What about the Juniors?

[!TLDR] The danger is not AI replacing junior developers, but juniors never becoming engineers who truly understand the systems they ship.

I started working in software professionally about three and a half years ago. I however was always fascinated by how things worked under the hood, so I've been using React for nearly ten years then, and earlier played with HTML, CSS, PHP, JQuery, and a bit of Node for years.

Back then, knowing how to code was the only thing that mattered. You had to. Nobody was going to bail you out of a broken useEffect, and copy-pasting from Stack Overflow only gets you so far, before you hit a wall.
Posting a question there? Be prepared for the most toxic answers ever.

Only actual understanding could get you through this.

That was my gym. That messy, frustrating, frequently humiliating gym where the only way out was through.

I think about this a lot, because it is the question I get the most in our Agentic Engineering Foundation Schools. Some version of: "You keep saying we need to understand how things work to supervise the agent well. But if everything is generated now, where does that understanding come from? Where do juniors actually learn?"

It is a good question. I don't think the industry has settled on a good answer.

The gym is full of agents

Entry-level developer employment for the 22-25 age group dropped nearly 20% between late 2022 and mid-2025 (to be fair here, that's not AI-related only, we had post-zero-interest-rate correction, overhiring after and during Covid, layoffs post and during Covid, etc.), according to Stanford's Digital Economy Studyhttps://stackoverflow.blog/2025/12/26/ai-vs-gen-z/. Entry-level job postings fell 25% year-over-year in 2024. 54% of engineering leaders plan to hire fewer juniors, explicitly citing AI assistants that let existing seniors cover more ground. https://www.cubic.dev/blog/the-junior-developer-paradox

Marc Benioff stopped hiring new software engineers at Salesforce entirely in 2025. The reasoning was plain: AI has made the team more productive, so fewer headcount needed.

From a spreadsheet perspective the case is obvious. Why pay $90k for a junior to write boilerplate when Claude Max costs a maximum of $200 a month? When Claude does the heavy lifting? The gym is full of agents now.

AWS CEO Matt Garman has a three-part answer to that question, and he called the "replace juniors with AI" idea one of the dumbest he has ever heard.

First: junior developers are often more fluent with AI tools than their senior counterparts, not less. The Stack Overflow 2025 Developer Survey backs this up: 55.5% of early-career developers use AI tooling daily, only 47.3% for experienced developers, but there are more of them who don't even plan to ever pick up on AI.
Over half of Gen Z employees are actively helping senior colleagues upskill on AI[^2], so do I.https://survey.stackoverflow.co/2025/ai#sentiment-and-usage-ai-select-ai-sel-prof-early

Second: juniors are already the least expensive people on the team, so they are not where the real cost optimization lives anyway.

Third, and this is the one that matters most: if you stop hiring juniors, at some point the whole thing explodes on itself. There is nobody left to promote. The pipeline is gone.

The apprenticeship model is quietly breaking

There is a pattern that has defined how engineers grow from beginner to competent for decades. You join a team. You get the boring work. You write the CRUD endpoints, fix the flaky tests, bump the dependency versions.
You mess up, a more experienced colleague reviews it, you (hopefully) understand what you did wrong and the picture of how systems behave becomes real in your head rather than theoretical.

That model required two things to work: junior-level work to exist, and seniors who had time to review it and explain it.

AI is attacking both. The boring work, which was boring precisely because it needed low judgment and high repetition, is exactly what agents are good at. And as agents pick it up, seniors are spending more time validating AI output instead of mentoring humans. That goes so far, that one analysis puts the increase in senior code review time at 91%, because AI-generated PRs are 18% larger on average and need much more scrutiny. https://www.cubic.dev/blog/the-junior-developer-paradox

The gym is shrinking, and the people who got strong in it are often too busy to notice.

By skipping the boring basics, the pipeline of deep skill transfer broke for generations. The concern is that we are doing the same thing in software right now, just faster.

How you use AI is the learning that matters

You're training apprentices? You're mentoring Juniors?

Anthropic ran a randomized controlled trial with 52 junior engineers, all learning a Python async library none of them had used before.
Half got AI assistance, half did not. The AI-assisted group scored 17% lower on comprehension afterward. That does not say much about AI itself. It says a lot about how the juniors used it.

Within the AI group, the split is what matters. Developers who used AI to ask conceptual follow-up questions ("why does this work?", "what happens when this fails?") scored 65% or higher on the evaluation. Developers who just used it to generate code and move on scored below 40%. https://www.infoq.com/news/2026/02/ai-coding-skill-formation/

Same tool. Same task. Twenty-five-point gap, purely from posture.

That is not a productivity finding. That is a learning finding. And I find this very interesting, since this underlines the point I have been making about Vibecoding vs. Agentic engineering.

The supervision paradox, which Lars Faye's article on agentic coding describes cleanly, assumes you already have the skills needed to supervise the agent. https://larsfaye.com/articles/agentic-coding-is-a-trap For a junior, that is the whole problem. They are trying to build those skills. If the mechanism for building them is generating code they don't understand and shipping it past a review process that is already overloaded, they may never actually develop them.

juniors become fluent in accepting suggestions but illiterate in understanding what those suggestions actually do. https://www.cubic.dev/blog/the-junior-developer-paradox

It's not as bad everywhere

Shopify went the other direction. They didn't quietly cut their internship program in response to AI. They grew it by a factor of ten, from roughly 100 interns a year to over 1.000.

The reasoning from Farhan Thawar, Shopify's VP of Engineering, is worth taking seriously. People coming out of school now have grown up with AI the entire way through. They are AI-native in a way that experienced engineers are not. Shopify made a bet on that, the same way they once hired mobile-native engineers when smartphones took over. https://coderpad.io/blog/hiring-developers/in-the-ai-era-shopify-is-investing-in-junior-engineers-not-cutting-them/

There is also a culture argument. Interns ask "why are we doing it this way?" with genuine curiosity rather than institutional skepticism. They push on assumptions that have calcified. Thawar's framing: the wrong answer to that question is "because we've always done it this way." The right answer is either a real explanation or an honest "actually, I never thought about whether there was another way."

On how to actually evaluate technical talent with AI in the mix:

I have an answer I don't love. The honest answer is: we don't know.

What Shopify is working with is a three-mode framework.

• No AI at all: can you write code by hand and understand what's happening underneath?
• AI optional: do you know when to reach for it?
• AI mandatory: the scope is too large for the time available and the candidate has to wield the tool effectively.

All three matter. Knowing when to set the tool down is a skill. So is knowing when not to.

but it's still bad

Back to the question from the school. The concern I hear from engineering leads is not abstract.
It is: we have juniors on the team who can build a feature in an afternoon using Claude Code, but when something breaks in production, nobody really knows what the code is doing or why.

That is a mentorship problem and a culture problem before it is a tooling problem.

Whoever submits the PR owns that code. AI wrote the first draft, but the junior's name is on the review, and the senior's name is on the approval.

The redesigned junior role that actually works in the AI era is closer to auditor than author. Read the code of the Agent. Understand the code, understand its plans and understand, why a plan will not work this and that way. Steer the agent. That's also the way to learn how code actually works, and how to debug it when it doesn't.

And deliberately keep some things AI-free. Not as punishment, but as gym. Pick one area per itration, one module, one service, where the junior writes it by hand, gets stuck, asks the agent for a conceptual explanation, and tries again.

The agent does not know what it does not know. You'll need someone on the team who does. That person does not appear from nowhere.

[^1]: Stanford Digital Economy Study on developer employment, via the Stack Overflow year-end retrospective [^2]: It's actually 62%