What makes a developer valuable today: what the data and this month's debate say

You open a job posting and it asks for “experience with AI agents”, “judgment to review generated code” and “product sense”. Not a word about the framework you studied last year. At your current job, the coding agent writes half of what you used to write yourself, and you start wondering what exactly you’re being paid to do.

Over the last two weeks, data and discussions appeared that help answer that. Developer postings are growing again, but only for a certain profile; one of the companies that did most to spread algorithm interviews apologized for them; and one of the most respected programmers in the field opened a debate about what you lose when you hand your work to an AI. Below is what they say, and six concrete skills that today separate a valuable developer from the rest, each with a way to practice it this week.

What the market says

On October 11, CNBC pulled together the latest labor market data (opens in a new tab) for developers in the United States, which tends to set the trend:

• Postings rose almost 15% since late February 2025, while overall postings across all fields fell 7%, according to Indeed. Note that they’re still 27% below pre-pandemic levels.
• The rise isn’t even. 71% of the increase came from roles for experienced people, and 37% from roles with “AI” in the title.
• For people starting out, it’s harder. Employment of developers aged 22 to 25 fell almost 20%, according to Stanford University.
• Agents are already part of daily work. According to GitHub, one in three pull requests (the requests to merge a change into the main code) now goes through an AI agent. A year ago it was fewer than one in ten.

The article’s own reading is that AI isn’t eliminating demand for developers, but changing what’s asked of each one: postings talk about verifying the quality of what AI generates, fixing its errors, and integrating it into existing systems.

What’s changing in interviews

On October 12, Vivek Ravisankar, co-founder and CEO of HackerRank, one of the platforms companies use most for technical exams, published an apology (opens in a new tab). He said his company helped normalize a kind of interview that became “months of grinding questions with little connection to the actual job”: LeetCode-style algorithm problems.

What he announced instead is more interesting than the apology. Assessments will move to tasks on a real repository, with an AI agent available, and what gets measured is no longer just whether the code works, but how you use AI and whether you can tell when what it generated is wrong. There was plenty of skepticism in the discussion, rightly so: this is a company that sells exams and is changing its product. But the direction matches the job postings: what gets evaluated is starting to look like the real job.

This week’s debate: tool or automation?

A day earlier, on October 11, Rich Hickey, creator of the Clojure language and one of the most listened-to voices on software design, published an essay titled “‘AI’ is Not a Tool” (opens in a new tab). His argument: a tool is something you wield yourself, with control and seeing the result as you work, like a hammer or a text editor. With a language model, instead, you ask for something and the model does it on its own. So, he says, it isn’t a tool you write code with, but “automation with which you avoid writing code”. And what worries him most is what gets lost: what you learn while wielding a tool is something about the world, and that doesn’t happen if someone else did the work.

The next day, a programmer who goes by Huahai replied with an essay (opens in a new tab) that was one of the most discussed of the week among developers, with over five hundred comments on Hacker News. It accepts part of the diagnosis, but not the conclusion: a model can be wielded, and wielding it well is a skill. The line that sums it up: “Tools can amplify poor judgment as readily as good judgment.”

Notice what the two have in common. Neither says AI is useless, and neither says using it is enough. Both say that if you hand off the work without understanding it, you lose. That’s the answer to what makes you valuable: not knowing how to use AI, which anyone learns in an afternoon, but having the judgment to direct it and to notice when it’s wrong.

Six skills that are worth more today

1. Reading and reviewing code you didn’t write

If a third of the code already goes through an agent, someone has to read it. A Harvard study discussed this same week showed that with agents people write 30% more code, but review has become the bottleneck. The person who quickly finds the bug hidden in a two-hundred-line change is the one who unblocks the team.

To practice it: next time the agent hands you a change, don’t accept it until you can explain every file it touched. This list tells you what to look for.

2. Explaining clearly what you want

An agent does what it understood, not what you meant. Whoever writes a clear request (what has to be done, what must not be touched, and how you know it’s finished) gets a result that works on the first try. It’s the same skill it always took to work with other people, but now you use it twenty times a day.

To practice it: before your next task with the agent, write five lines covering those three things. This article has a template.

3. Verifying, not trusting

HackerRank’s criteria and the job postings agree: knowing whether what the AI generated is right. That isn’t done by looking at the code with good intentions, but with tests. A few days ago, someone had an agent rewrite the entire TypeScript compiler without reading the code, and it worked because there were 184,749 tests saying what was right.

To practice it: on your next task, write the test before asking the agent for the code. If you don’t know what test to write, you don’t yet understand what you’re asking for.

4. Understanding the whole system, not just your part

An agent writes an isolated function very well. What it doesn’t see is that the function is called inside a loop and ends up making three hundred database queries for every page. Knowing how the pieces connect (the database, the network, the costs, what happens in production) is what lets you catch that kind of problem before it reaches users.

To practice it: pick one screen of your app and follow a request end to end: what queries it runs, which services it calls, and how long each step takes.

5. Knowing the business, not just the code

An agent doesn’t know your company bills differently at month end, or that a big customer depends on a field that looks unused. That information isn’t in the code; you know it because you talked to people. The easier it is to write code, the more it’s worth knowing what code to write.

To practice it: on your next task, ask why it exists before you start. The answer often changes what needs to be done.

6. Keep learning, even though AI could do it for you

This is Hickey’s point, and the easiest one to lose. If the agent solves everything new that comes your way, a year from now you’ll know the same as today, and your judgment for reviewing its work won’t grow.

To practice it: when you have to use a library or a concept you don’t know, write the first version yourself, even if it takes longer. Then ask the agent for its version and compare. That difference is exactly what you’re learning.

If you’re just starting out

The data doesn’t hide that the way in has narrowed. But what companies are looking for is also clearer than before: someone who uses AI with judgment, not someone it replaces. A personal project that shows how you reviewed what the agent did (what you corrected, what tests you wrote, why you made each decision) shows that better than a hundred solved LeetCode problems.

Writing code is no longer what’s scarce

For years, a developer’s value was measured by how much code they could write and how fast. Today an agent does that in minutes. What’s still scarce is knowing what to build, noticing when something is wrong, and understanding enough to direct the work. Those skills don’t go stale with the next model release, and that’s why they’re the ones worth practicing.