AI is fundamentally changing how junior developers learn, but it’s also making senior developers a whole lot less valuable if all they’re doing is taking well-defined tickets and turning them into code. As routine implementation costs go down, organizations will pay less for implementation alone. It’s an uphill move, and that’s the real challenge. AI-assisted workflows are making it much easier to replace developers of any seniority level who cannot clearly define problems, constrain technical solutions, verify generation, and truly own production outcomes. The recent HN debate is a good snapshot of the anxiety around junior developers but that exact same pressure is now moving up the chain to senior roles. Title and seniority count for very little when pure code production is a cheap commodity.
Implementation is cheaper
We see the implementation-only developer at all levels of an engineering organization. This is the developer who wants a pristine ticket, values themselves by the amount of code they push, refuses to work on a product that is not well defined and passes the tough decisions of system architecture to someone else. AI shines when the work is well bounded and implementation heavy. The limited conditions in which generative tools are most useful were highlighted in Copilot research, which showed developers using AI were able to complete a scoped, simple programming task much faster than those working without it. Similar speedups were found in a Google RCT on real C++ tasks, though that study had wide confidence intervals, and was heavily dependent on the developer’s familiarity with the task. Of course code still matters, but plausible code is simply cheaper to produce. The real work of software engineering is determining what to build in the first place, what constraints matter, and ensuring that the end product is safe to deploy.
Seniority has to prove itself
A study by METR challenges the assumption that experienced developers are inherently safe. We investigated 16 experienced open-source contributors working on their own mature repositories, randomly splitting 246 real-world issues into AI-allowed and AI-disallowed conditions with frontier models from early-2025. In this particular, complicated set-up, developers with AI access actually were 19 percent slower. This is especially noteworthy because prior to the study, those same developers predicted a 24 percent speedup and even after it was complete, still believed AI made them about 20 percent faster. The narrow scope of the study matters because it captured early-2025 tools in complex codebases with very high implicit quality bars, and their subsequent uplift update showed the tools did improve. But in practice, code generation is seldom the actual bottleneck of a system. The real cost is in understanding what should go in the target system and what invariants should hold and whether the generated output actually integrates correctly. Empirically, it says otherwise, but it feels much faster for developers. Although it feels productive for the day, writing lots of code is a poor proxy for delivery value, and perceived productivity is a poor proxy for delivery value.
Faster code, slower delivery
This very pattern is observed at the organizational level in the DORA research. Their research shows that a 25 percent increase in AI adoption correlates with improvements in local process metrics such as code quality, quality of documentation, and speed of code reviews. But at the same time, it correlates with a 1.5 percent drop in overall delivery throughput and a 7.2 percent drop in delivery stability. Faster code generation directly results in larger files and changesets that are much more difficult to review, more dangerous to merge, and more difficult to roll back when things break. The review process becomes a significant bottleneck. In such cases, teams need developers that can keep changes small, understandable and easy to revert. DORA findings show that while developers using generative AI report higher levels of flow and personal satisfaction, they spend less time on what DORA categorizes as valuable work. The time saved is often lost in downstream manual toil, meetings, and bureaucratic hurdles. The problem is the gap between feeling fast and actually delivering value to the system.
What I'm building
Delegate tasks. Get software.
Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.
Take a look at vroni.comI respect your privacy. Unsubscribe at any time.