I’ve been watching the same conversation loop for almost three years now. Every time a new coding assistant drops, someone declares the end of software developers. Then the hiring numbers come out and the story gets messy. Right now the mess looks like this: AI tools are generating a bigger share of code than ever, yet U.S. software-development job postings have climbed almost fifteen percent since late February while overall job postings dropped seven percent. That gap is what keeps me up at night in the best possible way. It forces us to stop treating “AI takes jobs” as a simple equation and start looking at who actually gets hired, who gets left waiting, and what the next decade might demand from anyone who wants to stay in the game.
Why Job Postings Are Rising While AI Writes More Code
The rebound did not appear out of thin air. After the pandemic hiring frenzy peaked in 2022, tech companies slammed the brakes hard. Postings for software roles cratered. What we are seeing now is a recovery, but a selective one. Even after the recent climb, software-development openings still sit roughly twenty-seven and a half percent below their pre-pandemic high. The climb itself is concentrated. Seventy-one percent of the increase between May of last year and May of this year came from senior-level roles. Thirty-seven percent of the same increase carried an AI-related title somewhere in the job description. That concentration tells a clearer story than the headline numbers alone.
Companies are not hiring more developers because AI failed. They are hiring because AI succeeded at making code cheaper and faster to produce. When output rises that sharply, the bottleneck shifts. Someone still has to decide what to build, evaluate whether the generated code is safe, connect the new pieces to legacy systems, and explain the whole thing to clients or product teams who do not speak in pull requests. Those tasks sit higher on the experience ladder. Junior developers who once spent months writing boilerplate now find that boilerplate already exists, polished and waiting. The work that remains is harder to hand off to an untested newcomer.
The New Shape Of Demand
In my view the most interesting shift is not the raw count of openings but the skill mix inside them. Employers keep repeating the same quiet message: we need people who can supervise the machine, not just feed it prompts. That means stronger judgment around architecture, security, and long-term maintainability. It also means softer skills that used to feel optional. Communication, the ability to translate technical trade-offs for non-engineers, and a knack for spotting when an AI suggestion is quietly wrong all matter more than they did five years ago.
One economist who tracks hiring platforms put it plainly: firms realize that even when AI handles the typing, they still need developers to refine products and help customers actually use the technology day to day. The demand is moving toward people who sit between the model and the business problem. That middle space is not entry-level work.
What The Data Actually Shows About Young Developers
Here is the part that stings. While overall software-developer employment is projected to grow about ten percent through 2035, adding roughly one hundred seventy-five thousand positions, the path for people aged twenty-two to twenty-five looks rockier. Employment for early-career software developers in that age band has fallen nearly twenty percent since generative tools became widespread. Older cohorts, by contrast, have expanded. The ladder is still there, but the bottom rungs are missing a few boards.
I’ve talked with recent graduates who spent months applying to roles that once hired dozens of juniors every quarter. Many of those roles simply evaporated or turned into “two years of production experience required” postings overnight. The same tools that let a senior engineer ship three times faster also let that senior engineer absorb the volume of work that used to justify an intern or a new hire. Efficiency cuts both ways.
Today’s entry-level talent becomes tomorrow’s senior experts. If employers do not build those pipelines now, they will not have the senior talent they need a decade from now.
That warning feels accurate. Pipelines do not rebuild themselves. When companies stop training juniors, they eventually face a shortage of mid-level people who understand the messy reality of production systems. The short-term productivity win can become a long-term talent drought.
How AI Coding Tools Changed The Daily Reality
A few years ago one major platform reported that on projects where its assistant was available, the tool was already writing close to half the code. More recent internal figures suggest the share has climbed higher still. One in three pull requests now involve an AI agent in some capacity. A year earlier that number sat below one in ten. If the pace continues, most code pushed to public repositories could soon be machine-generated, and a large portion may never be read line by line by a human.
That volume creates its own problems. More code means more surface area for subtle bugs, security gaps, and architectural drift. The humans who remain have to scale their oversight at the same rate the machines scale their output. Some days that feels manageable. Other days it feels like trying to proofread a library while the printer keeps running.
I’ve found that the developers who thrive in this environment treat the AI less like a magic wand and more like a very fast junior colleague who never sleeps and occasionally hallucinates. They set clear boundaries, review everything that matters, and keep ownership of the final decisions. The ones who struggle often hand over too much trust too early and then spend twice as long untangling the results.
Senior Roles Versus Entry-Level Pressure
The concentration of new postings in senior and AI-titled roles is not accidental. Organizations discovered they can produce more software with fewer people, but only if those people already know how the whole system fits together. A senior engineer who has lived through three major refactors can spot when an AI suggestion will create technical debt six months downstream. A new graduate usually cannot. That difference in judgment is exactly what the market is pricing right now.
At the same time, the long-term employment outlook remains positive. Official projections still show solid growth for the occupation as a whole. The catch is that growth will likely favor people who already have a track record. The classic path of “get a computer-science degree, join as a junior, learn on the job” has become narrower and steeper.
- Senior and staff-level openings account for most of the recent rebound
- AI-specific titles capture more than a third of the new demand
- Entry-level employment for the youngest cohort has declined sharply
- Overall profession is still projected to expand through the next decade
Those four points sit in tension with one another. The profession is growing, yet the on-ramp is shrinking. That tension is the real story.
What Skills Actually Move The Needle
If I had to advise someone starting out today, I would keep the list short and slightly uncomfortable. First, get comfortable using the tools. Ignoring them is no longer a viable strategy. Second, do not stop at using them. Learn to evaluate their output the way a senior would. That means understanding testing, security basics, performance trade-offs, and the cost of technical debt. Third, invest in the skills that AI still handles poorly: clear writing, stakeholder conversations, and the ability to frame a business problem so the technical solution actually matters.
Perhaps the most interesting aspect is how quickly the definition of “technical skill” is expanding. Knowing the latest framework still helps, but knowing when the framework is the wrong tool matters more. The same is true for system design. AI can propose a clean microservice layout in seconds. Deciding whether that layout will survive the next three years of product changes still requires human experience.
Employers have a parallel responsibility. If they keep hiring only people who already possess that experience, the pool will eventually dry up. Some companies are already experimenting with structured apprenticeship-style programs that pair new graduates with AI tools under heavy senior supervision. Those experiments are still small, but they feel like the only realistic way to rebuild the missing rungs.
The Longer View Through 2035
Official forecasts still look constructive. Employment of software developers is expected to rise from roughly 1.72 million to about 1.89 million over the next decade. That is real growth, even if the composition of the roles continues to shift. The open question is whether the growth will be evenly distributed or whether it will keep concentrating among people who already have five or more years under their belts.
History offers a few parallels. Earlier waves of automation in other knowledge fields rarely eliminated the occupation. They changed the daily work and raised the bar for entry. Spreadsheets did not end accounting; they ended the need for rooms full of people doing arithmetic by hand. The accountants who remained spent more time on judgment and client relationships. Something similar appears to be unfolding here, only faster.
Still, the speed creates risk. When the transition happens over decades, training systems have time to adapt. When it happens over a few product cycles, a generation of graduates can find themselves locked out before the market realizes it needs them later. That lag is already visible in the age-band data.
Practical Steps For Developers At Every Stage
For people already in senior roles the advice is almost too obvious: keep leaning into the parts of the job that require context the models do not have. Architecture decisions, cross-team coordination, and the quiet political work of getting stakeholders aligned remain stubbornly human. The more you treat AI as a force multiplier rather than a replacement, the more valuable you become.
For mid-level developers the window is still open but narrowing. This is the moment to volunteer for projects that force you to own the full lifecycle, not just the feature. Reviewing AI-generated code in production environments is one of the fastest ways to build the judgment that senior postings now demand.
For students and recent graduates the path is harder, yet not closed. Side projects that demonstrate real systems thinking still cut through. Contributions to open-source repositories where AI assistance is common but human review is rigorous can serve as a public portfolio of judgment. Internships that emphasize mentorship over pure output are worth more than they used to be. And yes, learning to write clear documentation and explain technical choices to non-engineers is no longer optional polish. It is part of the core skill set.
I’ve watched too many talented juniors freeze when asked to defend a design choice in a meeting. The code itself was fine. The ability to talk about the code under pressure was missing. That gap is exactly where AI cannot yet help.
Risks That Could Still Change The Trajectory
Nothing about the current rebound is guaranteed to last. If models continue improving at the same clip, the volume of code that needs human oversight could grow faster than the supply of people qualified to provide it. Security researchers already warn that the sheer quantity of new code is outpacing the ability to audit it thoroughly. At some point organizations may decide the risk is unacceptable and slow the rate of AI-assisted generation until better verification methods exist.
Another risk sits on the demand side. If economic conditions tighten again, the first cuts usually hit the newest hires and the most experimental teams. The same companies that are expanding senior AI roles today could freeze hiring tomorrow. The fifteen-percent rise in postings is real, but it is still early and still concentrated in a handful of sectors.
Finally, the cultural risk is real. When a generation of developers never gets the chance to make beginner mistakes under supervision, the institutional knowledge that used to transfer from one cohort to the next starts to thin out. That knowledge is hard to replace with documentation or model weights.
A More Nuanced Picture Than The Headlines
The simple narrative that AI will erase software developers never matched the data. The equally simple narrative that everything is fine and demand is booming also misses the mark. What we actually see is a profession that is growing overall, shifting toward higher experience levels, and leaving a visible gap at the entry point. The tools are powerful. The labor market is adjusting. The adjustment is uneven.
In my experience the people who stay calmest through these shifts are the ones who treat technology as a moving target rather than a finished threat. They update their skills without panic, they mentor when they can, and they keep asking the only question that still matters: what problem are we actually trying to solve, and does this new capability help us solve it better? That question has outlasted every previous wave of automation. I suspect it will outlast this one too.
The next few years will show whether companies rebuild the pipelines or keep relying on the existing stock of experienced talent. Either choice will shape the profession for a long time. For now the numbers say demand is real, the composition of that demand has changed, and the people who adapt fastest will keep finding work. The rest of the story is still being written, one pull request at a time.
Looking ahead, the interplay between rising code volume and the need for human judgment feels like the central tension of the decade. Developers who can hold both sides of that tension—speed and care, automation and accountability—will remain scarce and valuable. Those who treat the tools as either pure magic or pure threat will find the market less forgiving. The data we have today is incomplete, but it is already clear enough to reject the most extreme predictions on both ends. Software development is not vanishing. It is changing shape in real time, and the people inside the field are the ones who get to decide what the new shape looks like.
That responsibility is heavier than it used to be. It is also more interesting. For anyone willing to stay curious, keep learning the parts of the craft that still require a human mind, and help the next cohort find their footing, the next ten years still look full of possibility. The postings are rising for a reason. The question is whether enough of us will make sure the reason includes a future for the people just starting out.
