DEFINITION
What is context engineering?
Previously: what is prompt engineering?
Asking a model a good question is no longer a profession. Deciding what the model must know before it answers is. Gartner named that shift explicitly in July 2025.
Context engineering is the discipline of assembling, ordering and bounding the information an AI model receives with every task: the instruction, the business knowledge, the tools it may reach, the memory of earlier steps and the limits within which it may act. Where prompt engineering was about how you phrase it, context engineering is about what the model carries with it when it answers. W69 AI Consultancy in Amstelveen builds that context layer underneath AI platforms and agent systems.
Gartner named context engineering in July 2025 as the discipline replacing prompt engineering. PwC asked 4,500 chief executives in January 2026 what AI had delivered: 56 per cent said they saw no revenue or savings yet.
Why prompt tricks no longer help
Something changed between 2023 and 2026 that many organisations have yet to act on: the model got better at thinking, and with that the bottleneck moved to what the model knows.
Models now reason on their own
Chain-of-thought was the art of making a model think step by step by explicitly asking it to. Reasoning models now do that unprompted, and better than you can coax out of them. The trick has not disappeared; it has moved from your prompt into the model.
There is no single prompt any more
An agent finishing a task takes dozens of steps, consults systems and hands work on. At every step it is decided anew what goes along. Calling that a prompt is looking at one sentence in a conversation that runs all day.
Sending more makes it worse
A large context window is a space, not an instruction to fill it. The more you send, the greater the chance the model attends to the wrong part. The craft is in leaving things out: sending only what this step requires.
Where the context comes from
Five sources are combined, step by step, into precisely what the model needs this time. Whatever does not fit stays reachable but does not travel.
Context engineering in four steps
This is work for an architect, not a copywriter. The order matters: start at step three and you are building boundaries around knowledge that does not exist yet.
Establish what the model needs to know
Not what it needs to do, which is the easy half. Walk through a real task and write down which facts a new colleague would have to look up to do it properly. Those are exactly the facts that must end up in the context. What you cannot write down, an agent cannot know.
Put the knowledge layer underneath
Without one place recording what is true at your organisation, every instruction remains an attempt to retell your company by hand to a model that will have forgotten it again next time. This is the step that costs the most work and prevents the most trouble.
Set the boundaries
Which tools the model may call, which data it may reach, when a human has to step in and what must never travel with it. Boundaries belong in the context, not in a policy document living next to the system. This is where governance and engineering meet.
Measure it, every single time
Record a fixed set of realistic tasks with the answer you expect, and run that set on every change: a different model, an adjusted instruction, a new connection. Without that set you do not know whether an agent that worked well last week still does.
So what happened to prompt engineering?
The techniques were not thrown away, they became part of something larger. Here is where they ended up, so you recognise them when someone still brings them up.
Zero-shot and few-shot
Asking a model something with no example, or with a handful of examples attached. This lives on, except it is now called assembling examples in the context. The question moved from whether you include examples to which ones, from what source and how current.
Chain-of-thought
Making a model reason step by step to avoid mistakes. With the models of 2023 you had to coax this out yourself. Reasoning models now do it unprompted. What you still decide is how much time and compute a model may spend on a question, and that is a cost decision.
System prompts and personas
Giving a model a role and rules of behaviour. This has been absorbed into the instruction part of the context, where it now sits alongside the knowledge layer and the boundaries. A persona without a knowledge layer beneath it is a model that sounds convincing while knowing nothing about you.
In short: prompt engineering was a skill that briefly looked like a profession. Now that models reason on their own and agents keep working independently, the question is no longer how you phrase it but what the system carries with it. That is architecture, and it starts with a knowledge layer.
What people ask about this
Prompt engineering is about how you phrase a question. Context engineering is about what the model carries with it at the moment it answers: the instruction, the business knowledge, the tools it may reach, the memory of earlier steps and the boundaries. The prompt has become one component of that. With an agent taking dozens of steps there is not even a single prompt any more; a context is assembled anew at every step.
No, it has been absorbed into a larger discipline. Phrasing things clearly remains useful, but it is no longer a job in its own right. Gartner stated in July 2025 that context engineering replaces prompt engineering as the discipline that matters to organisations. What disappeared is the idea that the right phrasing can make a model know something it does not. You solve that with knowledge, not word choice.
Chain-of-thought is the technique of asking a model to write out its reasoning step by step so it makes fewer mistakes. With the models of 2023 and 2024 you had to coax that out yourself. Reasoning models now do it unprompted and often better than you can ask for. The technique has not disappeared; it has moved from your prompt into the model.
No, and this is the most common mistake. The more you send, the greater the chance the model attends to the wrong part. A large context window is a space, not an instruction to fill it. The craft in context engineering is in leaving things out: sending only what this step requires, and keeping the rest reachable through the knowledge layer or through tools.
By measuring it rather than sensing it. Record a fixed set of realistic tasks with the answer you expect, and run that set again on every change: a different model, an adjusted instruction, a new connection. Without such a set you do not know whether an agent that worked well last week still works well this week, and that is exactly what an auditor will press you on.
With the knowledge, not the text. As long as there is no place recording what is true in your organisation, every instruction remains an attempt to type that in by hand. That place is called a knowledge layer. Without it, context engineering is retelling your own company over and over to a model that will have forgotten it again next time.
Does your AI know what your organisation knows?
The AI Navigator™ maps where you stand, including the point where it usually comes unstuck: the context your models are supposed to run on.