The Judgment Was Always the Job

Himanshu Sharma, writing on X, on the fear that junior developers who lean on AI will never develop into senior engineers:

A junior engineer can vibe code (or even code manually) and still stay junior for years if they never learn how to reason about tradeoffs + debug production issues + improve code health over time. A senior or more experienced engineer, on the other hand, can use AI heavily and still be very much senior because the seniority is more about taste, judgment, and responsibility.

This is right, and it's the part that most of the "AI is deskilling an entire generation" discourse skips past. Seniority was never about typing speed or memorising API signatures. It's about knowing which problem to solve, which shortcut will cost you in six months, and when the "elegant" solution is actually the dangerous one.

Sharma proposes a middle path he calls "lazy coding" — write a rough version yourself, hand the mechanical parts to AI, keep your brain on the architecture. Specific advice follows: read the diff, debug manually sometimes, review AI output six times out of ten. All sensible.

Here's the problem. The habits that actually build engineering judgment — tracing a bug through three layers of abstraction, sitting with a bad data model long enough to understand why it's bad, feeling the real consequences of a wrong architectural call — are learned through friction. Not through voluntarily opting into friction on a self-imposed schedule. "Review it 6 out of 10 times" is a New Year's resolution, not a development methodology. Under deadline pressure, with plausible-looking AI output on screen, that number quietly becomes 2 out of 10, then 0.

The article also treats seniority as an almost entirely technical concept: taste, systems thinking, debugging instinct. But in practice, senior engineering is at least half communication, scope management, navigating ambiguity, and — critically — knowing which problem not to solve. AI tooling doesn't threaten any of that because AI tooling doesn't touch any of that.

The real question isn't whether individuals can still grow into senior engineers while using AI. Of course they can. The question is whether the default environment — ship fast, AI handles it, looks like it works — still produces enough friction to make that growth happen without deliberate, structural intervention from engineering organisations. Individual discipline is great. Organisational incentives are what actually scale.

The manual effort in building software is changing. The judgment effort isn't changing at all. That was always the actual job.