Agents for Ideation & Concept

Divergent thinking - deliberately generating many possibilities before judging any of them - is a distinct mode from the convergent thinking that selects and refines, and good designers move between the two consciously. The classic double-diamond in the figure names this rhythm: widen, then narrow, then widen again on the chosen direction, then narrow to the resolved concept. Agents are unusually good partners for the widening passes, because generating many varied possibilities on demand is close to what a generative model naturally does, and because a machine has neither ego nor fatigue nor fear of a bad idea - it will offer the obvious and the outlandish with equal willingness.

Concretely, an agent can help you diverge in several ways. It can generate organising concepts - many different ideas for what the project could fundamentally be about, from the pragmatic to the poetic. It can offer angles and lenses - 'approach this brief as a landscape', 'as a single room', 'as a sequence of thresholds' - that jolt you out of your first framing. It can surface references and analogies you had not thought of, drawing on a vast store of precedent. It can stress-test a direction by arguing the opposite, or by listing what could go wrong with a concept you like. And it can rapidly variate - take one idea and spin it into ten flavours - so you can feel the range of a direction before committing. For a solo practitioner, this approximates the productive argument of a studio; for a team, it is extra fuel for the charrette.

The value is real but specific: the agent is good at quantity, breadth and speed of possibility, and at breaking your fixation on the first idea. That directly addresses the two great enemies of good early design - the blank page and premature fixation. A designer who uses an agent to force themselves to consider more before deciding is, other things equal, likely to make a more considered choice than one who grabbed the first idea that came. The point of this widening is not that the agent will produce your concept - it will not - but that it enriches the field from which your judgement chooses. Diverge generously with the machine; converge deliberately with your own eye. That rhythm, held consciously, is how agentic ideation makes you a broader thinker rather than a lazier one.

The single most important habit in agentic ideation is to keep the agent on the divergence side of the diamond and yourself on the convergence side. The agent's job is to expand the field of what you are choosing among; your job is to choose, and to develop the choice into a concept with intent behind it. Blur that line - let the agent's options collapse into your decision - and you get design by default rather than design by judgement.

This matters because an agent's suggestions carry a quiet gravitational pull. They arrive fluent, organised and confident, and human beings anchor hard on the first plausible option they are shown, especially under deadline. If you ask an agent for a concept and take what it offers, you have not ideated - you have accepted, and you have accepted something with no author behind it, no intent, no relationship to this place and these people beyond the statistical. A useful counter-habit, shown in the figure, is to generate your own first: form your own instinct about the project before you ask the agent, so its options land as additions to your thinking rather than as replacements for it. Then treat what it offers as raw material to react to - to combine, reject, twist, or use as a springboard - rather than as a menu to pick from.

Directing for width rather than for an answer also changes how you prompt. Instead of 'what should the concept be', ask for range: 'give me ten genuinely different organising concepts, including at least two that are unconventional or uncomfortable, and for each say what it would prioritise and what it would sacrifice.' Push explicitly against the averaged middle - agents, left alone, tend toward the safe and generic because that is the statistical centre of their training, so you have to ask for the edges, the extremes, the deliberately strange. Ask for the trade-offs, not just the ideas, so you are choosing with your eyes open. And keep interrogating: 'what would the opposite of this be', 'what is this concept ignoring', 'push this further'. Used this way the agent becomes a widening instrument you control; used as an oracle you consult for the answer, it quietly narrows your thinking to its own comfortable centre while feeling like it expanded it.

Selecting and developing the concept - deciding which idea is right for this project and then giving it depth, intent and coherence - is the creative heart of design, and it is not delegable. This is not a sentimental claim about the human spirit; it is a practical one about what a concept has to do and what an agent cannot do. A concept has to be right for a specific place, a specific client, a specific set of constraints and hopes, and it has to be held with conviction through the hundreds of downstream decisions that make a project cohere. That rightness is a judgement saturated with things no agent holds: taste, an understanding of these people, a sense of place, professional responsibility, and the accountability to stand behind the choice.

So the agent proposes and the designer disposes, and the disposing is the design. When you converge, you are bringing your own eye to the field the agent widened - and your eye is doing something the agent cannot: judging what is good, not merely what is plausible or popular. An agent can tell you that a concept is common, or coherent, or fashionable; it cannot tell you that it is right, because right is a value judgement about this project made by someone who will answer for it. Guard this actively. Beware the concept that is fluent but has no author behind it - the plausible, generic idea that survives because nobody pushed back. Beware, too, the erosion of your own generative muscle: if you always reach for the agent before forming your own instinct, that instinct weakens, and the instinct is the thing you are actually being paid for and the thing that makes your work yours rather than the model's.

There is also an authorship and integrity dimension you cannot outsource. The concept carries your name and, downstream, your professional responsibility; a concept you did not really author is one you cannot fully defend or develop with conviction. And agent-generated ideas draw on a vast pool of existing work, which raises real questions of originality and influence you must hold consciously (Module 8.4) - using an agent to widen your thinking is healthy, but the leap to a concept, and the ownership of it, must be genuinely yours. The healthiest posture is confident partnership: let the agent be the most tireless, uninhibited brainstorming partner you have ever had, and remain, without apology, the designer who decides what is good and makes it so.

A further discipline separates agentic ideation from agentic production: at the concept stage the agent's fluency is at its most seductive and its least checkable, so you must be clear about what a concept from an agent is and is not. A concept offered by an agent is a provocation and a possibility, not a validated design. It has not been tested against structure, cost, code, buildability, climate or the messy realities of the site - and an agent will describe an unbuildable, unaffordable or unsafe concept with exactly the same confidence it brings to a sound one, because at this stage it is generating language about possibilities, not engineering a solution.

This has two practical consequences. First, treat exciting agent concepts as hypotheses to test, not answers to adopt. The figure's third row names the trap: a plausible-but-unbuildable concept is worse than a mediocre buildable one if you fall for it, because you will invest in a direction that cannot survive contact with reality. When a concept genuinely excites you, that is the moment to interrogate it - can it be built, can it be afforded, does it meet the brief and the code, does it actually serve the client - rather than the moment to fall in love. The feasibility work is downstream and it is real; the concept has to earn its way through it. Second, the concept stage is not where verification is most acute, but it is where the seeds of later problems are sown - a concept that ignores an inescapable constraint will fail expensively later, so a quick reality check even at this stage is worth the discipline.

Balanced against this caution is the genuine upside, which is worth holding onto. Used well, agents let you explore more of the possibility space before committing - to test more directions cheaply, to catch a better idea you would not have had time to reach, to break a fixation that would have led to a worse building. The goal is not to ideate less carefully but to ideate more widely and then judge more rigorously: widen with the machine, converge with your eye, and pressure-test the survivors against reality before you commit. That is the rhythm that turns a fluent option-generator into a real creative asset - and it keeps the concept, and the responsibility for it, exactly where it belongs, with you.