Who Trains the Junior When the Agent Does the Grunt Work?

A couple of years back I ran two intern cohorts, first-year PHP developers, one in 2023 and one in 2024. The way they got good was completely predictable. You handed them the unglamorous work, the small endpoints, the form validation, the fixes nobody fought over, and each week they got a little sharper without anyone formally teaching them.

I thought about those cohorts last month, when I handed a junior a first-year task and caught myself halfway through the sentence. The agent could produce all of it before he had finished reading the ticket. And I sat with an uncomfortable question that the intern years had never made me ask: if the machine does the work I would normally give a beginner, how does the beginner ever become anything else?

The grunt work was never valuable because it was hard. It was valuable because it was how you learned. You wrote the boring thing, got it a little wrong, watched it break, and something settled into place that no amount of reading could teach. The tedium was doing the teaching, quietly, the whole time.

Nobody designed it this way, but every craft used to have a bottom rung made of repetition. You did the simple thing badly, then less badly, then without thinking, and by the time it was automatic you had built the judgment that lets you tell good from bad. The junior who wired up a hundred forms is the senior who can glance at the hundred-and-first and know it is wrong.

AI is very good at exactly that bottom rung, and it is pulling the ladder up behind it. The work that used to build a beginner's instincts is now the work most easily handed to a machine. What is left for the junior is not the doing. It is being handed the machine's output and asked to judge it, which is a different and much harder skill.

That is the trap in miniature. Reviewing well is a senior ability. You are asking someone to catch mistakes in work they could not yet have produced themselves, before they have built the pattern that makes the mistakes visible. The doing came first for a reason. It was how the judgment got made.

I watched this land in a review not long ago. The junior approved an agent's change, and it was mostly fine, but he had missed how it handled an empty result, the kind of thing you only learn to distrust once it has burned you. He had not been burned by it yet, because the tasks that used to do the burning are now done by something that never learns and never teaches.

He was not careless. He simply had not been given the years that make you careful. And that is the part that worries me about a team built on prompting and skimming, with nobody underneath who came up the hard way. A team like that has no one who can catch the machine when it is confidently wrong, because catching it is a skill you only earn by having done the work yourself.

There is a longer-term worry underneath the daily one. Seniors are not minted, they are grown, slowly, out of juniors who did a lot of ordinary work. If the ordinary work stops happening, the supply of people who can actually judge the machine's output starts to thin a few years out, right when we will need them most. The shortcut that saves a week now may be quietly mortgaging the next generation of judgment.

The old apprenticeship never had to be planned. Learning happened as a by-product of doing, so you got it for free. Remove the doing and the free learning goes with it, and you are left needing to build the experience on purpose, which almost nobody is set up to do.

So I think the job quietly gains a duty rather than losing one. If you want a junior to develop judgment now, you have to construct the experience deliberately: give them the agent's work and sit with them through where it is subtly wrong, walk them through the reasoning a burned senior would apply, make them predict what will break before you show them.

It is slower. It looks less productive on any dashboard, because the fast path is to let the agent do it and move on. And it is probably the most important thing a lead can do right now, because the alternative is a team that can operate the tools but has nobody who can tell when the tools are lying.

None of that is a course you can send someone on. It is the patient, unglamorous work of turning a machine's output back into a teaching moment, over and over, until the instincts form the way they used to form on their own.

So here is what I keep turning over, and I would genuinely like to know how others are handling it: on your team, who is teaching the thing that used to be taught by doing, now that the doing is gone?