I spent the first chunk of my career doing the work nobody else wanted.
The rubbish tickets. The repetitive fixes. The small jobs that took ages, needed almost no imagination and somehow always landed on the junior engineer’s desk. I knew exactly where I sat in the pecking order, and I can’t pretend I enjoyed the view.
But I’ve been wondering lately whether I’d be the same professional if someone had taken those first five to seven years away from me.
Not because the tickets mattered. Some were spectacularly dull. What mattered was everything happening around them. You made mistakes. You watched better engineers work through problems. You learnt when to ask for help, and occasionally discovered that the thing you’d been certain about for three hours was wrong from the first line.
You also learnt to turn up and do decent work when the work wasn’t interesting, which is a skill nobody puts on a CV and every employer quietly depends on.
Now we’re building agents that can do most of that work.
And I can’t decide whether we’re about to hand the next generation an extraordinary head start or quietly remove the ladder they were meant to climb.
The apprenticeship nobody called an apprenticeship
When I was at university, computer science was a sensible choice if you liked technology and wanted a decent job. You got the degree, joined a company as a junior developer, spent a few years getting things wrong under the supervision of people who had already got those things wrong, and gradually became useful.
Nobody expected a graduate to operate like a senior engineer on their first Monday. Companies paid juniors to learn while contributing something, and the industry got experienced engineers out the other end. It was an apprenticeship in everything but name, funded by the backlog.
In October, Anthropic published numbers suggesting the backlog has stopped paying for it. In its own technical departments, juniors fell from 21% of staff to 14% since early 2025. The junior share of technical hires fell faster, from 27% to 9%. Managers said they were putting more weight on autonomy and judgement, and less on domain expertise.
That’s one company, and a slightly unusual one, given it builds the tools doing the displacing. The researchers are careful to call the trend descriptive rather than causal. But it isn’t an isolated data point. Stanford researchers found in 2025 that employment for 22 to 25 year olds in the most AI-exposed occupations had fallen about 13% relative to less exposed work, while older workers in the same jobs were fine.
The question doesn’t depend on one statistic anyway. If a senior engineer with a set of agents can do the work that used to occupy several juniors, what commercial reason does a company have to hire the juniors?
I know the answer we’re supposed to give. We’ll free young engineers from drudgery so they can tackle more interesting problems. They’ll learn faster, build bigger things, and everybody wins.
I can see it. If I’d had today’s tools at 22, I’d probably have progressed faster. Less time wrestling with trivial bugs, more time understanding why systems behave the way they do.
But that assumes somebody would have hired me.
Which is the bit the optimistic version tends to skip.
Would I hire them?
Here’s the uncomfortable part. If I ran a business with 50 engineers and found I could get much the same output from 15 experienced people and a decent set of agents, I wouldn’t keep recruiting traditional junior developers out of respect for tradition.
I’d still hire young people. I’d want their energy, their curiosity and their willingness to question how things have always been done. There’s real value in a smart graduate who hasn’t yet learnt which ideas they’re supposed to consider impossible.
But I’d hire them into a different job. Not three years clearing the backlog, because the agents would already be doing that. Something closer to directing agents, reviewing what they produce and owning progressively harder outcomes, with an experienced engineer close enough to catch the mistakes before production does.
Which makes me part of the problem. I’ve spent the last six months building exactly the kind of agent workforce that changes this calculation. I’ve seen the productivity gains and I expect them to grow. Asking companies to preserve inefficient work as a training scheme would be an odd position for me to take, and I’m not going to take it.
Still, there’s a difference between making someone more productive once they’re inside a profession and removing the economic reason to let them in.
We keep treating those as the same thing.
Why this isn’t just another Stack Overflow
The obvious objection is that we’ve been here before. Frameworks, better IDEs, Stack Overflow and cloud platforms all automated chunks of junior work, and the industry kept hiring graduates.
The difference is who the tools helped. Those made the junior faster. Agents increasingly do the junior’s work outright, and the person they make faster is the senior engineer who used to need the junior.
One shrinks the learning curve. The other shrinks the number of people allowed onto it.
The other side of my own argument
I wrote recently that AI is coming for the middle of the org chart, and that younger people have an enormous opportunity because frontier models give them capabilities that used to take years to build.
I still believe that. A 22-year-old with good instincts and the right tools can research, prototype and test ideas well above their pay grade, and turn up to a meeting better prepared than plenty of people who outrank them.
But a young person can be more capable than ever and still find nobody wants to employ them. We’ve made the individual more productive while making the entry-level job less commercially attractive. That’s a peculiar result, and not one the market looks keen to correct on its own.
Nor is the answer to tell every graduate to start a company. Some will. Most won’t want to, and that’s fine. An industry can’t sensibly rely on everyone becoming a founder just to get their first five years of experience.
So who trains the people who’ll eventually replace today’s senior engineers?
The senior engineer problem of 2035
Good senior engineers aren’t just people who write better code. They have judgement. They’ve watched changes fail in production, worked under ugly constraints, seen promising ideas fall apart on contact with reality and learnt which apparently minor details deserve worrying about.
Most of that judgement comes from consequence. From owning something, getting it wrong and having to put it right while someone waits.
An AI assistant can explain a mistake brilliantly. That isn’t the same as having made it while somebody’s system was down and the clock was running.
Perhaps we’ll build better apprenticeships. Small teams could take graduates who learn by directing agents, reviewing their output and owning increasingly difficult outcomes. Universities might shift from teaching people to write every line towards teaching them to understand systems, interrogate outputs and exercise judgement. That could be a better education than the one I had.
But somebody has to design it, fund it and give those graduates real responsibility. It won’t happen because the technology exists. Companies, reasonably enough, optimise for the work they need done this quarter, not for the senior engineers the industry will need in ten years. Anthropic’s own researchers make a version of the point: policies that protect current employees do nothing for the people who were never hired.
What I’d tell a student
When someone asks whether I’d recommend software engineering to a young person today, my first, slightly unhelpful answer is electrician. Possibly plumber. Nobody’s agent is rewiring a fuse box yet.
I’m joking. Mostly.
I’d absolutely encourage a young person to understand technology properly. Build things. Learn how software actually works. Form a technical point of view and use these tools as aggressively as you can.
What I wouldn’t do is sell them the old promise that a computer science degree leads to a junior developer job and a predictable climb through the ranks. That route is getting much less certain, and pretending otherwise does students no favours.
For teams working out what that learning should look like, my guide to AI training for employees covers the practical side.
I might be wrong about the speed. I’d quite like to be wrong about some of the consequences.
But the question is no longer whether AI can do the grunt work I spent years learning from. Increasingly, it can. The question is what replaces that experience once the job that provided it stops existing.
An industry that stops training juniors can enjoy a very productive few years. Then, somewhere around 2035, it’ll go looking for senior engineers and discover it forgot to make any.
Sources
