will be the most successful. It’s still unclear if that will be sufficient to preserve the profession as a whole. As of right now, the tools are producing more code than anyone requested, and someone has to sort through it all in the background.