Two days ago, Pieter Levels posted that software is dead. His view is that we are now in the era of personal software, where everyone builds what they need, whenever they need it, or their AI agent builds it for them. The post got a lot of agreement. On the same day he also said he had cancelled almost all his SaaS subscriptions.
I understand why the idea feels true right now. Agents write real code. I use them every day. But the claim falls apart once you look at what it actually asks people to do, and the best reply I read put it in one line: good restaurants don’t compete with your air fryer and microwave.
I want to explain why I agree with that, and why I think “software is dead” gets the economics of software almost exactly backwards.
What the argument asks of a normal person
The strong version of the claim says you tell your agent what you want, the agent builds the software on the fly, and you never need a real product again.
Look at the trade this asks for. Today you open an app and it works. In the new world you write a prompt, wait some minutes, and get something that may or may not do the job, delivered with the confidence of a horoscope. Then at the end of the month you pay for the tokens.
Those tokens are not cheap, and they are getting more expensive in total even as the price per token falls. A guide I read recently puts agentic coding tools at $200 to $2,000+ per engineer per month in token spend, far above the $30 to $60 seat price people used to assume. Uber’s CTO confirmed the company used up its whole 2026 AI coding budget in four months. That is a company with a real budget and a lot of engineers. Now imagine a small business owner in Lagos or London or Sydney paying in dollars to rebuild an invoicing tool that already exists for a few dollars a month. The invoice for the invoicing tool is the bigger invoice.
There is also the time. A product you open in one click beats a product you wait fifteen minutes for, every single time, for most people.
Why the restaurant still exists
I like the air fryer comparison because it is fair to both sides. The air fryer is a good tool. It made home cooking better and faster. Lots of people eat at home more because of it.
But nobody closed their restaurant because air fryers got popular. Nobody leaves a good restaurant saying, “nice, but my air fryer could do this in twelve minutes with some foil.” When you pay a restaurant, you are not paying for heat. You pay for years of practice, sourcing, consistency, and someone else washing the pots. The heat was never the expensive part.
Software works the same way. The first version of a product is the cheap part. The expensive part is the years of edge cases that real users find. Steve Yegge, one of the loudest voices for vibe coding, still points to Sourcegraph as a code search engine with ten years of enterprise bug fixes that you almost certainly would not try to clone yourself. When the people who champion vibe coding say that, I take it seriously.
Everyone has a camera now
The camera is another comparison I find most useful, because we already ran this experiment.
In 2008, Canon released the 5D Mark II. Professionals loved it. It shot 21 megapixel photos and was one of the first DSLRs to record full HD video, and the body alone cost about $2,700 before you bought a single lens. Today the phone in your hand matches or beats it on many of those specs. It also edits, adds filters, and posts to the internet before the moment is over. If “everyone has the tool” meant “nobody needs the professional,” photographers should have disappeared about ten years ago.
They did not. Wedding photographers still get booked months ahead. Brands still pay for product shoots. Your uncle with the newest phone is at every wedding, and he is not the photographer. He is the reason the photographer has to say “excuse me” forty times during the vows.
What the cheap camera removed was the cost of taking a picture. What it did not remove was knowing where to stand, reading the light, catching the moment once, editing a few hundred photos into a story, and being reliable on a day that will not happen again. People still pay for that, because that was always the actual job.
AI Code generation is like the phone camera. It makes making software cheap, and that is a good thing. It does not make judgment, reliability, and years of fixes cheap. Those are what people pay for when they pay for good software.
Everyone building the same thing
The part of the argument I find funniest is the aggregate view. If every person’s agent builds software on demand, most of those agents will build the same few things over and over. Millions of people pay to generate the same to-do app, the same habit tracker, the same file sharer.
Your agent spent three days and $500 of tokens building you a habit tracker? Mine did too. Let’s celebrate how personal our habit trackers are. The only habit either of them built was a monthly token bill.
I have built personal software myself. A recent one is Ferry, a small file transfer app I made because I wanted a simple way to move files between my devices. That is exactly the kind of thing the “software is dead” crowd describes. But the sensible next step was not to keep it to myself. I made it free on my own site, so the next person who needs it can open it and use it instead of paying an agent to make it again.
Building once and sharing the result is not a new idea. That is just software. The moment personal software is good enough to share, it turns back into a product.
What I think the future looks like
I don’t think software stays the same. I agree with the original reply that good software moves toward three things: a strong core, a headless service, and a fluid interface.
A strong core. High-quality software that takes a lot of time and money to make keeps its value. Agents make it cheaper to build the easy parts. They do not make a decade of hard-won correctness cheaper to copy.
A headless service. “Headless” means the product works without its own screen, so an agent can call it directly. This is already happening. Stripe hosts an official MCP server at mcp.stripe.com and also publishes an open-source local server through its Agent Toolkit. Linear also runs its own MCP server. I see the same thing in my own work. I built a demo where an OpenAI agent works inside a user’s GitHub through Kinde Connected Apps. The agent did not rebuild GitHub. It called GitHub. I also build Kinde components for Convex that agents use as auth and billing building blocks. In both cases the agent is a new kind of customer for the software, not a replacement for it.
A fluid interface. Human-facing software will let people reshape the screen with plain language, so each person gets a view that fits them. At I/O 2026, Google said AI Mode passed one billion monthly users and presented generative UI as the next step for Search. The interface becomes personal. The core under it does not need to be rebuilt by every user.
What is actually in trouble
I want to be honest about the real pressure. In the first week of February 2026, close to $1 trillion came off software and services stocks, and a Jefferies analyst named it the “SaaSpocalypse.” A big part of the fear is that agents weaken the per-seat pricing model that SaaS growth stood on for a decade.
That pressure is real, and some companies deserve it. If your product is a thin screen on top of a spreadsheet, an agent can replace you. If your pricing counts human seats while agents do the work, you need a new model. But that is a change in how software gets sold and used. It is not the death of software. The products with nothing inside them are in trouble. The products with a real core are about to get a lot more customers, and many of those customers will be agents.
Everyone got a camera, and photographers are still busy. Everyone got an air fryer, and the good restaurants are still full. Everyone is getting an agent that writes code, and I expect good software to do the same thing it has always done: get used by more people than ever.