Own the data, rent the methods

Look at how fast each layer turns over. Frontier models ship every few months, and the lead changes hands between providers more or less every time. Training methods churn just as fast: reinforcement learning from human feedback gave way to direct preference methods, which will give way to something else, and the optimizer underneath is a parameter you swap, not a commitment you make. Agent frameworks have a shelf life measured in months. This year's orchestration library is next year's migration.

If your advantage is that you use the best of any of these, you are renting your advantage, and the lease is short. The moment a better model or method or framework arrives, your edge resets to whatever everyone else can also adopt that week. Renting is fine, even necessary. It is just a bad place to keep your moat.

Concretely: a competitor can buy the same frontier model you use, install the same open-source framework, and read the same paper on the same training method, all in an afternoon. Whatever advantage lived in those choices is now shared. None of that afternoon buys them the record of how your company actually does its work.

One layer does not churn: the data produced by doing the work. It comes in three kinds, and none of them is generated by a method.

Execution traces are the complete record of how a task was done: what context was retrieved, what the agent did, what a human corrected, and what was finally accepted. Context graphs are the growing map of which data and which decisions matter for which work. Expert preference data is every correction, acceptance, and rubric grade from the people who know what good means. None of it comes out of an algorithm. It comes out of doing the work, and it accumulates. A year of it is very hard to replicate, because you cannot go back and buy last year's decisions.

And it is not only a matter of volume. The data is specific to one company's work, so even a competitor sitting on more total data does not have this data: the decisions and corrections of these teams, on these accounts, under these standards. It is non-fungible by construction. There is no market where you can purchase the thing, because the only way to produce it is to have been in the room.

Notice the relationship between the two. Every method in the stack consumes data as input. A training method needs labeled examples. A retrieval system needs a corpus and feedback on what was actually useful. An evaluation needs a definition of good to measure against. The methods are hungry, and they are interchangeable. What they are hungry for is the scarce thing.

So the edge is not in having the best consumer. Everyone will have a good consumer soon enough, because consumers are what the field ships. The real edge is in owning the supply, the input every one of those methods needs and none of them can manufacture on its own.

This is a design principle, not a slogan, and it shows up in how the platform is built. Wherever a layer has a durable part and a swappable part, the durability goes in the data and the method becomes a pluggable backend.

The institutional knowledge an agent accumulates is the durable asset; how it is retrieved and presented to the next agent is a layer we can replace without touching the store underneath. The expert rubrics and the real task trajectories are the durable asset; how they are compiled into a reward model is a backend we can swap as the training frontier moves. When the next method arrives, and it will, it plugs into data that is already there, already richer than a competitor's, and the switch costs nothing structural.

Owning the data does not mean ignoring the methods. The opposite. Because the advantage does not live in any one model or framework, you are free to adopt the best one the week it ships, with no lock-in to defend and no sunk cost to protect.

The company that depends on a method hesitates to leave it. The company that depends on data switches methods without a second thought, and routes its work to whatever is best right now. Renting well is an advantage in itself. It is only dangerous when renting is also where you keep your moat.

Here is the part that runs against intuition. Fast method churn is good for this strategy, not bad. Every time the field invents a better method, a company anchored to methods has to re-tool around it. A company anchored to data just swaps a backend and runs the new method over a larger pile of the scarce input than anyone else has.

The faster the methods improve, the more the advantage accrues to whoever owns the data they all need. You want to be the landlord in a town where everyone keeps building newer houses. The houses get better every year, and you still own the land they all sit on.

Models, frameworks, and training methods will keep changing, faster than anyone can fully productize. The execution data, the traces, the context, the corrections, does not change. It only accumulates, and it cannot be bought after the fact, because it is a record of decisions that were made while real work was happening and are gone the moment they are not captured. Own the data. Rent the methods. The methods will date. The data compounds.