Agentic AI · Service Design

What is context
engineering?

An AI model remembers nothing and sees only what it is shown. Context engineering is deciding what that is — and what it must never be. A plain-language explainer for business and service designers, about the quietest experience decision in AI.

Read the explainer Skip to the design angle

New vocabulary from the agentic AI deep dive

Part 1

The model sees one thing: its context window

Why the quality of an answer is decided before the model says a word.

An AI model has no memory of its own. It does not remember yesterday’s conversation, it has not read your knowledge base, and it does not know who the customer is. Every single time it answers, everything it “knows” in that moment arrives in one package of text: the context window. Its standing instructions, the conversation so far, documents fetched for the occasion, the customer’s history, and finally the question itself — all stacked into one bundle, read fresh, every time.

Nothing outside that window exists for the model. Ask two systems built on the same model the same question and you can get a generic shrug from one and a precise, personal answer from the other — the difference is entirely in what each window contained. That is why teams stopped obsessing over the wording of questions and started engineering the window itself.

In one sentence: context engineering is deciding what an AI system gets to see at each moment — and what it doesn’t — so that its answer is grounded in the right information.
The context window Instructions Conversation so far Retrieved documents Customer history Your question read fresh, every time AI model no memory of its own The answer only as good as the window If it isn’t in the window, it doesn’t exist for the model.
Fig. 1 — One moment, one window: everything the model knows arrives in a single package.

Part 2

The craft is selection, not accumulation

The obvious instinct is to give the model everything — the whole knowledge base, the full history, every policy. It doesn’t work. The window is finite, and attention dilutes: bury the one detail that matters under forty pages of the ones that don’t, and the model will miss it. Too little context and the answers turn generic, or worse, invented. Stale or wrong context is the most dangerous of all — the model will use it with complete confidence.

So context engineering is editorial work. Not “how much can we give it?” but “what does this moment need?” Four decisions come back at every step:

Selection

What earns a place in the window right now — and what is noise that will drown the signal?

Order

What should the model read first, and what deserves the emphasis — the policy, or the customer’s last message?

Freshness

Is it still true? An outdated price list in the window becomes a confident wrong answer out of it.

Boundaries

What must never enter — private data the moment doesn’t need, history the customer wouldn’t expect it to have.

Part 3

In an agent, the harness sets the table

With a chatbot, you curate the window yourself without noticing — you type the background, you paste the document. An agent working towards a goal has no such luxury: it takes many steps, and at every step its window has to be rebuilt — the goal, what has been done so far, what came back, which document matters now. That assembly is done by software: it is the memory component of the harness, working faster than any human could curate.

This is where context engineering stops being invisible and starts being the experience itself. Assembled well, the customer feels a service that remembers them and picks up where the journey left off. Assembled badly, they repeat their story at every step — or the agent acts confidently on information that is no longer true.

The same agent given given A starved window Contains: the instructions · the latest message — nothing else Missing: the journey so far, the customer’s file, what step three found → “Could you give me your order number again?” An engineered window Adds: the journey so far · results of earlier steps · the one document that matters Selected for this moment — not everything, the right thing → picks up exactly where the customer left off
Fig. 2 — Same agent, two windows, two very different customers’ days.

Part 4

What to remember

  1. 01

    The model only knows what is in its context window — each time, every time. Nothing outside it exists.

  2. 02

    Context work is editorial: selection, order, freshness and exclusion — not accumulation. More is not better.

  3. 03

    In agents, the harness rebuilds the window at every step. Memory is something someone designed, not something the AI has.

  4. 04

    When a service feels like it knows you, that feeling was engineered — someone decided what it should see at that moment.

For business & service designers

What the service knows
is an experience decision

We already make this decision for people. What should a front-line employee know about the customer standing in front of them? What is on their screen at that moment, what did the previous department pass along, what are they not allowed to see? Every service blueprint answers those questions somewhere. The context window is the same decision, made explicit and executable — and made thousands of times a day.

Which means the window is where our knowledge of the customer either reaches the agent, or doesn’t. Look at what goes into one, and the craft is familiar.

Retrieved documents → content & knowledge strategy
The agent answers with whatever it is given to read. Curating that knowledge — current, consistent, written to be understood — is content strategy with new stakes: a stale article becomes a confident wrong answer.
Customer history → journey continuity
What the window carries between steps and channels is what makes a journey feel continuous. Nothing says “we don’t know you” like being asked for your story twice.
Order & emphasis → what this moment needs
Deciding whether the policy or the customer’s distress should weigh more in the window is moment-of-truth prioritisation — a judgement about the situation, not a technical setting.
Exclusions → privacy & trust
What the agent must never see is a promise to the customer. Deciding which data stays out of the window — and being able to say so — is trust design, not compliance paperwork.

What will we do in the future?

The same answer as everywhere on this site: much of what we already do, applied to a new material. Four places where context work belongs on a designer’s desk:

  1. Map the window per journey step

    For each moment an agent touches the journey, ask: what should it see here? What did the customer already tell us? Journey maps become window specifications.

  2. Audit where customers repeat themselves

    Every “as I said before…” in your research is a context gap — a place where information existed but never reached the window. We already know how to find these; now they are fixable.

  3. Curate what the agent drinks from

    Help articles, policies, product data — the agent will repeat them verbatim to customers. Owning the quality and freshness of those sources is now front-line experience work.

  4. Own the exclusion list

    Sit with privacy and legal to decide what never enters the window — and treat it as a customer promise worth communicating, not fine print.

The question to take with you

An agent can only care about what it can see.
Who decides what it sees?

Next topic

What is an AI harness?

The behaviour of an AI agent is decided mostly by the structure built around the model — its instructions, tools, memory and guardrails.

Read the explainer →