The Vibe Coding Paradox: Why the AI Revolution Is Creating More Engineering Jobs, Not Fewer

I've been asking our portfolio companies an obnoxious question. The answer I kept getting was the last one I expected — and it contradicts almost everything you've read about AI (for now)

It’s been a while since I posted here. But a series of conversations with our portfolio founders this past month — and a genuinely sharp Aaron Levie episode on 20VC — rattled something loose that was too counterintuitive to leave in my notes. Turns out the loudest voices in the AI narrative are often the furthest from the ground truth.

Posted May 12th, 2026 · Reflections from the portfolio

I have to confess: I sometimes ask the annoying VC question — which seems, of the time, given the markets and high growth we are seeing in so many companies, fastest 100M is the chart everyone is trying to say out loud.

My latest one: “You’re already moving fast — what would it actually take to go 5x faster?” This is common among most VCs, as the growth rates and absolute revenue of many AI-centric companies are rising rapidly.

I expected the usual suspects on why not. Sales cycles are stretching as companies move upmarket. Reps that need another quarter to ramp. Go-to-market motions that need tuning. Product gaps and AI adoption of products within the legacy product set, and international and global challenges. There is a true market pull, but we can’t meet the demands, or we are not demand-constrained but supply-constrained. These are real answers, true answers — answers I’ve heard dozens of times across dozens of portfolio calls.

Instead, across five separate conversations — five different founders, five different stages of growth — I heard nearly the exact same thing. And it stopped me cold.

Four out of five said: we need more engineers.

4/5portfolio founders named engineering capacity as their #1 growth constraint

5×faster growth in demand for AI-driven features vs. one year ago

500K–1Mnew “agent operator” roles Levie predicts will emerge in the near term

Let that land for a moment. These are AI-native companies. Built on frontier models. The poster children, by every media account, for the “one founder with a prompt can out-ship a ten-person team” narrative.

And their single biggest constraint isn’t capital, distribution, or product-market fit. It’s senior engineering talent.

“Everyone is Wrong; We Will Have More Developers in Five Years” — Aaron Levie, CEO of Box ($1B+ ARR)

Now, I know what you’re thinking. You’ve read the headlines. Spotify’s best developers haven’t written a line of code since December. Boris Cherny at Anthropic ships 30 PRs a day without touching a keyboard. Shopify’s River agent is handling a “ludicrous amount” of the company’s engineering. How do I square those stories with founders telling me they need more engineers?

Two things are true simultaneously, and I’ll be honest — I don’t have inside access to those organizations. But I’d bet a significant amount that there is a ton of nuance buried underneath those headlines that never makes it into the coverage. “Our engineers haven’t written code in six months” is almost certainly a compressed, PR-friendly version of something far more layered: a handful of elite engineers at a frontier AI company using bleeding-edge internal tooling, working on greenfield problems, with token budgets that most companies wouldn’t sanction. It describes an exception — a remarkable, genuinely impressive one — not the standard operating condition of a 200-person engineering org shipping enterprise software against a real SLA. The headline travels. The asterisk doesn’t.

The more important question isn’t whether AI can eliminate coding for the best engineers at the best AI labs. It’s what happens to the other 99% of software that the world actually runs on — the enterprise systems, the integrations, the legacy infrastructure, the security layers, the products that millions of people depend on and that break in creative and expensive ways. That software is not getting built faster by reducing engineering investment. It is getting built faster by giving great engineering teams dramatically more powerful tools — and then letting them run.

The narrative vs. the ground truth

The market’s story is elegant. Almost too elegant. “Vibe coding” was supposed to democratize software creation so completely that the traditional engineer becomes optional. One visionary, one laptop, one good prompt. The code practically writes itself.

Like most compelling narratives in venture, this one contains a kernel of truth wrapped in a husk of hyperbole.

The kernel: yes, the cost of generating a line of code has genuinely collapsed. A developer pairing with Claude, Cursor, or Copilot can accomplish in an afternoon what once took a week. The floor of individual productivity has risen in a way that would have seemed impossible three years ago.

“We haven’t removed humans from the loop — we’ve just changed where they enter the loop.”

Aaron Levie, CEO of Box · 20VC, April 2026

The husk: none of that has reduced the demand for elite engineering talent. If anything, it has supercharged it. Aaron Levie — who has led Box through the cloud revolution, the mobile era, and now the AI era — named this episode title with precision: “Everyone is Wrong; We Will Have More Developers in Five Years.” That’s not a hedge. That’s a conviction bet from a CEO who has watched the same cycle play out three times.

More fuel requires a bigger engine

Here is the structural reality the hype cycle keeps missing: frontier models are fuel, not factories. You can have all the fuel in the world — without an engine to convert it into motion, it just sits there. The engine that takes raw AI capability and turns it into a product that a Fortune 500 company will bet their operations on is still a human engineering team. A very skilled one.

The “token maxing” phenomenon is the most vivid illustration of this dynamic. Across Silicon Valley, companies are actively pushing engineers to consume as many AI tokens as possible — treating token burn not as a cost problem but as a signal of ambition and exploration. Levie said he would rather his team waste tokens, because it means they are experimenting. Jensen Huang went further, saying at a recent event that he would be “deeply alarmed” if an engineer earning $500,000 a year wasn’t using the equivalent of $250,000 worth of tokens.

“For me right now, I’m like, ‘Yeah, we should probably waste a lot of tokens — because that means we’re trying new things.’”

Aaron Levie, CEO of Box

Think about what that means structurally. If you’re burning more tokens, you’re generating more code, more features, more product surface area. That surface area has to be reviewed, architected, secured, scaled, and maintained by someone who understands what it does and why. The AI's output is the input to the engineering problem. More fuel doesn’t reduce the engine requirement. It expands it.

The bottleneck moves — it never disappears

Levie’s conviction about job growth is rooted in something economists call Jevons Paradox: when a technology makes a resource cheaper to use, total consumption of that resource tends to increase, not decrease. Steam engines didn’t reduce coal consumption — they made coal so useful that global demand exploded. The same dynamic is playing out in software right now, and in virtually every other knowledge industry.

“We’re going to use AI to accelerate output in one area, and then eventually you run into a new bottleneck somewhere else in the process that still requires humans.”

Aaron Levie · Benzinga, April 2026

His legal industry example is concrete and instructive. AI has made generating legal content — letters, patent drafts, contract red-lines — trivially easy. Law firms are now drowning in AI-generated material from clients who can produce in minutes what once took days. The bottleneck didn’t vanish. It moved downstream, to the lawyers who must review, validate, and respond to all of it. Levie’s prediction: there will be more lawyers in five years, not fewer.

“Automation is gonna actually just force us to see the next set of bottlenecks that are in all of these industries.”

Aaron Levie · 20VC, April 2026

In our portfolio, I watch this in real time. AI makes it faster to ship features. So companies ship more features. More features means more security surface area, more integration complexity, more edge cases requiring architectural judgment. The bottleneck moved up the stack — from writing code to knowing which code to write, why, and what happens when it breaks at 2am for your most important enterprise customer.

The three walls companies hit at scale

The talent moat problem

Recruiting the top twenty or thirty AI engineers in any given specialty remains a bloodbath. AI tools haven’t loosened that market — they’ve concentrated it. When one great engineer can genuinely do the work of five average ones, every company wants the great ones more desperately, not less. Levie is direct about a broader myopia in how the tech industry frames this: “We are so myopic and self-interested and we think that the entire industry is the tech industry.” The demand for engineering skill is expanding far beyond Silicon Valley — into healthcare systems, financial infrastructure, legal operations, and manufacturing. The talent war is getting bigger, not smaller.

The feature complexity trap

Every enterprise customer now expects deep, custom AI integration as a baseline expectation, not a competitive differentiator. That shift has created a massive backlog of high-complexity features that vibe coding cannot clear once you move past the prototype stage. Levie’s framing on workflow redesign lands exactly here: you can’t just bolt AI onto an existing process. As he put it, “the workflow needs to be redesigned for agents, not for people.” Executing that redesign well — at enterprise scale, with real security and compliance requirements — is architectural work that requires senior judgment, not more tokens.

The prototype-to-production cliff

Building an impressive demo has never been easier. Frighteningly easy. But the cliff between “impressive demo” and “production system that an enterprise will bet their operations on” is steeper than it looks from the outside. Security, resilience, latency under load, graceful degradation, compliance requirements — none of these yield to a clever prompt. And as Levie observed about software evolution generally: “upgrading software is a multi-year effort — there’s not some magical moment where you can just secure everything.” That effort requires people who can hold the full system in their heads across that entire arc.

The new jobs nobody is talking about

Here is where the narrative gets genuinely interesting — and where most of the public discourse has completely missed the story. This shift isn’t just about preserving existing engineering roles. It’s about a wave of entirely new roles that didn’t exist eighteen months ago.

Levie’s prediction is striking in its specificity. Somewhere between 500,000 and one million new jobs will emerge in what he calls the “agent operator” category — people who manage, orchestrate, direct, and govern AI agents running inside enterprises at scale. These are not traditional software engineers. They’re not traditional operations people. They sit at the intersection of deep technical understanding and domain expertise, and they represent an entirely new labor market.

“There is gonna be 500,000 to a million jobs created — and it’s basically some kind of agent operator.”

Aaron Levie · 20VC, April 2026

The talent war is about to expand its surface area dramatically. It’s not just about hiring ML engineers anymore. It’s about who can identify, develop, and retain the people who can actually run AI-powered organizations at scale. The companies building that capability now are creating a structural advantage that will be very hard to replicate in two years.

What this means for founders right now

The companies I’m most excited about in our portfolio aren’t treating AI as a cost center — a mechanism for doing the same things with fewer people. They’re treating it as a force multiplier — a way to give their best engineers the leverage to do things that simply weren’t possible before.

That distinction sounds subtle. It is not. It is the difference between an engineering culture that is quietly demoralized (”we are being replaced”) and one that is genuinely energized (”we have superpowers — what do we build?”). The senior talent you need most reads that energy instantly. They will not work for companies that see them as a cost to be optimized away.

Levie frames the design challenge with precision: “If we actually want to get real leverage from automation, we need to start to redesign the workflow.” That redesign — not the prompting, not the token burning, but the hard architectural work of rebuilding processes around AI agents — is the highest-leverage thing a technical founder can be doing right now.

So: do not let the hype cycle talk you into under-investing in your core engineering function. The narrative that AI makes talent redundant is being written by people who have never shipped an enterprise-grade AI product under real security requirements, for customers who will call at 2 am when something breaks.

The talent war isn’t over. It just moved up the stack — away from implementation, toward architecture. Away from writing code, toward knowing exactly which code to write, why, and what happens when it fails at the worst possible moment.

That judgment has never been more valuable. And the people who have it know it.