Two years ago the rule was: whoever can write good prompts wins. Today in projects we almost always observe the opposite — the best prompts fail because of bad context, while mediocre prompts shine with good context.

From sentence structure to architecture

Context engineering is the art of providing an AI system, at the right time, exactly the information it needs for its task — and nothing beyond that. That sounds trivial but is an architecture topic: which documents, database extracts, tool descriptions and history end up in the model's context window? In what order, at what density? Modern models can process huge contexts, but research and practice show the same phenomenon: what sits in the middle of an overfilled context tends to be overlooked. More is not better. More precise is better.

Why agents sharpen the issue

With a single chat prompt the model forgives a lot. An agent that executes dozens of steps over hours forgives nothing: every superfluous block of text travels through every follow-up action, dilutes the model's attention and costs money on the side. In our projects a principle has therefore become established: an agent's context is curated like a good handover to a new colleague — the essentials, up to date, findable, without legacy baggage. This includes clean knowledge sources, clear tool descriptions and consistent summarisation of long histories.

A craft, not a trick

The good news: context engineering can be learned and measured. You see the effect directly in success rates, processing times and cost per case. It is the least spectacular topic of our series — and perhaps the one with the greatest leverage.

If you want your AI solution to seem smarter without switching the model: the answer often lies in the context. We are happy to take a look together.