Software Engineering Job Postings Are Up 18% Since May 2025 — The Most AI-Exposed Job Is Accelerating
Citadel Securities data shows software engineering postings up 18% since May 2025. The most AI-exposed occupation is seeing demand accelerate, not collapse.
Software Engineering Job Postings Are Up 18% — And That’s the Confusing Part
Citadel Securities data shows software engineering job postings up 18% from the May 2025 inflection point. The Federal Reserve confirms it: software engineering employment is at its highest level since November 2023. If you’ve been following the AI displacement narrative, this should feel like a contradiction. The most AI-exposed occupation in the US economy is seeing accelerating demand, not collapse.
This post is about understanding why that’s happening, what the data actually says, and what it means if you’re an engineer or AI builder trying to read the next few years correctly.
What the numbers actually say
Start with the macro picture. The unemployment rate in March 2026 was 4.3%. In March 2020, it was 4.4%. That’s not a typo. Despite two years of aggressive AI adoption, widespread deployment of coding assistants, and constant headlines about software jobs being automated away, the headline unemployment number has barely moved.
The Wall Street Journal, citing a LinkedIn analysis of job posting data, reported that AI created 640,000 jobs between 2023 and 2025 in the US — including new white-collar roles like head of AI and AI engineer. These aren’t reclassifications. They’re net new positions.
And then there’s Stripe Atlas. Q1 2026 startup incorporations were up 130% year-over-year. Stripe just hit 100,000 all-time incorporations through Atlas. That’s not a jobs-destroyed story. That’s a new-demand story.
Not a coding agent. A product manager.
Remy doesn't type the next file. Remy runs the project — manages the agents, coordinates the layers, ships the app.
None of this means the displacement concerns are wrong. It means the picture is more complicated than either side is admitting.
Why the doom narrative felt so airtight
The case for software engineering displacement was never crazy. OpenAI and University of Pennsylvania researchers estimated that roughly 80% of US workers could have at least 10% of their tasks affected by LLMs, and about 1 in 5 workers could see 50% or more of their tasks affected. Anthropic’s Economic Index found that 49% of jobs have already had at least a quarter of their tasks performed using Claude. Microsoft researchers looked at 200,000 Bing Copilot conversations and found the most common AI work was gathering information and writing — exactly the kind of tasks that fill a junior engineer’s first year.
If you’re a software engineer, those numbers hit close. Writing code, summarizing requirements, drafting documentation, routing tickets — a lot of what fills a workday is exactly what AI is best at absorbing.
The logical conclusion seemed obvious: fewer engineers needed, lower salaries, shrinking headcount. And some of that is happening in specific pockets. But the aggregate data isn’t cooperating with the narrative.
The Jevons problem (and why it keeps surprising people)
University of Chicago economist Alex Imms wrote an essay called “What Will Be Scarce” that’s been circulating in AI circles. He applies Jevons’ paradox to AI: when a resource gets cheaper, total consumption of that resource often goes up, not down, because the lower cost unlocks demand that previously couldn’t justify the expense.
Ezra Klein picked this up in his New York Times essay “Why the AI Job Apocalypse Probably Won’t Happen” and gave a personal example. When he started his podcast, he was its only researcher. Now he has a team. Has AI made his job easier? His answer: not in the least. He does more challenging episodes because he can. Every enthusiastic AI adopter he knows is working harder than ever because there’s more they can do.
This is the pattern showing up in software engineering data. Sequoia partner Konstantin Beuler flagged it directly: the acceleration in software job postings “violates the displacement narrative.” The argument from Merzmik Ahmed at Emergence AI is blunter: “Code is digital brick. If bricks get much cheaper and easier to lay, you don’t use fewer builders. You build what was previously too expensive, too slow, too bespoke, or too annoying to justify.”
The key distinction is between elastic and inelastic demand. Software is elastic. When the cost of building software drops, the amount of software people want built expands. Payroll processing is inelastic — you need to run payroll once a month regardless of how cheap automation gets. Software development is the opposite: every cost reduction reveals a backlog of things that were previously too expensive to build.
This is why the Claude Mythos benchmark results on SWE-bench — 93.9% on the coding benchmark — don’t straightforwardly translate into “fewer engineers needed.” A model that can close 93.9% of software issues autonomously doesn’t eliminate the demand for software. It makes software cheaper to produce, which means more of it gets produced.
The travel agent pattern (and why it’s not the whole story)
There’s a counterargument worth taking seriously, and it comes from the travel agent analogy.
Expedia didn’t erase travel agents overnight. Online booking changed the economics of routine work first, and nothing seemed to change immediately. The visible break came later, during downturns, when the industry had to admit what had already shifted. The agents who survived moved toward complex trips, corporate travel, luxury travel, emergency problem-solving — the parts of the work that simple booking paths couldn’t handle.
A lot of knowledge work is sitting in a similar position right now. AI doesn’t have to replace your whole job to put you on thin ice. It only has to pick away at enough pieces that when the next shock comes — a recession, a budget freeze, a reorg — the company asks the question it’s been avoiding: why is this role bundled this way?
The engineers who are seeing accelerating demand are not the ones doing what AI does well. They’re the ones doing what AI can’t do yet: holding ambiguous requirements, making architectural calls where the tradeoffs aren’t obvious, reading what’s actually happening in a client relationship versus what’s being said in the ticket. The 18% increase in job postings is real, but it’s not evenly distributed across all engineering work.
Anthony Pompliano, who previously believed AI would steadily reduce engineering headcount, updated his view publicly: “The number of software engineers being hired has been increasing. The number of open software engineer roles is growing. The number of new college grads who get hired has increased 5.6% over the last 12 months.” He added that companies in his portfolio are aggressively hiring to keep up with demand for their products — and if AI makes employees more productive, companies want as many productive units of labor as possible.
That’s the Jevons argument in practice. More productive engineers → lower cost to build software → more software gets built → more engineers needed.
What’s actually driving the new demand
Three things are creating demand that didn’t exist before.
The agentic era unlocks a new category of software. When AI can handle routine coding tasks, the constraint shifts from “can we build this?” to “what should we build?” That question requires engineers who can scope, architect, and evaluate — not just implement. The Claude Code source code leak revealing hidden features is a good example of how quickly the tooling is evolving: the tools themselves are becoming more capable, which means the humans directing them need to be more capable too.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
AI-native products require engineering to build. Every company building AI products needs engineers. The 640,000 jobs LinkedIn identified between 2023 and 2025 weren’t all at AI labs — they were distributed across companies building AI-powered features into existing products. Platforms like MindStudio handle orchestration across 200+ models and 1,000+ integrations with a visual builder, but someone still has to design the workflows, evaluate the outputs, and connect the system to real business logic. That work is engineering work.
Startup formation is accelerating. The 130% year-over-year increase in Stripe Atlas incorporations means more companies exist that need software built. Derek Thompson noted that AI startups are seeing faster revenue growth than is historically normal. More startups with faster revenue growth means more engineering hiring.
The part the data doesn’t resolve
Here’s where I’ll give you my actual opinion: the macro data is real, but it’s a lagging indicator.
The travel agent pattern is also real. The 18% increase in software engineering job postings doesn’t tell you which engineers are seeing that demand. It’s plausible — likely, even — that demand is concentrating at the top of the skill distribution while eroding at the bottom. Junior engineers doing routine implementation work are in a different position than senior engineers making architectural decisions.
Ezra Klein made a point in his essay that’s worth sitting with: “A world where AI displaces 8 million workers might be harder to handle than a world where it displaces 80 million workers. A mass unemployment event would force a wholesale restructuring of our economy. We are crueler when displacement is more limited.” The China trade shock cost roughly 2 million jobs. Small in aggregate, devastating in specific communities, and largely ignored by policy. Targeted displacement is the scenario we’re least prepared for.
The macro numbers being fine doesn’t mean everyone is fine. It means the aggregate is absorbing the disruption so far. That’s different from “nothing is changing.”
How to read the next 18 months
If you’re an engineer trying to make sense of this, a few things are worth tracking.
Watch the composition of job postings, not just the count. An 18% increase in postings is meaningful, but if the postings are shifting toward senior roles and away from junior roles, that tells a different story than the headline number. The Federal Reserve data showing software engineering jobs at their highest since November 2023 is encouraging, but it’s worth asking what kinds of jobs.
The agentic era is creating new roles faster than it’s eliminating old ones — for now. Agent ops engineers, context librarians, eval engineers, coordination architects — these roles are emerging because someone has to direct and evaluate the agents. Understanding how Claude Code’s effort levels work is the kind of operational knowledge that’s becoming more valuable, not less. The engineers who understand how to get reliable output from AI systems are in a different position than the engineers who are being replaced by those systems.
Hire a contractor. Not another power tool.
Cursor, Bolt, Lovable, v0 are tools. You still run the project.
With Remy, the project runs itself.
Spec-driven development is the next abstraction layer. The history of programming is a history of raising the abstraction level: punch cards to assembly to C to TypeScript. Tools like Remy take this further — you write a spec in annotated markdown and it compiles into a complete TypeScript stack with backend, database, auth, and deployment. The source of truth shifts from code to spec. Engineers who understand how to write precise, well-structured specs are positioned well in this world; engineers whose value was in typing out implementations are not.
Startup formation data is the leading indicator to watch. Job postings lag reality by months. Startup incorporation rates lead it. If Stripe Atlas Q1 2026 incorporations are up 130% year-over-year, those companies will be hiring engineers in 6-18 months. The demand pipeline looks healthy.
What the 18% number actually tells you
The Citadel Securities data is a useful corrective to the doom narrative, but it’s not a reason to stop paying attention. The correct read is something like: the aggregate demand for software engineering is holding up better than the displacement narrative predicted, the Jevons effect is real and operating in software, and the engineers seeing the most demand are the ones doing work that AI can’t yet do reliably.
That last part is the actionable piece. The question isn’t whether AI will affect software engineering — it already is, significantly. The question is which parts of software engineering are elastic (demand expands as cost falls) and which parts are inelastic (the work just goes away). Architectural judgment, ambiguous requirements, cross-functional coordination, evaluation of AI outputs — these look elastic. Routine implementation of well-specified features looks inelastic.
The comparison between GPT-5.4 and Claude Opus 4.6 on coding tasks is a useful lens here: both models are genuinely capable at implementation tasks, which means the value of human engineers doing pure implementation is declining. The value of human engineers who can direct, evaluate, and architect is not.
The 18% increase in job postings is real. So is the travel agent pattern. Both things are true, and the engineers who are going to be fine are the ones who understand the difference between them.
