Engineering career ladders depend on a broken assumption: juniors learn by doing work seniors don't want. That assumption has been quietly breaking for two years. At QCon London, Alasdair Allan quantified the collapse in a talk titled "Engineering Progression When AI Ate the Middle."
The data is uncomfortable across the entire AI-productivity debate. A METR randomised controlled trial found experienced developers were 19% slower with AI tools while believing they were 20% faster — the tools degraded performance while masking that fact. Anthropic's internal study of junior engineers showed they scored 17% lower on code mastery without finishing tasks faster. "They traded learning for nothing," Allan said. On real open-source codebases, zero out of fifteen AI-generated PRs were mergeable despite passing automated tests. AI implements functionality. It does not yet do craft.
Allan identifies three connected stages. First: AI stunts skill formation. Junior engineers skip the hands-on debugging, incident response, and legacy-code work that built pattern recognition in previous cohorts. Second: that same AI converts engineering work into supervision. Anthropic's engineers use Claude in 59% of daily work but fully delegate only ~20%, leaving a judgment gap that demands skills juniors are not accumulating. Third: AI slows entry-level hiring. The work persists — Allan notes no decrease in work for engineers over 25 — but the junior positions that produced senior engineers are vanishing. "People weren't being sacked; they just weren't being hired."
The operational pressure lands hardest in what Allan calls "blackfield" systems: legacy code under load, slated for deprecation but never executed. The spec was never written or was written and made obsolete by undocumented decisions. Business rules live in code conditions that outlived everyone who understood them. AI can read code and test documentation. It cannot read production traffic. It cannot detect which code paths are load-bearing in ways the code doesn't declare. That context lives in senior engineers' heads — heads filled by fifteen years of scut work.
The organizational feedback loop demands attention. The same AI tools generating short-term productivity gains are degrading the pipeline that produces people capable of supervising those tools at scale. Adding friction: 80–90% of engineering questions now go to AI rather than colleagues, bypassing the learning that comes from struggling through a problem with a mentor. The answer arrives without the journey.
Allan's prescriptions acknowledge their limits — "we're at the diagnosis stage, not the solution stage" — but translate into concrete choices. Structured rotation through fundamentals, modeled on medical residencies where scut work is required to build judgment. Promotion criteria shifted from velocity to demonstrated understanding: watch how engineers think, not what they ship. Internal documentation as infrastructure — written for a senior engineer who can code but doesn't know the codebase, not as an AI summary. And explicit framing that AI is a tool, not a teacher; using it to bypass understanding is borrowing against future senior capacity.
| Prescription | Concrete action | Rationale |
|---|---|---|
| Structured rotation through fundamentals | Require hands-on debugging, incident response, and legacy-code work — modelled on medical residencies | Scut work builds the pattern recognition AI cannot supply |
| Shift promotion criteria from velocity to understanding | Evaluate how engineers think, not only what they ship | Masks the learning deficit created by AI-assisted output |
| Internal documentation as infrastructure | Write docs for a senior engineer who can code but doesn't know the codebase — not as AI summaries | Preserves context that lives in senior engineers' heads |
| Frame AI as a tool, not a teacher | Make explicit that using AI to bypass understanding borrows against future senior capacity | 80–90% of questions already bypass mentor-led struggle |
The hard limit: the people who can construct good AI contexts are those who already carry context from years of craft. That pipeline is breaking. Teams restructuring career ladders now to protect the formation layer — even at the cost of short-term velocity — will retain senior engineers to supervise agents a decade from now.