Deedy Das, a partner at Menlo Ventures who invests in AI and enterprise software companies, dropped a blunt thread on X this weekend that's been rattling around Silicon Valley since. His take: most software engineers are caught in an 'identity crisis bordering on depression' as AI coding tools reshape what it means to actually write software for a living. The catalyst isn't the technology itself—it's how different engineers are choosing (or being pressured) to use it, creating a growing chasm inside engineering teams that's fueling burnout and resentment.

Two Flavors of Engineer

Das breaks the new reality into two camps. First up: what he calls 'lazy' engineers—workers who lean heavily on AI to write code, answer questions, prepare updates, and handle tasks with minimal actual engagement. Then there are his 'craftsmen,' experienced veterans who end up bearing the full cognitive load of understanding, reviewing, and fixing the ever-growing flood of AI-generated output landing in their pull requests. It's a division that's becoming structural at companies pushing aggressive tokenmaxxing strategies to boost productivity metrics—but leaving senior engineers drowning in technical debt they didn't create.

The Review Bottleneck Nobody Talks About

Here's where it gets gnarly. As AI makes code generation trivially easy, the actual bottleneck has shifted downstream to review and maintenance—work that requires deep context, architectural understanding, and institutional knowledge that AI tools simply don't have. 'The craftsmen are tired,' Das wrote. 'Very tired. The entire burden of review falls on the craftsman. The burden of understanding.' This echoes what's been called 'botsitting' in other contexts: supervising AI systems, fixing their hallucinations, validating outputs that look correct but miss edge cases only a human would catch.

AI Sprawl Is Making It Worse

The problem extends beyond individual team dynamics. Business Insider's Amanda Hoover recently documented what she calls 'AI sprawl': workers juggling multiple AI tools simultaneously, duplicating work across different platforms, and generating ever-increasing volumes of output without clear evidence that companies are becoming substantially more effective. Engineers report buried pull requests while bugs increasingly slip into production—code that looks syntactically valid but contains logic errors a human reviewer would catch in seconds.

Larger Organizations Hit Hardest

Das notes this tension tends to emerge most acutely in bigger, older companies with higher talent variance—organizations typically 10+ years in operation where engineering teams span juniors to principals with wildly different expectations around AI adoption. 'It happens,' he said of the dynamic. 'A lot.' Not every company sees this fracture, but it's common enough that anyone watching the human side of AI integration closely should be paying attention.

Key Takeaways

  • Tokenmaxxing strategies are creating a two-tier engineering culture: minimal-engagement AI users versus overloaded craftspeople doing review work
  • The actual bottleneck in modern software development has shifted from code generation to code review and maintenance—work that requires irreplaceable human judgment
  • Larger, older organizations with diverse skill levels face the most acute version of this tension
  • Engineers report mounting pull request backlogs while production bugs increase, suggesting quality control is struggling to keep pace with AI-accelerated output

The Bottom Line

The uncomfortable truth here is that we're watching the craft of software engineering get disaggregated in real-time—and not everyone is going to land on their feet. Tokenmaxxing looks great in productivity dashboards until you realize you're just moving work downstream onto your most experienced people who didn't sign up for quality-assurance duty at scale.