On September 25, Thomas Ptacek published a post on sockpuppet.org titled “What Even Is An OS Now?” It opened with the news that he is leaving Fly.io — the company he co-founded — to build a phone with Kurt Mackey. It spent the day on the front page of Hacker News, where 174 comments accumulated under it. And buried in the middle was the line everyone quoted: “Every mainstream phone on the market in 2026 was planned (torturiously, to death) in 2023.”
Read the reaction and you’d think he’d found a crack in Apple’s wall. Incumbents are frozen in three-year roadmaps. The app grid is a 2008 artifact. The AI-native operating system is up for grabs, and the person who grabs it will be someone who wasn’t in the room in 2023.
That’s the consensus. It is also, I think, backwards — and Ptacek’s own post contains most of the reason why.
The Three-Year Cycle Isn’t a Bug You Can Route Around
The received interpretation of that 2023 line is that the incumbents are behind, and behind means vulnerable. But the three-year cycle isn’t a failure of imagination. It is the operating system.
An OS is a compatibility promise issued simultaneously to silicon vendors, carriers, developers, payment networks, accessibility regulators, and the enterprise IT departments that have ten thousand devices in the field and no appetite for a surprise. Its defining feature is that it does not move at the speed of anything. AI capability turns over every six months, which is exactly why the one layer that must not turn over every six months is the layer least likely to be replaced by one.
You can ship an AI-native assistant, an AI-native browser, an AI-native IDE. “AI-native OS” is a category error — like a car that is also a road.
Four Things Only the OS Can Hand Out
Here is what an agent that acts on your behalf actually needs: a hardware root of trust that some server will believe, an identity a counterparty will accept without a human in the loop, a payment credential, and the standing right to refuse an app. Every one of those four is issued by the operating system. Nothing else can issue them.
Now notice what more capable AI does to the value of those four things. It raises it. The better your agent gets, the more consequential it becomes who signs its credentials and who can veto what it installs.
Which is why the hard part of building a phone is not the kernel. It’s eSIM provisioning across the carrier labs, FCC equipment authorization, a secure element the payment networks will certify, and a developer ecosystem whose members have already amortized their software over a billion-plus devices.
“I’ve watched a demo device fail certification over how it handled a profile switch during a roaming handoff,” a certification engineer at one of the big two carriers’ device-testing labs said, standing in the corridor while the next test rig was loaded. “Nobody in that room was talking about the operating system. The OS was fine. The OS was the easy part.”
The Squeeze Is in the Middle of the Stack
If the OS becomes the agent’s principal, the loser is not Apple. The loser is every party that currently owns a customer relationship without owning the glass — airlines, banks, retailers, insurers, the whole subscription economy built on the assumption that you open an app to receive a service. When the agent does the opening, the brand in the middle becomes a tool call with a price attached.
That is the actual disruption story, and it is not the one the disintermediation crowd is selling. They are modeling the operating system as a technical artifact losing relevance. It is a rent-collection mechanism acquiring a new tenant, while the tenant’s needs make the mechanism more indispensable rather than less.
The people who should be updating their priors are not incumbent product managers. They are the ones who have spent two years telling limited partners that platforms are commoditizing.
The Part of the Post That’s Actually Alarming
Give Ptacek credit for the honesty of writing that he is “talking his book” — that he is building one of these things, so of course he thinks the category is open. But notice what the genre requires. The “the OS is up for grabs” argument is only ever made by people who need it to be up for grabs.
His earlier post this year on LLM-generated vulnerability research, which LWN.net picked up, is the more interesting piece of writing, because it names a cost with no revenue attached to it. If agents write exploits faster than anyone can write patches, then the operating system’s other job — the boring one, shipping security updates to a billion devices for a decade — gets harder and less fundable at precisely the moment it matters most.
That is not a platform shift. That is a bill. And nobody is raising a Series A to pay it.
Build the phone. Build several. But spare me the eulogy for the operating system, because the operating system is the only thing standing between your agent and a counterparty who has to decide whether to trust it. The more autonomous the software gets, the more of our lives get routed through the layer that can say no on our behalf — and that layer has a three-year roadmap, a certification lab, and a hardware key stored somewhere you will never see.
The incumbents are not behind. They are downstream of the only bottleneck that AI cannot compress.