"AI is coming for junior devs first." I think that's backwards.

A junior developer argues that AI is not coming for junior developers first, but rather transforming the role into one focused on judgment, verification, and taste. The developer shares practical habits, such as maintaining an instruction file for coding agents and skeptically reviewing agent output, to help juniors thrive in an AI-assisted workflow.

You hear it at every meetup and under every thinkpiece: AI is coming for the junior developers first. The logic is always the same. Juniors do the simple work, agents do the simple work now, so there's no room left for juniors. I'm a junior developer, and I run coding agents every working day on real tickets. I think that logic has the facts right and the conclusion wrong. Here's the part that's true. If you treat the agent as a faster typist and compete with it on typing, you lose. It is faster than you, and that isn't close. The old junior job, the one that was mostly typing with a learning problem attached, really is shrinking. Here's the part people miss. The moment you stop competing with the agent and start driving it, the job turns into judgment, verification, and taste. Those are learnable, and a junior who practises them on real work builds them fast, because every task the agent does can come with its reasoning attached if you insist on it. You get a running explanation of the codebase, all day, at whatever pace you can absorb. Two concrete things that moved the needle for me. A fresh session starts from zero. It doesn't know your stack, your conventions, or that you're still learning and want the reasoning spelled out. So I stopped re-typing that every morning and put it in one file the agent reads before every session: How I work - I'm still leveling up. When you explain something, show the reasoning, not just the answer, I want the why. - Capture fast, sort later. New notes hit an inbox folder. - This is work only. Personal stuff lives in a separate place. That's a trimmed piece of mine. The rule I use to decide what goes in it: if I type the same instruction into a chat twice, it moves into the file. Once might be a fluke. Twice means I'll be typing it forever otherwise. My favourite cautionary tale from my own notes. I had an agent generating summaries of some numbers. One line read, roughly, "180 plus 95 plus 60 across the three categories, 310 total." Read that again. It's 335. The agent wrote it with complete confidence, and I pasted it into my own notes without checking, because it looked right. That's the trap. Agent output always looks right, whether it is or not. Well written and correct used to travel together. With an agent they don't, so if you don't check the numbers, nobody has. The skill nobody warned me would become the core of the job is reading work you didn't write, sceptically. Neither of these is clever. That's sort of the point. The junior who gets good at the unclever, unglamorous parts feeding the agent context, checking its output, writing down what you learn ends up in a strong spot, not a threatened one. I wrote a short book about the whole setup: the instruction file, a plain-markdown "second brain" the agent maintains, durable memory, hooks, multi-agent audits, and the guardrails that stop it doing something you can't undo. It's called The Agentic Junior, written from the junior's side, with the failures left in. Link in my profile / below if you want it. Either way, the two habits above cost nothing to try this week.