Issue #27: The Apprenticeship Layer Is Being Absorbed

A ticket comes in, a junior takes a day and ships something half-working, a senior explains what went wrong. Describe the same ticket to Claude and it merges before lunch. That trade wins on every metric a team reports this quarter, and it also quietly removes the only training that junior was going to get. Researchers at Seoul National University interviewed 14 engineers in South Korea about exactly this decision, and the results are stark: one founder who cut his junior headcount said flatly there was 'no impact whatsoever,' and when asked what a junior could do better than a Claude subscription answered, 'I don't think there's anything.'

The same founder, earlier in the same interview, described how he became competent: years of what he called 'development grunt work' — writing code as told, debugging daily, testing whatever came back. The mechanism that produced him is the mechanism his company no longer runs. Korean IT job postings dropped 43 percent between 2023 and 2024. In the United States, programmer employment declined 27.5 percent over roughly the same stretch.

The uncomfortable reframe in the study is that minor bug fixes and routine implementation were never assigned to juniors because the work was valuable. It was assigned because doing it badly, and being corrected, is what produces judgment. A senior at a mid-sized game company put the operational consequence in one line: 'The cost of code generation is so cheap by using GenAI, but the cost of code verification is so expensive.' Verification is the scarce input now, and the pipeline that produced verifiers has been switched off.

The students in the Seoul study describe the deficit in the same language their seniors use. One said he does not know what not to do, having seen only the good examples. Another said, 'I don't know what I don't know.' A senior at a re-commerce company arrived at the same conclusion from the hiring side: 'The reason someone is called a senior is that they know how work should not be done, but now juniors can't learn from the bad examples.'

The classroom half of the pipeline is not compensating. One student described giving up on unaided assignments because 'everyone except me was using GPT, so everyone except me was getting nearly perfect scores.' The same student took two courses in the same field, one that permitted chatbot use and one that did not, and reported identical grades: 'I definitely studied much more in the machine learning course. The deep learning course just felt like I got better at using GPT.' The transcript looks the same either way, which is the part that should worry anyone reading resumes.

Engineers hiring in 2026 are inheriting the graduates this system produced. The failure mode is not incompetence — it is confident output with no felt sense of when the output is wrong.

One finding in the Seoul study complicates the doom narrative and deserves more attention than it has received. A senior at a large enterprise with a formal onboarding pipeline reported that junior work had changed very little. Absorption ran furthest, the researchers noted, where nobody had written down how juniors were supposed to develop. The collapse being documented is not really a collapse of junior work; it is a collapse of undocumented mentorship.

One re-commerce team's answer costs almost nothing to copy. Newcomers begin with a task engineered to be trivial — finding a missing period in a production project, committing the fix, opening a pull request, getting reviewed, and pushing to production. Later assignments arrive with less scaffolding, until the newcomer is asked to find a problem and solve it without being told how. The engineer describing it calls the design 'a kind of learning by doing.'

The implication for engineering leaders is direct. If the ladder inside your company lives only in the heads of the seniors who happen to be around this quarter, the agents will absorb it. If it lives in a document, a checklist, and a graduated task list, it survives. Two adjacent industries already made this choice. FAA Safety Alert for Operators 13002 tells carriers to preserve opportunities for manual flying so pilots retain skills the autopilot performs. Nuclear plant operators face a parallel requirement under 10 CFR 55.59, mandating periodic simulator training on emergencies that routine automated operation never exercises. Software has no equivalent and has not asked whether it wants one.

At QCon, Tessi co-founder and CTO Ben Greene framed the current moment as what 'startups can teach us about surviving the coding agent apocalypse,' and offered a mindset rather than a tool list. His first rule is to start simple, grounded in what he called Gall's Law: 'A complex system that works is invariably found to evolve from a simple system that worked. A complex system designed from scratch never works and cannot be patched to make it work.' Agents ratchet the pressure to produce up to 11 while being, in his words, 'the worst offending enablers imaginable.'

His second rule maps directly onto the verification problem the Seoul researchers identified. 'Comprehension is key,' Greene argued, because change is inevitable and the linchpin of change is truly understanding what you have already built. His description of agent-generated code is worth quoting: 'It's like hiring a team that copies code off the internet at the speed of sound and always tells you exactly what you want to hear. It terrifies me.' The failure mode is not bad code; it is confident code you did not read, in a system you cannot now modify without heroics.

Greene's practical stance — treat the agent as an over-eager contractor whose output you own, refuse to ship code you do not comprehend, keep the initial system small enough to hold in your head — is what verification looks like as a daily habit rather than a virtue. It is also, notably, teachable. Which is the point.

A useful sanity check on the 'coders are finished' story comes from India's labour data, which shows the salary premium migrating rather than disappearing. TeamLease's Jobs and Salaries Primer FY2026-27 reports that the biggest projected raises are no longer flowing to traditional software developers. Electrical Engineers lead at 11.2 percent, followed by Quality Control Inspectors at 10.9 percent, Project Engineers at 10.7 percent, System Administrators at 10.4 percent, and IT Support Executives at 10.3 percent. Associate Software Engineers come in at 9.7 percent — healthy, but below the roles that install, maintain and verify the systems around the code.

The report's framing is worth taking seriously for engineers thinking about where to invest attention. 'Every AI application ultimately depends on physical infrastructure — servers, chips, factories, networks, power systems and industrial automation,' it notes, and employers are increasingly rewarding professionals 'who keep increasingly digital businesses running.' System Administrators and IT Support Executives appear among the fastest-growing roles across EV infrastructure, healthcare, power, banking, logistics and telecommunications.

The pattern is consistent with the Seoul diagnosis. When code generation gets cheap, the scarce input becomes the judgment to verify, integrate and operate what the generators produce. Markets do not have to know the theory to price it. India's shift is the same signal read from the payroll side: the premium is moving toward people who can tell whether the system actually works.

If verification is the new scarce skill, and if nobody is being handed the work that produces it, then the question for individual engineers is not which agent to adopt next. It is which deliberate practice you are willing to schedule now, on your own initiative, to remain the person who can tell when Claude is wrong.

A workable starting list, drawn from the sources above: read every line of agent-generated code before merging it, and refuse to ship what you cannot explain to a colleague. Keep one project — however small — where you write everything by hand, the way pilots are required to hand-fly. If you lead a team, write down the onboarding path that currently lives in your head; the Seoul study's one exempt company is the one that did. If you are early-career, seek out the graduated, real-stakes tasks the re-commerce team described, and treat any employer who does not offer them as a place where your ceiling is already set.

The apprenticeship layer is being absorbed on the quarterly-metrics timescale and the bill arrives on the career timescale. That gap is where individual choices still matter. The engineers who thrive across the next few years will be the ones who noticed the gap and put something on the calendar.