Blog
AI is already writing a significant share of production code, forcing teams to rethink what makes a great engineer. According to Stack Overflow’s 2025 developer survey, 84% of developers now use AI coding tools. In Y Combinator’s Winter 2025 batch, a quarter of startups reported codebases that were 95% AI-generated.
The cultural shorthand for this shift, popularized by Andrej Karpathy, is “vibe coding”: describing what you want in natural language and letting AI handle code generation while engineers guide the work and review the results.
It’s fast, powerful, and it is changing the nature of software engineering.
For engineering leaders, the question is no longer whether to adopt AI coding tools — that decision has been made. The real question is how to build teams that excel in this new environment. The best engineers aren’t just coding. They are deciding what work AI should take on, how to break it into workable chunks, and how to verify that the results are correct, secure, and maintainable.
When much of the code comes from AI, the developer’s role shifts from producing every line to deciding what to build, how to build it, and how to know if it’s any good.
This often means taking on responsibilities that once sat primarily with tech leads and architects: breaking down features into AI-friendly steps, supplying the right patterns and constraints, reviewing for correctness and performance, and ensuring the pieces work together once deployed. These skills have always been valuable but they are becoming part of the daily reality for more engineers, not just those in formal leadership roles.
It is tempting to think this shift levels the playing field between junior and senior engineers. A smart new graduate with strong CS fundamentals can be highly effective with AI-assisted coding — but experience still brings advantages in judgment, trade-off decisions, and seeing the long game. The gap between good and great engineers has never been about raw output. It has been about judgment, problem solving, and making the right decisions under real-world constraints.
The best engineers in the AI era, regardless of tenure, combine deep fundamentals with the ability to guide and challenge AI output. They spot inefficiencies the model misses, debug from first principles when generated code behaves unexpectedly, and anticipate the architectural ripple effects of small changes. Just as important, they bring original thinking to problems AI can only remix from past patterns. They design novel solutions, reframe problems, and push beyond what is in the training data. They know how to set up AI for success, when to step in, and how to create what the model cannot.
AI can make an average team faster, but speed without direction is a liability. Without strong oversight, you risk code that passes tests but fails in production, inconsistent styles that erode maintainability, or hidden complexity that becomes a long-term tax. AI without a guiding hand is not leverage — it is acceleration toward bigger problems.
Engineering leaders now need to hire and develop people who pair strong fundamentals with the judgment to direct, evaluate, and integrate machine-generated work. That means framing problems so AI has the best chance of producing accurate, maintainable output, critically reviewing for subtle flaws, anticipating how components will behave in combination, and adapting approaches when output misses the mark.
And yes, they still need to code. For credibility. For skill retention. And for the many problems AI cannot yet solve well.
Soon, AI coding tools will be everywhere. The differentiator will not be having them — it will be how effectively your team uses them. The strongest teams will plan AI work deliberately, review it with the same rigor as hand-written code, and keep their fundamentals sharp enough to step in when needed.
Vibe coding is not the end of engineering skill. It is a shift in where that skill shows up. The engineers who will define the next decade are not just quick with a prompt. They are precise planners, uncompromising reviewers, and system-level thinkers who know when to let AI accelerate their work and when to take full control.
© RunLLM, Inc. 2025
.avif)