Something has shifted in software engineering, and if you haven't felt it yet, you're either working somewhere slow to adapt or you're in denial. Vincent, an engineering manager at Sanity who still codes, dropped a field report on Hacker News this week that cuts through the hype and gets into the actual mechanics of what happens when AI becomes your main development surface.

The Four Generations Framework

Vincent frames it as generational abstraction: assembly gave way to high-level languages (C, Java, Python), which gave way to frameworks and cloud services (the Stack Overflow era), and now we're in generation four—AI-first development. Each shift made implementation cheaper while moving the hard problems up the stack. "The code still matters," Vincent writes. "It is just no longer where the hardest constraint sits." The new bottleneck? Judgment, context-setting, and review. He asked his team of senior engineers (each with 10+ years experience) how much code they write by hand now. One said one percent. Another tried writing directly, then went back to the agent because it "no longer felt like the intuitive path." That's not a failure of skill—that's a workflow shift happening in real time.

The New Development Pipeline

The pipeline Vincent describes looks nothing like the code-first workflows most teams still operate on. It starts with shaping context, producing a spec, then letting an agent build and iterate until it's ready for human review. But here's the uncomfortable part: that human review becomes sharper, not softer. "You may understand the goal and the architecture, but the implementation can contain decisions you did not personally make." Small pull requests are non-negotiable. Vincent trains agents to produce PRs around 500 lines—1,000 max. Ten thousand lines? "The process has failed. No one is reviewing that properly." One recent stack showed what this looks like in practice: Claude created the entire PR chain, verified CI, and responded to AI reviews built into the pipeline. The engineer reviewed it. That was the job.

The Uncomfortable Failure Modes

Let's talk money because someone has to. Vincent's AI spend runs around $2,000 a month. "That is a lot of money but for me, the productivity gain is worth it." For most organizations rolling this out broadly, that's a real operating question—not a rounding error. Teams need to treat AI usage like infrastructure: measure it, budget for it, and ask whether it's producing better outcomes, not just more output. The creepier failure mode is acceleration of existing habits—good or bad. "AI does not create good products or engineering judgment," Vincent warns. "It accelerates whatever habits are already there." That means strong engineers get stronger. Engineers who were handed narrow tickets and executed without thinking about outcomes? AI makes that pattern louder, faster, and more dangerous. "The result can look impressive, but you still end up with code and documents that humans need to review, test, understand, and maintain."

Review Is Now the Hard Part

When you wrote code by hand, understanding came from writing it. With agents, that's gone. Vincent puts it plainly: "If you do not review carefully, you may be approving work you have never properly internalized." The review bar cannot drop just because the code was cheap to produce. Testing matters more, not less. Linters matter. CI matters. Automated reviews built into PR pipelines matter. And here's something nobody's talking about enough: the cognitive load shift. "Writing code did two useful things: it gave us a rhythm, and it forced us to internalize the system as we went." With agents, you must meet all of that later, in one review pass. Vincent describes it like CPU scheduling—allocating cycles, waiting on IO, deciding which agent gets context, which task is blocked, which thread should be killed before it wastes more time.

What This Changes for Teams and Careers

The work didn't disappear—it moved up a level. Valuable skills are now judgment, taste, verification, product thinking, and systems thinking. Syntax matters, but it's not enough. "Knowing what good looks like matters more." Smaller teams can do more because you need fewer people to move work from problem to implementation to release. The constraint becomes clarity, not capacity. For Vincent personally, this changed his career shape. "For years, I felt a real split between management and individual contributor work." AI-first development let him be an engineering manager and still ship production code—not as nostalgia or a side project, but as part of how he operates. "It did not just make me faster. It changed the kind of engineering career I could have."

Key Takeaways

  • Code is cheap now; judgment is expensive—protect the right things
  • Review quality cannot drop because AI produces code faster
  • Small PRs remain critical—10,000-line agent outputs are unreviewable failures
  • AI accelerates existing habits, good or bad—audit your team's thinking patterns before you audit their code
  • The cognitive load shifts from writing to reviewing and context-setting