AI Was Supposed to Kill Dev Jobs. The Postings Went Up.

For two years the story about software careers had exactly one shape: AI writes the code, so the code writers are next. It was on every timeline, in every layoff post-mortem, in the quiet math every junior developer was doing about their own future. The robots learned to program; draw the obvious conclusion.

This week Indeed's Hiring Lab published a number that runs directly at that story. Since February 2025 — which happens to be almost exactly when Claude Code and the current wave of agentic coding tools arrived — software development job postings are up nearly 15%. Over the same window, job postings across the whole economy fell about 7%.

So the sector everyone said AI would gut is the sector hiring against the trend. Indeed's own headline floats the reframe out loud: AI's effect on jobs may be "flipping from destruction to creation." As someone who writes software for a living and has spent two years reading my own obituary, I want to enjoy this. And I mostly do. But I've learned to distrust a number that feels this good until I find the number sitting next to it.

The number nobody puts next to the good one

Here's the rest of the Indeed data, the part that didn't make the celebratory reposts:

• 71% of the increase in postings is for senior roles.
• 37% of it is jobs that mention AI in the title.
• And even after the bounce, software postings are still about 27.5% below their pre-pandemic level.

Read those three together and the shape of the recovery changes completely. This isn't "software hiring is back." It's "senior, AI-fluent software hiring is back, from a floor that's still a quarter below where we started." The market didn't reopen the door for everyone. It reopened it for people who already have a decade of judgment and a working mental model of how to drive these tools.

Which makes brutal sense the moment you think about what agentic coding actually did to the shape of the work. It didn't remove the need for engineers. It removed the need for the tasks juniors used to be hired to do. The first-year job was never really "write the code" — it was "write the small, well-specified, low-stakes code, and learn judgment by doing it a thousand times." That rung is exactly what an agent now does before lunch, cheaply, without needing to be mentored through it. The ladder didn't get taller. Someone sawed off the bottom three rungs and called it productivity.

The direction this actually points

Strip the optimism and the doom off, and the data is describing a real, specific shift in what our profession is, not whether it exists:

The job is moving from authorship to orchestration and judgment. The postings going up are for people who can decide what to build, whether the machine built it right, and who is accountable when it didn't. That's senior work. It was always the valuable part; AI just stripped away the manual labor that used to disguise it as "coding."

The on-ramp is broken and nobody has fixed it. This is the part I can't wave away. If the entry-level task is the task we automated, then the path from "knows how to code" to "has the judgment we're now exclusively hiring for" no longer runs through the job that used to teach it. We're going to spend the next few years discovering that "just hire seniors" is not a strategy, it's a countdown — because seniors are grown, not minted, and we just automated the farm.

And it's not a local story. The same rise showed up in Canada, the UK, and Australia. Whatever this is, it's the profession changing shape, not one economy having a good quarter.

So, calmly

I think this is genuinely good news, and I think you should hold it carefully. Software isn't a dying field — the postings are the least ambiguous evidence we've had that it's the opposite. But "the field is growing" and "the field is growing for me" are two different sentences, and the gap between them is exactly the 71% and the 37%.

If you're already senior, the data says the leverage of your judgment just went up, and the tools are a force multiplier pointed in your favor. Lean in. If you're early — and this is the part the cheerful headline owes you — the honest read is that the market is asking you to skip a rung that used to be climbable, and no one has built the replacement yet. The tools that squeezed the on-ramp are also, if you're relentless about it, the fastest way anyone has ever had to manufacture judgment: build real things with them, break them, read what they produce with suspicion, and get to "I can tell when this is wrong" faster than the ladder would ever have let you.

The robots learned to program. It turns out the job was never only programming. We're about to find out, expensively, whether we remember how to teach the rest of it.