What Should I Learn in the AI Era? — Everything You're Curious About, Asked to AI Agents on Loop
More often now, people who have started building working web services on their own by combining Claude Code, OpenAI Codex, and Cursor ask me: “what should I learn in the AI era?”
“On their own” here means front end, back end, light authentication, and data persistence all wired up end to end, standing up in a few days without a dedicated engineer to lean on. Work that would have taken several engineer-months a few years ago is now something one person can run through by orchestrating AI agents.
My answer is: everything you get curious about. Whenever something (a technology, a concept, a mechanism) catches your attention, turn it into a question to an AI agent on the spot, and keep chasing the new terms that appear in the reply. This post explains why I keep giving that answer and how the approach only holds together because AI agents are the ones answering on the other side.
The Background Behind the Question
The people asking me this are usually not classically trained engineers, but they have reached a state where Claude Code or Codex can carry their ideas all the way to running code. Business-side teammates, designers, and domain experts are writing code through an AI agent for themselves.
They can ship something that runs, but they stall on the next question: where to invest their learning time. Should they study a framework end to end, learn the computer-science fundamentals, go deep on one language? The branches are many. On top of that, since an AI agent will explain any framework or fundamental concept on demand, even the motivation for sitting down with a textbook is harder to construct than it used to be.
The Answer: Ask an AI Agent About Everything That Catches Your Attention
What I actually recommend is not deciding the scope of study up front. While you are already interacting with an AI agent, pick up every word, concept, error message, or citation in an answer that catches your attention, and turn it into a follow-up question right there.
For example: you deploy a UI you built with Next.js and get curious about the Cache-Control header; you notice infer inside a TypeScript type definition and want to understand what it does; you see CVE-2025-XXXXX on a PR and want to know how CVE numbering works. Every level of tiny snag counts. Don’t stop after one round trip: dig deeper into whichever new term appeared in the previous answer, and keep going vertically.
People often push back that not scoping this will just diverge. I think divergence is exactly what you want here. Where you spend your time once you diverge is what surfaces your inclination.
Why This Works Only Because the Answerer Is an AI Agent
“Ask about everything that catches your attention” only becomes viable because there is effectively an unlimited answerer on the other side. Asking a human colleague or mentor about every small thing burns their time; books and official documentation don’t reshape their explanation to your current level of understanding.
An AI agent will re-explain, at whatever granularity you ask for, taking your prior turns as context. You can point at the part you didn’t follow, ask for a different analogy, have it write example code, run it locally, and ask again about the behavior you observed. The iteration cost dropped enough that you can do this without worrying about the other side’s time. That’s the property that makes “no fixed scope” a realistic learning approach.
The other prerequisite is not trusting the AI agent’s answer on its own. You have to build in the verification step. Cross-check against official documentation (Next.js documentation, MDN Web Docs), actual source code, and actual execution output. Without that, what you’ve “learned” plateaus at “what an AI agent told me,” which is not the same as knowing it.
How to Chase What Catches Your Attention
Three things I keep in mind when I do this.
First, don’t defer the question. The moment you tell yourself “I’ll look that up later,” you almost always don’t. Spending five to ten minutes right there to get even a rough feel for the answer is enough. Once you have the rough feel, the term becomes a search key inside your head for later.
Second, keep asking about the new terms that show up in the answer. If the reply is “that’s about HTTP caching’s stale-while-revalidate,” don’t stop there. Follow up with “what does stale-while-revalidate actually do, when does it apply, and what side effects come with it?” Without that follow-up, you end up knowing the name of a thing but not the thing itself.
Third, run whatever can be run locally. Take the example code the AI agent gave you, run it as-is, tweak it, watch the behavior change, and ask about why the change had that effect. Some topics you can learn by reading; others only click once you’ve watched them run. For the latter, running them yourself gives you far more per unit of time.
The Topics You Keep Circling Back to Are Your Inclination
If you let yourself diverge like this for a while, a pattern shows up: there are areas you keep spending disproportionate time on. The finer details of UI/UX, authentication and authorization design, performance measurement, database schema design, team process and delivery workflow, or business-side decision-making.
The topics that survive after chasing countless small curiosities are, I think, a more honest signal of your inclination than an externally assigned career path or a currently fashionable role. Once you can put your inclination into words, the next investments (which textbooks to read cover to cover, which communities to join, which engineers’ feeds to follow, which roles to apply for) narrow down.
Engineering role divisions are reshuffling for a while because of AI agents. Measuring your own interests only against the role definitions that were solidified five years ago (front-end, back-end, SRE as distinct silos) feels cramped. Using the topics that show up in your own hands-on work as the axis for your inclination fits the current environment better.
After Your Inclination Comes into Focus
Once your inclination has words attached to it, the way you learn shifts. My own sequence tends to be: read the official documentation end to end, work through one or two canonical books in the area, keep following engineers who publish primary-source material on it, and rewrite what I’ve learned in my own words in a place like this site CodeNote.net.
At this stage the role of an AI agent changes. It becomes a sparring partner for chewing through primary sources faster, rather than the main source itself. The purpose has shifted from “find where my interests live” to “become someone who can make reliable judgment calls inside that area,” so the shape of the question changes too, from asking for a good analogy to asking the agent to point at the exact spec or implementation the claim rests on.
Once your inclination is set, narrowing the scope and going deep is the move that pays off. But to get there, going through the “chase everything that catches your attention” phase first is what leaves you able to answer “where should I invest?” without needing someone else to tell you.
That’s all from why I keep telling people “chase every small curiosity through an AI agent, and let the topics you keep returning to become your stated inclination,” from the Gemba.