Return to Sender

In the fall of 2005, the Commonwealth of Massachusetts tried to kill a file extension.

It was not a frivolous attempt. The state's IT division, under CIO Peter Quinn, announced that executive-branch agencies would standardize on open document formats, on the entirely reasonable grounds that public records locked inside a proprietary format are not really public. Microsoft's format was named as forbidden. The practical effect was that Massachusetts agencies could not upgrade to the forthcoming Office 2007. The state had procurement power, a public-interest argument that was difficult to argue against, an international standards body behind the alternative, and the full attention of every open-source advocate in the country.

By July 2007 the same IT division had revised the policy to recognize Microsoft's OOXML as an acceptable open format. The ban was over. A state government, holding every card that mattered, lost to a file extension in under two years.

I have been thinking about that story since September 16, when Anthropic folded its agentic Cowork product into the main Claude app and shipped Claude Docs and Claude Slides alongside it. The coverage arrived within hours and it arrived in one shape: this is a shot at Microsoft. Axios ran it as raising the stakes for Microsoft. Computerworld framed it as making Claude stickier. The consensus formed fast and it formed around a question nobody bothered to finish asking.

A shot at Word. Fine. But at which part of Word?

Start with the thing that makes the war metaphor collapse on contact: Anthropic sells Microsoft the gun.

Claude models run inside Microsoft 365 Copilot — Microsoft announced them itself, across Copilot Studio, Power Platform, and agent modes in the Office applications — underneath a reported thirty-billion-dollar Azure compute commitment. Anthropic shipped a Claude integration into Word before it shipped anything resembling a Word competitor. The challenger rents its compute from the incumbent, sells the incumbent its engine, and competes with the incumbent at the application layer, all in the same quarter. Whatever that is, it is not Lexington and Concord. It is a supply relationship with a product overlap, and the two firms are each other's vendor, customer and rival at three different altitudes simultaneously.

This matters beyond the pleasure of catching a metaphor being lazy, because the war framing smuggles in an assumption: that there is a single thing called Word and that somebody is going to win it. There isn't. Word is at least four separate businesses wearing one icon. There is the editor — the surface where text gets typed and formatted. There is the application — the specific program you had to acquire in order to touch the file. There is the distribution — which is not really Word at all but Microsoft 365, a bundle whose actual purchase decision is about identity, mail, device management and compliance, with the word processor riding along as a rounding error. And there is the format — .docx, the thing your counterparty opens, your regulator accepts and your court admits.

Claude Docs is a genuinely good assault on the first of those. It is not aimed at the other three, and the history says the other three are where Word actually lives. To see why, it helps to have a name for the distinction.

A production layer is anything whose value is a function of how hard it is to make. Writing code. Drafting a document. Rendering an image. Laying out a page. Production layers are priced by scarcity of capability, which means they are exactly what a supply shock destroys — that is the entire argument of the means-of-production piece, and the reason the producer surplus in commoditized work gets competed away.

A coordination layer is anything whose value is a function of how many other people already use it. A file format. A protocol. A unit of measurement. An instruction set. A plug shape. Coordination layers are not priced by how good they are, and this is the part that trips up almost everyone who has ever tried to displace one: a better coordination layer does not win. Being superior is not the variable. Being already-agreed-upon is the variable, and the cost of changing it scales with the number of parties who would have to change at the same moment.

The cleanest illustration of this I know sits about two hours south of here. In 1890, a chemistry professor at the University of Wisconsin named Stephen Babcock solved a problem that was quietly strangling American dairy. Milk was bought by weight, which meant a farmer who watered his milk or skimmed the cream got paid the same as a farmer who didn't. Honesty was a competitive disadvantage. Babcock's test was a cheap, fast, repeatable way to measure butterfat, and once you could measure butterfat you could price milk by what was actually in it. Wisconsin's dairy dominance is usually told as a story about cows, climate and immigrant cheesemakers. It is at least as much a story about a measurement standard.

And here is the detail that makes it a parable rather than an anecdote: Babcock refused to patent it. He gave it away, it won grand prizes at the Paris and St. Louis expositions, and it became the thing every creamery and every buyer in the country agreed on. He understood something about the layer he was building that the enclosure instinct gets wrong. The test's power came from universal adoption, and charging for it would have reduced adoption, which would have reduced its power. A coordination layer earns its moat by being everywhere, not by being owned.

Now put the two layers next to each other and the thesis writes itself. AI is a supply shock to production. It makes things cheaper to make. Agreement was never expensive to make — agreement is expensive to change, and nothing about a cheaper writing tool touches the coordination cost of persuading every counterparty, regulator, court and archive on earth to accept a different file on Tuesday. The shock lands on the layer priced by production cost and rolls straight past the layer priced by consensus.

A thesis this tidy deserves to be run against the record rather than asserted. So here are four moments when the cost of producing software or documents fell off a cliff, scored against the same four layers each time. Flip through them.

Take them in order, because each one teaches a slightly different lesson.

1969. IBM unbundles under antitrust pressure on June 23, software gets an invoice line for the first time, and the independent software industry is born — a story this blog has already told at length. Production, application and distribution all get rearranged violently. The System/360 instruction set does not move, and its direct descendants are still executing inside IBM mainframes today, sixty-two years later. The business model got demolished by the Justice Department. The coordination layer underneath it shrugged.

1995. Netscape promises to reduce Windows to “a poorly debugged set of device drivers.” The line is usually credited to Marc Andreessen, though he later attributed it to Bob Metcalfe and described his own version as a retweet, which is a very 2012 way to cite a 1995 insult. The prediction was directionally correct and roughly twenty-five years slow — precisely the pattern in the prediction-problem post. The browser did become the application platform. Windows did become substrate. But notice what actually happened at the bottom of the stack: HTTP and HTML were not commoditized. They became the new coordination layer, and the reason the web beat every proprietary online service of that era is that it was an unowned standard in the Babcock tradition.

2006. Google buys Writely and ships Docs and Spreadsheets, and every reasonable person declares Word dead. Two decades later the scoreboard is a denominator mess of exactly the kind we took apart in the AI-debt piece: counted by domains, Google Workspace leads on roughly 50 percent to Microsoft 365's 45; counted by enterprise seats and revenue, Microsoft holds something like 58 percent of the enterprise segment, around 446 million paid seats, and about three-quarters of the Fortune 500. Both numbers are real and they are measuring different things. Google won the count. Microsoft kept the enterprise, the money and the format — and the free editor that was supposed to end Word instead ended paying for an editor, which was never the business.

That is the same window in which Massachusetts tried the direct approach. Not competition — legislation. A sovereign buyer with mandate authority declared the format non-compliant, and the format won anyway, because the state discovered what every challenger discovers: it was not fighting Microsoft, it was fighting every other party its agencies had to exchange documents with, none of whom were subject to Massachusetts procurement policy. You cannot unilaterally leave a coordination equilibrium. That is the definition of one.

2026. Which brings us to now, and to the row that looks different. Production has collapsed hardest of any of the four — drafting is now the cheapest step in producing a document rather than the expensive one. The application layer is under genuine pressure for the first time, and for a reason worth stating precisely: if an agent can write the format directly, the program that existed to let a human manipulate the format becomes optional. That is a real threat and it is not the threat the headlines described. Distribution held, because an admin toggle beats a procurement cycle. And the floor did not move at all.

Run the distinction one layer down and it produces a claim that sounds wrong until you hold it still.

The current excitement about agentic desktop tools — Cowork and its competitors working directly in your files — reliably produces some version of the question: does this eventually replace the operating system? It is the Netscape prediction with new staging, and the answer is no, but the interesting part is why no. A kernel is a coordination layer, and the applications running on top of it are production layers. Scheduling, memory, the driver ABI, the syscall interface, the hardware contracts, the certification regimes — none of that is valuable because it is hard to write. Plenty of people can write a kernel. It is valuable because every driver, every binary, every peripheral and every compliance auditor has already agreed to it.

The applications are the opposite. A word processor, a PDF editor, a mid-tier vertical desktop tool — these are priced by how hard they were to build, which is precisely the quantity collapsing. So the prediction that falls out of this framework is the reverse of the popular one: Windows and macOS survive this turn comfortably by becoming more boring, while the application layer sitting on them gets hollowed out. The OS is atoms-adjacent in the sense that matters — hardware-coupled, certification-bound, brutally path-dependent. The apps are pure bits, and AI is a supply shock to bits.

There are also plain mechanical reasons an agent cannot be an OS, and they are worth keeping in your pocket for the next time someone suggests otherwise. An operating system answers in microseconds and is deterministic; an agent answers in seconds and is probabilistic, and you cannot build a window manager on a thing that occasionally does something else. Agent sessions run remotely and meter their consumption against a rolling window. OS operations are free at the margin. You cannot meter ls.

What agents are credibly contesting is the shell — the surface where a human states intent and work comes back. That is a genuinely valuable position, historically the position that captures the user relationship while everything underneath commoditizes into plumbing. It is worth fighting over. It is just not the operating system, and calling it the operating system makes a defensible claim easy to dismiss.

1. A coordination layer only holds while translation is expensive. This is the one that could actually kill the thesis, and I do not have a complete answer to it. The reason everyone must agree on a format is that converting between formats is lossy, fiddly and unreliable. Remove that and you remove the need for agreement — not by storming the standard but by making it unnecessary. If a model reads and writes every format flawlessly, including the parts that historically broke (tracked changes, comment threads, deep styles, embedded objects), then “everyone uses .docx” stops being a constraint and becomes a preference. Coordination moats do not usually get attacked successfully. They get dissolved by universal translation. The counter-worry to the counter-worry: the last thirty years of format conversion suggests the final ten percent of fidelity is where all the difficulty lives, and legal redlining is exactly the workflow that cannot tolerate a lossy round trip.

2. Babcock cuts both ways. I used an unowned standard to make the case, and .docx is not one — it is a standard with a corporate parent and a commercial interest in where it goes next. By my own argument that makes it weaker than the butterfat test, not stronger. An owned coordination layer has a defector problem the unowned kind does not: every other party knows the owner can change the terms, which is precisely why ODF existed in the first place. The honest version of my claim is narrower than the clean one. Formats with corporate parents are durable because of inertia and installed base, not because they have the deep structural immunity a genuinely public standard has.

3. Distribution is attackable, just not by this attacker. The one moment Microsoft's bundle actually bled was when Google attacked it from the identity side — free mail plus an identity people already had, with the suite attached. That worked. Anthropic does not sell identity, mail, calendar, storage or device management, which means it is attacking the most defensible configuration from the least advantaged position. But someone who does sell those things could partner their way into the same move, and a separate AI budget line that routes around the annual seat renewal is a quieter version of the same attack.

4. This might be a category shift, in which case I am arguing about buggy whips. Satya Nadella's line about business applications collapsing into “CRUD databases with a bunch of business logic” in the agent era applies just as well to documents. A document is a container that exists because humans needed something portable to pass around and mark up. If the unit of work becomes a tracked state, a decision, an answer — something an agent maintains rather than something a person mails — then arguing about who wins documents is arguing about the wrong noun entirely. I think this is slower than its advocates believe, for the reason every prediction in this genre is slow. I do not think it is wrong.

The practical reading, for anyone who has to make a decision rather than enjoy an argument, is a sorting exercise. Look at what your organization does and ask of each piece: is this priced by how hard it is to do, or by how many other people have already agreed to it?

Everything in the first pile is on the clock. That is not a reason to panic, but it is a reason to stop treating capability as a moat — the capability is being handed out at list price to everyone including your competitors, which is the Red Queen problem and the reason adoption is table stakes rather than an edge. Everything in the second pile is worth far more than your org chart currently says it is, and is probably not even on the balance sheet. The interchange format your industry runs on. The data schema your partners integrate against. The certification only you hold. The measurement everyone in your category quotes. Those are Babcock tests, and most companies own at least one without having noticed.

The mistake is not adopting AI too slowly or too fast. The mistake is spending the adoption on the production layer, where the returns get competed away, while leaving the coordination layer unattended — or worse, trading it away for convenience, which is what happens every time an organization abandons its own schema to fit inside somebody else's.

Massachusetts had the better format, the better argument and the power to mandate it, and got it returned to sender inside of two years, because a format is not a document standard — it is a treaty, and you cannot sign one alone. Anthropic is not going to beat that with a better editor. Neither is anyone else. The editor was never what was being defended.

When the cost of making something falls to zero, stop asking who makes it best. Start asking what everyone already agreed to, and who owns that.

This post extends an argument this blog has been circling: Own the Means of Production on where value goes when supply shocks hit, Full Circle on the sixty-year wheel these turns ride, Software Is Dead on why the direction is right and the timing is always slow, and Ch-Ch-Changes on what it costs to be locked to the layer that keeps moving.