Context is the new defensibility

What do you believe about the future that others do not?

I was asked this question recently as part of an application for a VC position. The answer has kept growing on me, so I’m writing it down here in longer form.

The model race is already over. Not because progress has stopped, but because it no longer matters in the way most people think. A model from two years ago was already capable enough to power most products for the next decade. The real competition was never about who builds the smartest model. It was always about who captures the most context.

Context is the new defensibility.

Most people still use AI as a sophisticated search engine: prompting, asking, retrieving. A smaller group has figured out how to make it an extension of themselves. The difference between those two groups is not intelligence or technical skill. It is context. The people getting superhuman output are the ones who have invested in feeding the model their world: their thinking, their work, their decisions. The model did not get smarter. It got more situated.

This is why platforms of registration — the systems that hold data, documents, and institutional memory — are suddenly the most strategically important layer in tech. Not the AI labs. Not the foundation models. The systems that know the most about you, your company, and your customers. Context compounds. Models are commodities.

Why legacy can’t catch up

Here is where I diverge from the consensus: most people understand this problem but underestimate how structurally hard it is to solve — and more importantly, they are missing where the solution is already emerging.

Legacy software captured context accidentally. It was a byproduct of people doing their jobs inside a system. Retrofitting that context for AI is an enormous, expensive, often impossible undertaking. The data is siloed, inconsistent, and was never designed to be machine-readable in a meaningful way.

I’m living both sides of this right now.

On one side: a production management platform I’m building as a freelance project, using coding agents. I don’t know how to code in this stack — not in the way a trained engineer does. And yet the velocity is absurd, and I’m holding several streams in parallel. What changed is not my technical skill. It’s that I’ve gotten better at providing context — what the business is, how it works, what matters — and letting the agent translate that into software.

On the other side: a codebase I recently stepped into, where documentation is outdated and the important decisions live in people’s heads. And here is the part I don’t get to dodge — when I left my previous job, the handover documentation I wrote wasn’t good enough either. Not because I was careless. Because context that accumulates accidentally, in a thousand micro-decisions nobody thought to record, cannot be retrofitted after the fact. You either capture it as you build, or you lose most of it forever.

That asymmetry is the whole argument.

AI-native is structurally different

Something different is happening now. A new generation of products is being built from zero with AI — not just AI-assisted, but AI-generated, with humans acting as architects, designers, and context providers. Functional domain experts capture what the business actually is, how it actually works, what actually matters. AI translates that into product. The result is software that is not just functional. It is structurally legible to machines, because machines built it.

This is why service-as-software is the most interesting idea in enterprise right now. Not because it lowers development costs, though it does, but because it breaks the FTE-to-output constraint while simultaneously creating a context-rich product that gets more valuable over time. The business logic is not buried in someone’s head or scattered across legacy code. It is encoded, explicit, and ready.

And that is the setup for the agentic wave. Agents do not work well in opaque environments. They need to understand the systems they operate in. AI-native products, built by AI, are the first software that agents will actually be able to work with, not around.

Most people are watching the models. The real game is being played one layer below.