It's the Harness, not the Horse - on How AI Models are Commoditizing
There is a particular kind of person who, upon meeting a champion racehorse, immediately asks how fast it can run a quarter mile. Magnificent animal. Tremendous lungs. Genuinely the fastest thing on four legs you will ever stand beside. And also, if you hand it a field and a plow and ask it to feed a village, completely useless until somebody invents the harness.
We are all, collectively, standing in the paddock asking about the quarter mile.
For three years the entire discourse around artificial intelligence has been a horse race in the most literal sense. Which model is smartest. Which one nudged ahead on some benchmark by 1.3 points. Which lab dropped the new weights at midnight to ruin a competitor's launch morning. We have treated raw model capability as the whole game, as if intelligence were a substance you could bottle and sell by the liter, and whoever bottled the strongest stuff would win civilization.
This was always a slightly silly way to think about it. It is becoming an actively expensive one.
The horses are starting to look suspiciously similar
Here is the uncomfortable thing nobody at the foundation model labs wants printed on a billboard. The horses are converging. The gap between the best frontier model and the merely excellent one has gone from a canyon to a crack in the sidewalk. For the overwhelming majority of actual work that actual people need done, the question "which model" now produces a shrug, because four different ones will all do it competently and the cheapest one will do it on Tuesday for a tenth of the price.
This is what commoditization looks like. Not a single dramatic moment. Just a slow, grinding sameness, where the thing that was supposed to be your unrepeatable moat becomes a line item you renegotiate every quarter. Raw intelligence is going the way of bandwidth and storage and compute before it. Still essential. Still impressive. Increasingly something you rent by the token from whoever undercut the others this week.
The labs know this, which is why they have all quietly stopped selling you a brain and started selling you a worker. And a worker is not a brain. A worker is a brain plus a harness.
So what, exactly, is the harness
The harness is everything that turns a thing that knows into a thing that does.
A model on its own is a brilliant, amnesiac consultant locked in a sealed room with no phone, no internet, no memory of yesterday, and no hands. You can slide questions under the door and beautiful answers come back. Genuinely useful. Also fundamentally a brain in a jar. It cannot check whether it is right. It cannot look anything up. It cannot press a button, send an email, query a database, or notice that the thing it confidently told you ten seconds ago was nonsense.
The harness is the set of inventions that let it out of the jar.
The first piece is tools. Give the model the ability to call out to the world and it stops being a know-it-all and starts being a do-it-all. It can run code and see whether the code worked. It can search and read the result. It can hit an API, get a real number back, and base its next move on reality rather than on a confident hunch. This is the difference between a person describing how to ride a bicycle and a person riding one. The describing is impressive. The riding pays the bills.
The second piece, and this is the one the engineers get misty-eyed about, is the agentic loop. The loop is deceptively simple. The model makes a plan, takes an action, observes what happened, and then, crucially, decides what to do next based on what it just learned. Then it does it again. And again. Plan, act, observe, correct. It is the single most underwhelming-sounding idea in the entire field and also the one that changed everything.
Because here is the dirty secret of intelligence, biological or artificial. Most of what looks like genius is just error correction running fast enough that you do not see the errors. A model that answers once is taking a single exam under fluorescent lights with no calculator. A model in a loop is a craftsperson who measures twice, cuts once, looks at the cut, swears quietly, and adjusts. The first is graded on luck. The second is graded on persistence. Persistence wins, every time, and it does not even need to be the smartest one in the room.
The third piece is the plumbing, and this is where MCP earns its keep. For a while, connecting a model to your tools and your data was an artisanal nightmare. Every integration was hand-forged. Every company built the same fiddly connectors over and over, slightly differently, all of them brittle, none of them reusable. The Model Context Protocol is the deeply unglamorous decision to agree on a standard plug. That is it. That is the whole revolution. A common socket so that any model can talk to any tool without a team of engineers welding the two together by hand.
It is the USB of cognition, and yes, that comparison is unflattering, and yes, USB also quietly reorganized the entire physical world while everyone was busy arguing about processor speeds. Standards are boring. Standards also win wars. Ask anyone who bet against the shipping container.
Where the value actually went while you were watching the leaderboard
Follow the money and the story tells itself. The benchmark obsessives keep waiting for the one model so smart it makes everything else obsolete. Meanwhile the businesses quietly printing money are the ones that figured out the model was never the product. The harness is the product.
The value migrated, the way it always does, away from the raw ingredient and toward the thing that makes the raw ingredient useful in your specific life. Nobody pays a premium for flour. People pay a premium for the bakery. The model is flour now. Astonishing, world-changing flour, available from sixty suppliers, fungible by the kilo. What people will actually pay for is the system that knows their data, remembers their last conversation, plugs into their messy real tools, runs in a loop until the job is genuinely done, and hands back something they can use without checking every line.
That system is mostly harness now. The horse inside it is interchangeable, and increasingly, everyone knows it.
The part where intellectual work quietly stopped being what it was
Strip away the hype and the doom in equal measure and here is what actually happened to knowledge work. For a few centuries the bottleneck on intellectual labor was the supply of capable minds and the hours they could stay awake. You wanted more analysis, more code, more research, more drafts, you hired more brilliant tired people, supplied them with coffee and waited.
The harness broke that bottleneck, and it did not break it by making one superhuman genius. It broke it by making competent, tireless, looping doers cheap and plentiful and able to actually touch the world. The leverage did not come from intelligence getting deeper. It came from intelligence getting hands, getting memory, getting a standard way to plug into everything, and getting the patience to try, fail, notice, and try again without ever getting bored or asking for a raise (until tokenomics takes over that is).
This is genuinely a big deal, and it is also exactly the opposite of the story we were all told. We were promised a single oracle of terrifying brilliance. We got, instead, an enormous number of pretty-good workers in well-designed harnesses, and it turns out that is what reorganizes an economy. Not the lone genius. The fleet of competent doers that never sleeps.
The takeaway, for those keeping score in the wrong column
If you are an investor still funding the forty-seventh foundation model on the theory that this one will be 4 percent smarter, consider the possibility that you have bought the world's most expensive racehorse to enter in a race that quietly turned into a logistics company.
If you are a builder, the lesson is almost insultingly practical. Stop chasing the model. The model is a commodity you will swap out twice a year regardless. Pour yourself into the harness. Into the tools, the loop, the memory, the standard connections, the unglamorous plumbing that turns a clever guesser into a dependable colleague. That is where the moat is, because that is where your actual users, your actual data, and your actual workflows live, and none of those swap out for a tenth of a price on Tuesday.
And if you are just a person watching all of this with a healthy mix of awe and dread, here is the genuinely reassuring part. The thing reshaping your work was never a mind so vast it left you behind. It was a harness. Harnesses are built by people, improved by people, pointed at problems chosen by people. The horse runs fast. Somebody still has to decide where the field is, why it needs plowing, and what gets planted.
That somebody, for now, is still you. Hold onto the reins.