Context, not the model, is the ceiling of your AI agent

The test that settles it takes five minutes. Take a case it got wrong, and manually supply the model with everything a competent colleague would have had in front of them. If it answers correctly, the subject is context. This test saves weeks of migrating models to no effect.

The most useful comparison is not the brain, it is onboarding a new joiner. You do not expect a brilliant recruit to guess your pricing structure, your commercial exceptions, or the name of the client who must not be chased on a Friday. You tell them, you give them access to the right tools, and you review their work at the start.

An agent is in the same position, with two differences: it never asks for clarification when something is missing, and it shows no sign of hesitation. Where a new hire comes and asks a question, the agent fills the gap. That is exactly why an incomplete context produces confident errors rather than requests for clarification.

Context windows are growing fast. In August 2026, Pokee AI released Pokee-Isaac 28B, a 28-billion-parameter agentic model announced with a window of up to 10 million tokens and presented as deployable on a single GPU such as an RTX 4090. Its stated commercial argument is explicit: running inside the customer boundary.

A necessary caveat: the long-context results published — around 95 on the RULER benchmark between 256K and 4M tokens — come from the vendor's own internal evaluations, published in its own technical report. We are not aware of independent reproduction. It is a claim, not a verified fact, and this article does not treat it as one.

The interest for a European company is not the number. It is the combination of modest size plus very long context plus local execution: it makes credible a scenario that was not credible a year ago, namely an agent reasoning over a substantial document set without those documents leaving your infrastructure. For sectors where data cannot leave, that matters far more than any benchmark point.

Beware the shortcut "big window, so no selection needed". Sending your whole corpus on every request is expensive, slows responses and dilutes the useful signal. A serious architecture selects, whatever the window size — which is what we set out in our guide to RAG for the enterprise.

An agent given access to every client file so it can "be more helpful" can cite one client to another. The answer is not to add a discretion instruction to the prompt — an instruction in natural language is not an access control. Partitioning happens at the permission and query layer, upstream of the model.

The same reasoning applies to inbound content: a document uploaded by a third party enters the context and may contain instructions. That is the injection mechanism covered in our analysis of AI agent scope and permissions.

• Run the five-minute test on three real errors before considering any model change.
• Write the agent's job description: what it must know, where that lives, who keeps it current.
• Hunt down stale sources. An obsolete document still indexed does more damage than a missing one.
• Partition through access, never through the prompt.
• Measure on your own cases. Public benchmarks say nothing about your business; twenty annotated real cases say a great deal.

This is the core of our work: wiring an agent to the right sources, with the right partitioning, hosted in Europe. See our custom AI agents and, on choosing self-hostable models, our comparison of the best open-source LLMs.