An open context layer

Strip an agent down and what makes it useful is not the model, those are increasingly a commodity, a few interchangeable providers a config change apart. What makes it useful is what it can see: your data, your documents, your systems, the record of what it has done before. A frontier model with no access to your world is a very expensive autocomplete. The context layer is where the value is, and, not coincidentally, where lock-in likes to hide.

You do not own the model. What you own is the context, the connected, accumulated record of your own work.

What the context layer is

It is three things that get flattened into one word. There is the window, the tokens in front of the model right now, finite and ephemeral. There is the corpus, the documents, records, and live systems the agent can reach. And there is the accumulated record, what this agent has learned from doing the work, the small high-value memory almost nobody builds. Most stacks wire up the first two and skip the third, which is why an agent that has run a task four hundred times approaches the four hundred and first exactly as it approached the first.

Context is the multiplier, and it is measurable

That context makes a model useful is usually asserted. It has now also been measured, on a task built to isolate exactly this variable. Researchers at Stanford took a frontier model, Gemini 2.5 Pro, and asked it to reproduce the confidence that climate scientists assigned to statements in the IPCC's Sixth Assessment Report, a four-way classification where random guessing scores 25 percent. Reasoning alone, the model scored 48.7 percent. The same model, handed retrieved evidence for each statement, scored 89.3 percent. Cohen's kappa, which corrects for lucky guesses, moved from 0.32 to 0.86.

The detail that matters for anyone deciding where to spend is what did not work. The same authors swept the reasoning budget, the fashionable lever, giving the model more tokens to think. It did not close the gap, and past a modest point it made calibration worse. Their conclusion, in their words, is that "information access, rather than reasoning depth or inference budget, may be the critical bottleneck" for this class of task. One benchmark in one domain is not a law of nature, and calibration is a narrow slice of what agents do. But the shape of the result is the thesis stated as a number: the cheapest way to make a model useful on your problem is usually not a bigger model or a longer think, it is the right context in front of it.

Why open matters more than large

The industry keeps selling bigger windows as the answer, and bigger is genuinely useful. But it answers the wrong question. The one that decides whether an agent is useful in your business is not "how much fits in the prompt", it is "what can it reach, and can you move it." That is a question about openness, not size:

• Connect to any source. Your database, your files, your SaaS tools, your internal APIs. Context that stops at the platform's own walls is a demo of someone else's data, not yours.
• Model-agnostic. The same context serves whatever model you point at it. When a better or cheaper model ships next quarter, switching is a config change with a number attached, not a rebuild, because the value lived in the context, not the coupling to one provider.
• Self-hostable. The context layer holds your most sensitive material by definition. Being able to run it on your own infrastructure is the difference between "we can use this" and "legal said no."
• Yours to leave with. The accumulated record is an asset. If you cannot export it, it was never really yours.

If context is the multiplier, the question of who controls it becomes the whole game. The encouraging development is that the interface for connecting models to context is standardizing in the open rather than fragmenting into proprietary silos. The Model Context Protocol, an open standard for exposing data and tools to any model, was open-sourced by Anthropic in late 2024 and adopted by its two largest competitors inside five months.

That matters because it decouples the two decisions vendors prefer to sell as one. The connection to your world can be a portable standard, while the model behind it stays a swappable commodity. The lock-in that remains is the older, cruder kind: the cost of physically moving the data you have accumulated.

Egress is a lock-in instrument, and regulators now treat it as one. Moving data out of AWS to the internet runs about $0.09 per GB for the first 10 TB each month, a charge that scales with exactly the asset you would most want to take with you. The EU Data Act makes the point in law: cloud switching charges, including egress incurred when leaving a provider, are being phased out and are fully prohibited from January 12, 2027. Context you cannot move cheaply is context you do not fully own.

Retrieval is a judgment problem

Once the corpus is bigger than the window, something has to choose what goes in, and that choice is the actual product. Semantic similarity is a reasonable first pass and a poor final answer, the most similar document is often not the most useful one, and the most useful one is frequently a short note that says "this approach failed last time, here is why." Good selection weighs recency, provenance, and outcome, which means it has to be learned from results. Context and proof turn out to be the same problem wearing different clothes.

The asset you actually build

Here is the honest answer to what a company owns when it builds on someone else's model. Not the weights. Not the architecture. What it owns is the accumulated context: the structured, retrievable record of its own work, getting more valuable every time the system runs. Models improve on the provider's schedule. Your context improves on yours. Over a decade, the organizations that treated their operational record as an asset from day one will be the ones that got the most out of all of this, and that is very hard to do on infrastructure you cannot see into or take with you.

How we think about it

Recursiv treats context as open infrastructure, not a prompt trick. Connect any data source; point it at any model through a provider-agnostic layer; run it managed or fully self-hosted; export it whenever you want. Memory is a stored, queryable thing with an owner and a scope, and agents write to it deliberately, including their own evaluations of how the work went. The window will keep growing, and that is great. It just answers a different question than the one your business is actually asking.