Ruby on Rails creator David Heinemeier Hansson says the most experienced engineers, not juniors, are extracting the most value from AI coding agents, flipping the industry's expectations.
The dominant narrative around AI-powered coding tools has been one of democratization: lower the barrier to entry, let anyone build software, and watch junior developers close the gap with veterans. David Heinemeier Hansson, the creator of Ruby on Rails and co-owner of 37signals, is seeing something entirely different in practice. The people getting the most out of AI agents at his company are the ones who already know exactly what good code looks like.
Speaking on The Pragmatic Engineer podcast, Hansson explained that senior engineers at 37signals are working in parallel with multiple AI agents simultaneously, critically evaluating the output, and making fast decisions about whether the generated code meets production standards. The key word there is "evaluating." The agent does the heavy lifting, but the senior developer acts as a quality gatekeeper, redirecting and refining in real time. This is precisely the skill set that made them senior in the first place.
Hansson's central point is one that many AI tool vendors would prefer to sidestep. When code needs to run reliably at scale, serving millions of users, someone has to verify that the AI's output actually works. Junior developers, he argues, simply lack the mental models and battle-tested intuition to spot subtle bugs, architectural flaws, or security vulnerabilities buried inside AI-generated suggestions. They can run the code, see it compile, and assume it is correct. That is not nearly enough. The danger is surface-level correctness masking deep structural problems.
This concern is not hypothetical. As Business Insider reported, Amazon implemented a 90-day safety measure requiring two-person code reviews for changes to consumer-facing services, specifically tied to its internal AI coding tool, Kiro. When a company operating at Amazon's scale feels the need to slow down and add human checkpoints, it tells you something about the current maturity of these tools in production environments.
From skeptic to agent-first adopter
What makes Hansson's position particularly noteworthy is the trajectory. In January, he publicly stated that AI had not yet reached the competence level of a junior engineer. Six months later, his daily workflow is what he calls "agent first on everything." That is a significant shift for a developer known for his pragmatic, sometimes skeptical, stance on industry hype cycles. The difference is that he is not using AI to replace understanding. He is using it to multiply existing expertise. The agents handle boilerplate, generate first drafts, and explore options. The senior developer directs, validates, and ships.
This pattern has broader implications for how startups and technology companies should think about AI tool adoption. The temptation is to hand these tools to the least experienced team members and expect productivity gains. In reality, the ROI appears inverted right now. A senior engineer who can confidently review, accept, or discard AI-generated code in seconds gets a genuine multiplicative boost. A junior engineer who cannot reliably spot what is wrong with generated code may actually move slower, spending more time debugging opaque AI output than writing from scratch.
For founders and engineering leaders, the practical takeaway is straightforward. If you are investing in AI coding tools, pair them with your most experienced people first. Let those veterans develop best practices, build internal review patterns, and establish guardrails before rolling the tools out broadly. The technology is evolving fast, but the validation bottleneck remains firmly human, and it favors depth of experience over speed of adoption.
