For more than a decade, the standard advice to anyone worried about the future of work was the same: learn to code. It was treated as the one safe bet, the skill that would keep you employable no matter what changed. That advice is now pointing in the wrong direction.
The person saying so most plainly is Dario Amodei, the chief executive of Anthropic, the company behind the AI assistant Claude. In recent interviews, Amodei has argued that writing code will be among the first kinds of work that AI takes over — not eventually, but soon, and at scale. He draws a careful distinction. Coding, the act of writing lines of code, is not the same as software engineering. Designing systems, understanding what real users actually need, and managing teams of AI agents — that work, he argues, will last much longer.
His advice to young people starting out is correspondingly direct: learn to use AI as quickly as possible. He has also warned that as many as half of entry-level white-collar jobs could be disrupted within one to five years. The point is not to frighten anyone. It is that the ground is moving under the entry-level job itself, and the safe harbor is not a credential in programming. It is the ability to get work done with AI.
The skill is harnessing, not coding
What does that ability actually consist of? It is not a technical specialty. It is the skill of taking a task you already do — in your field, whatever it is — and figuring out how to get an AI agent to do part of it, well, under your direction. An agent, in plain terms, is software that can carry out a multi-step task on your behalf: search, read, sort, draft, check, and report back, while you supervise. A researcher who uses agents to search, sort, and summarize thousands of pages. A teacher who automates grading rubrics and drafts lesson variations. A small business owner who sets up an agent to track invoices, answer routine inquiries, and flag the ones that need a human. A paralegal who has an agent pull the relevant passages from a stack of filings. None of these people needs to become a software developer. They need to know what they want, to describe it clearly, to check what comes back, and to correct it.
That is a different kind of competence than the one our schools and training programs were built to teach. It is not about memorizing syntax or passing a certification exam. It is about judgment: knowing what a good result looks like, spotting when the output is wrong or shallow, and having the patience to iterate. The agent does the execution. You do the thinking. And the thinking is the part that cannot be automated away, because it is the part that decides what the work is for.
The real requirement is discomfort
If there is a barrier to this skill, it is not technical. It is psychological. Working with AI means being comfortable with being uncomfortable — starting before you fully understand the tool, trying something, watching it fail, adjusting, and trying again. Most adults are trained out of that posture. We are trained to be competent before we act, to know the answer before we speak, to avoid looking foolish. AI work rewards the opposite: a willingness to not know things, to ask naive questions, to treat confusion as the starting point rather than a sign of failure.
This is worth saying plainly because it is the part nobody sells. The vendors sell ease — press a button, get a result. The reality is messier and more interesting. The people who will do well with AI are not the ones who master a particular interface. They are the ones who can tolerate ambiguity, who can hold a problem in mind while the tool does its work, and who can tell the difference between a plausible answer and a correct one. That is critical thinking, applied to a new medium. Amodei has made the same point: as AI takes over more of the execution, the scarce resource becomes the human judgment that steers it.
There is also a practical side to the discomfort. When you hand a task to an agent, you are giving up direct control over how it gets done, and you have to build the habit of verifying what comes back. That habit — checking the sources, reading the output, testing the result against what you know — is the discipline that keeps the human in charge. Without it, the convenience of AI becomes a trap: fast output, unchecked, gradually replacing your own judgment. The skill is not just using the tool. It is staying awake while you use it.
Platform agnostic, by design
One more thing matters, and it matters a great deal: this skill is platform agnostic. It does not belong to Anthropic, or OpenAI, or Google, or any other company. The tools will change — they already change every few months — and any skill tied to one vendor’s interface has a short shelf life. What transfers is the underlying ability: breaking a task into steps, giving clear instructions, evaluating output, building small automations, and managing agents the way you would manage any team. Learn that, and you can pick up whatever tool is in front of you. Anchor yourself to one platform, and you will have to start over every time the market moves.
There is also a practical reason to keep it platform agnostic. Your work is yours. The records, the methods, the judgment you build up over years should not live inside a single company’s product. The harness — the way you direct AI to serve your purposes — should be portable, because your career will outlast any particular tool.
So the advice has changed. Not “learn to code.” Learn to harness. Start with one task you already do — the repetitive one, the one that eats your afternoon — and automate part of it. You will not know what you are doing at first. That is the point. The discomfort is not a sign that you are behind. It is the skill itself.
