The great inversion: what if engineers replace everyone else?

The whole conversation about AI and software has one arrow pointing in one direction: agents replace engineers. Every thread, every panel, every LinkedIn hot take assumes the developer is the one about to be automated away.

What if the arrow is pointing the wrong way?

Here's the uncomfortable inversion. The person who masters agents doesn't get replaced by them — they use them to absorb everyone else's job. Not "engineers survive." Engineers expand. The threatened role becomes the orchestrating one.

This isn't wishful thinking from someone with a computer science degree. Five things, each with a name and a number attached, convinced me it's worth taking seriously.

1. The skill that survived isn't coding — it's judgment

Remember "vibe coding"? The term Andrej Karpathy coined in early 2025 for describing an app in plain English and shipping whatever the model produced. It raised the floor beautifully. Then it hit production, and the floor gave way — security holes, undocumented business logic, code nobody could maintain.

By Sequoia's AI Ascent this spring, Karpathy had moved past it, drawing a hard line between vibe coding and what he now calls agentic engineering, a framing IBM has since adopted: vibe coding raises the floor, agentic engineering preserves the ceiling. Coordinating fallible, stochastic agents against a real spec — writing the evals, reviewing the diffs, owning the architecture and the security boundary — turns out to require exactly the taste and rigour engineers built over years. The multiplier is the point: at the top tiers, technical mastery becomes more of an advantage, not less.

So the thing AI made cheap was typing code. The thing it made more valuable was knowing what good looks like — which decides who ends up directing the agents and who ends up managed by someone who does.

2. Three roles are collapsing into one — and it has a name now

Shipping a digital product used to need three people: a product manager to decide what to build, a UX designer to shape it, a developer to make it real. Marc Andreessen, in a January 2026 interview on Lenny's Podcast, describes what's replacing them as a "three-way standoff" where each role now believes it can do the other two with AI — and argues all three are essentially right, because AI is genuinely good at all three functions. The job that emerges he calls the builder.

This isn't only a VC's slide. Speaking at Davos, Satya Nadella said Microsoft-owned LinkedIn had merged product management, design, program management and front-end engineering into a single "full-stack builder" role — which he called the biggest structural change to software teams he'd seen since the 1990s.

Andreessen is careful that you can reach the builder track from any starting point — coding, product, design, even support. But notice the gravity: product and design decisions ultimately have to survive contact with technical reality, and the engineer already lives there. The absorption doesn't have to run one way, but it has a strong current.

3. The one-person company stopped being a joke

The clearest evidence is in revenue, not research. Sam Altman's 2024 line about a betting pool for the first one-person billion-dollar company sounded like hype. Then Anthropic's Dario Amodei was asked the same and answered "2026," at 70–80% odds. Then it got concrete: as Forbes and Inc. reported off an April New York Times profile, Matthew Gallagher's telehealth startup Medvi posted $401M in first-year sales with a two-person headcount, running AI across code, marketing, and customer service.

I'd flag Medvi carefully, because it cuts both ways. The same coverage surfaced an FDA warning letter, a spam class action, and AI-generated "doctors" in its ads — critics argued the AI mostly turbocharged the marketing, not the medicine. Which is exactly the tell: a technical solo operator can now compress a marketing and support org into an agent stack — and can just as easily compress the failure modes into one unsupervised point. The capability is real. So is the governance hole.

4. The bottleneck moved from the code to the product

When writing code stops being the constraint, the constraint moves upstream — to clarity of vision and knowing what users need. That sounds like it favours the product manager. In practice it pulls the engineer toward product, armed with agents that query analytics directly. Product School cites Andrew Ng suggesting the developer-to-PM ratio could flip toward 2:1, a sharp reversal from the old norm of several PMs' worth of product work per developer — more developers, each absorbing more of the product job themselves.

The binding skill becomes deciding what to build. The engineer who can also build it holds both halves at once, and the gap between "person who decides" and "person who executes" starts to close inside a single head.

5. The high-context generalist eats the org chart

Andrew Ng has a phrase for where this lands: the high-context generalist. Picture a two-person team covering five functions — engineering, product, marketing, legal, design. The old answer was "you can't, hire more people." The new one is that a deeply technical person uses frontier AI to be genuinely competent in the four areas outside their specialty while keeping the technical depth that anchors everything.

What's load-bearing here isn't "generalists beat specialists." It's that a specific kind of specialist — the technically grounded one — is best placed to go wide without going shallow. Take the technical depth away and the AI-assisted marketing, the AI-assisted legal, all of it gets brittle fast.

The honest counter-argument

I won't sell this cleanly, because it isn't clean. Three things could break the thesis.

The first: domain expertise may not absorb as easily as code did. An engineer with an agent can draft a marketing campaign; whether it's a good one is the taste a marketer also spent years building. The absorption may be real for execution and shallow for judgment — the exact trap vibe coding fell into, one layer up.

The second: this could be a transition state. Roles consolidated during the internet era too, then re-specialised once tools matured and volume grew. "One engineer does everything" may be the awkward middle, not the destination.

The third, and the one I sit closest to given where I work: the solo-operator dream is a governance nightmare in a regulated setting. Even the boosters concede the limit — Gartner analysts note today's agents still need narrowly-scoped decisions to stay reliable, and the consumer-software cases that work best are the ones with the least regulatory surface. In a bank, someone still has to own accountability when an agent-run process touches a client or a control. That constraint doesn't disappear because the demo is impressive — Medvi is the cautionary version of exactly that.

Bottom line

The next phase of this shift may not be won by whoever builds the smartest agent. It may be won by whoever the agents make most powerful — and the early evidence says that's the person who already understands systems well enough to direct them without being fooled by them.

"Will AI replace software engineers?" might be the wrong question entirely. The sharper one: which profession does the agent-fluent engineer absorb next — and is your role one of the ones being quietly folded into someone else's screen?