The Rung That Got Removed.

What was actually doing the teaching

It is worth being precise about the mechanism, because the loose version of this argument ("juniors learn by writing code") is easy to dismiss and slightly wrong. Writing code was not the teacher. The teacher was a specific loop: a junior committed to a decision, a senior contested that decision, and the junior had to either defend it or rebuild with the objection absorbed. What transferred in that exchange was not syntax. It was the senior's model of which tradeoffs matter, delivered at the exact moment the junior had skin in the outcome.

That loop needed the junior to have made the decision. Not to have typed it, made it. Everything that makes review pedagogically expensive for the senior and valuable for the junior comes from that one property, and it is the property that generated code removes, silently and completely.

Inspection has a ceiling, and it is lower than it looks

Training juniors to review generated code is the obvious response and it is a good one as far as it goes. Reviewing is a real skill, it is increasingly the job, and someone who is excellent at it is valuable now. The limit is specific rather than general: an inspector who has never built the thing eventually stops being able to see a whole class of defect, because the defects that matter most are the ones that are only visible if you know what the alternative would have been.

A missing index, a retry that will stampede, a schema that forecloses next year's feature: none of these look wrong on the page. They look like ordinary code. They register as wrong to someone who once made the opposite choice and lived with the consequence. That is the part inspection cannot teach, and it is the part that separates a reviewer from an engineer.

Why this is a CTO problem, not a career-advice problem

The version of this discussion aimed at juniors — should I worry, what should I study — is a different question with a different answer, and this site covers it separately in will AI replace junior developers. The organisational question is the one with nobody's name on it.

Consider the timing. A hiring pause on juniors produces no visible harm in year one, mild benefit in years two and three, and a shortage in year six that cannot be resolved by paying more, because every peer organisation made the same decision in the same eighteen months and is now bidding against you for a cohort that was never trained. Individually rational, collectively expensive, and structurally invisible until it is not. Meanwhile the same pressure is running inside the organisation: capability is built by doing, not by hearing about it, and the doing is what got automated.

A separate cost lands sooner. If the promotion signal used to be "built something hard alone" and everyone can now produce output alone, the signal stops discriminating. Organisations that do not replace it will promote on output volume by default, which is the one metric generated code inflates for free.

Four moves that keep the rung

Each of these costs velocity. That is not an objection to them, it is the point: the apprenticeship was always paid for out of velocity, invisibly, and the only thing that changed is that the invoice now has to be signed deliberately.

Reserve a class of work as human-authored. Chosen because it teaches, not because AI cannot do it. The second criterion is the trap: it shrinks every quarter and encodes the assumption that the point was output rather than learning.

Make the junior commit before the agent does. Have them state the design decision and defend it verbally, then let the tool implement it. This restores the one property that made review teach — the reasoning under critique is theirs — and it costs about fifteen minutes.

Review the judgment, not only the diff. The senior's question stops being "is this code correct" and becomes "why did you accept this, and what did you check". That question is answerable only by someone who thought about it, so it also detects when nobody did.

Measure the pipeline as an output. How many people moved a level this year, and how many seniors did the organisation grow rather than hire. An untracked number is an unmanaged one, and this is the first thing surrendered under delivery pressure precisely because its absence is not felt for years.