The future of architecture is context engineering
Why architecture is the discipline that makes AI work
Across government and regulated sectors, the organisations getting value from AI share one habit. They build context on purpose. For the CIOs and architects who own this work, that changes the roles, the ways of working and the competencies worth building now.
Somewhere in your organisation, a team in finance built an agent over a weekend. A researcher wired up something clever for triage, or marking, or drafting, and told no one. A vendor switched on a feature you never asked for, and it is now rewriting your documents. Three more are running that you have not heard about yet.
This is the texture of AI in most large organisations right now. Capability spread unevenly. Real ingenuity in pockets. A lot of energy, very little of it pointing the same way. We call it random acts of AI. Once you have the phrase, you see it everywhere.
Context is the work, now named
In the last wave of transformation, the thing you had to fix was data. Platforms could not run without clean records and reliable pipelines. The equivalent now is context. Not the documents, which most organisations have indexed to death, but the reasoning behind them. Why a decision went the way it did. What the trade-offs were. The hard-won knowledge that sits in a senior practitioner’s head and never reaches a system.
That discipline has a name. Context engineering. Information with judgement. Data plus the why we did it that way. Designing, capturing and maintaining the organisational, sector and role context an agent needs to make a good decision.
A general-purpose agent without that context behaves like a chatbot. Give it the context and the same model respects your legislation, your risk posture and the things that make your organisation specifically yours.
For anyone who has spent a career in architecture, this is familiar ground. Capability models. Service models. Reference architectures. The line of sight from strategy to portfolio to backlog and back. Context engineering is that discipline carried into the agent era.
Architecture has always produced the diagram. The value shows up when the diagram changes how people work. Context engineering is what carries it across that gap. And now it is something you can put a price on.
“+1 for ‘context engineering’ over ‘prompt engineering’. People associate prompts with short task descriptions you’d give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window.”
What seven leaders kept coming back to
We tested the pattern in a room we convened with seven senior leaders responsible for AI, digital and transformation across government. Two hours, no vendors pitching, a frank conversation about what was working and what was not. The agreement was close to total. Everyone had the pilots. Everyone had the shadow use, the enthusiasm and the published frameworks. What almost no one had was a way to make the whole thing cohere.
The conversation kept circling back to coherence, and to a quieter problem underneath it. People are busy keeping the front of the organisation running, and almost none of that effort is happening strategically.
What changes for the architecture function
Three things change for the people who run this.
The roles change. Two are already appearing, in our own practice and in the organisations moving fastest. The Context Architect owns how memory and context are designed and curated, so an agent inherits the right knowledge rather than guessing at it. The forward-deployed architect, a pattern borrowed from Palantir and OpenAI, sits where the work happens and builds a rough working prototype in place, so a sponsor can see what production would feel like before committing. Pair the architect who can build with the portfolio lead who can hold the conversation at the right altitude, and use cases get found and proven quickly.
The work splits into three life-cycles, run in parallel. The software life-cycle of the agents and platforms. The context life-cycle of the prompts, the knowledge, the examples and the patterns. The structural life-cycle of the roles, the governance and the processes around them. Most organisations run one of the three well. Running all three together is the capability worth building, and it is hard to copy.
The ways of working change. Models are deprecated on a cadence measured in months, so anything you build has to be tested and validated again whenever a model changes. At that pace, hand-crafting a prompt for every agent stops being viable. What scales is a method. A workbench that assembles prompts from modular, tested components. A library of context you can version and reuse. Tests that run on their own and tell you when a model update has quietly changed an agent’s behaviour. This is engineering discipline applied to context, and it is what lets a system survive the move from a demo that dazzles in the room to something that holds up in production.
“The issue is no longer whether people will use AI. They already are. The question is whether we can bring that energy into a coherent system before the organisation fragments around it.”
Own the context, or rent it
There is a strategic reason to own this rather than rent it. If a vendor owns your context, the vendor owns your operational knowledge. The model is rented and replaceable. The context is the part that is genuinely yours, worth keeping inside the organisation, in your own infrastructure, portable across whichever models you point at it next. For government, that is a question of sovereignty before it is a question of procurement.
Done properly, context compounds. The first engagement carries the cost of setting it up. The second inherits the memory. The third runs at close to peak. The organisations that start now will hold a position late movers cannot easily buy back.
The competencies worth building
If you are watching this shift and wondering what to build in yourself, a few sensibilities matter more than any tool. Curiosity about the why behind decisions, because that is the raw material of context. Comfort working in the open, because context engineering only works when knowledge is shared rather than hoarded, and anyone who keeps context private to defend a seat becomes the bottleneck. Judgement about what good looks like, because someone has to tell quality from slop, and there is a great deal of slop. A willingness to redesign the work itself rather than bolt an agent onto the process you already run. The tools will keep changing. These hold.
“You can have all the best agents and platforms, but if you can’t keep the context system tight and fit-for-purpose, the whole thing unravels.
This is the work we do as architects, now wearing its proper name.”
Where the value goes
The human role moves up. As context gets encoded, people shift towards curation, judgement, quality assurance, and deciding which context to serve, to whom and when. The judgement a senior practitioner carries stops being locked in one head and becomes reusable. That is the cognitive load paradox in practice. You take on the effort of encoding the context once, and you free the judgement for good.
The leaders in the room kept returning to a related point. As organisations get this right, they move from institutional knowledge, the kind that walks out the door when someone resigns, to institutional awareness, a live and managed system that knows how the organisation actually runs.
The harder questions that follow are human ones. Where does the productivity dividend go. How do you give the next generation a way to learn judgement when the apprenticeship model many of us grew up with no longer fits. Those questions held the table longer than any tool did.
From here on, the work changes
We are building our own practice around exactly this, in the open, including the parts that are not finished yet. The architecture function is being re-architected. From here on, the work changes.