Agents for Research & Analysis

Design research has a shape that fits an agent almost perfectly. It is a goal you can state clearly ('find and compare precedents for a 200-bed district hospital on a constrained urban site'), it decomposes into steps (search, gather, read, extract, compare, summarise), it needs tools (web search, document reading, sometimes a database or a PDF full of standards), and it produces an artefact you can inspect (a briefing, a comparison table, a set of citations). None of it requires the irreducible human core of design - the judgement about what is good, the creative leap, the duty of care - so almost all of it can be delegated and checked rather than performed by hand. That is why research is where most designers should take their first serious agentic step.

Consider what a research agent can realistically take on across a project. Precedents: finding comparable buildings, extracting what worked, organising them by relevant criteria rather than just by style. Codes and regulations: locating the clauses that apply to this building type in this jurisdiction - though this is the highest-verification category, treated below. Products and materials: comparing systems, assemblies and finishes on performance, cost, availability and sustainability, pulling data from many manufacturer sources into one comparable table. Market and context: typologies, user needs, cost benchmarks, local practice. Synthesis: turning a heap of gathered material into a structured, readable briefing that surfaces patterns and tensions you would want to know about.

The practical gain is not only speed but reach. A single designer, or a small Indian studio without a research department, can now cover ground that used to need a team or a subscription library - reading more widely, comparing more options, catching a precedent or a product they would never have had time to find. The grind of first-pass research, the part most people do grudgingly and incompletely under deadline, is exactly the part an agent does tirelessly. Handled with the verification discipline this lesson insists on, that is a genuine expansion of what a small practice can do well - more thorough groundwork, more considered decisions, more time left for the design itself. The catch, always, is that reach without verification is just a faster way to be confidently wrong.

Here is the hard truth that every user of a research agent must internalise: the agent's output is a draft to verify, never a fact to trust, and the polish of the output tells you nothing about its truth. Language models generate plausible text; a plausible-sounding citation, statistic or code clause is exactly what they are built to produce, and they will produce it whether or not it corresponds to reality. This is not a bug that will soon vanish - it is inherent to how these systems work, and it is worst precisely where it matters most: specific numbers, named sources, exact clauses, precise quotes.

Three failure modes deserve names. Fabrication (hallucination): the agent invents a source, a study, a product or a figure that does not exist, presented with total confidence. Always open a cited source and confirm it says what the agent claims - a real-looking URL or a plausible author is not evidence. Misreading: the source is real but the agent has summarised it wrongly, blended two things, dropped a crucial condition, or stated as general a claim that was heavily qualified. Staleness: the fact was once true but is now superseded - a withdrawn standard, a discontinued product, an old price, a changed regulation. Agents are especially prone to citing older material as current, so date-check anything time-sensitive.

The workable discipline is not to distrust everything equally but to triage by stakes, as the figure shows. Low-stakes framing and orientation (which precedents to look at, general context, ideas to explore) can be spot-checked. Claims and data you will use (quoted figures, product specifications, cost benchmarks) require opening the primary source and confirming it. Safety, code and binding facts (a fire clause, a structural limit, a byelaw setback, anything that becomes a commitment) must be verified against the actual standard or authority - here an agent citation is a pointer to check, never a substitute for reading the source, and this is the professional's own responsibility. A useful habit: require the agent to cite sources for every substantive claim and to flag its own uncertainty, then treat those citations as a to-verify list, not as proof. The agent has done the finding; confirming what is true remains your job, and it is where your professional judgement earns its keep.

The quality of agent research depends heavily on how you direct it, and a little craft here pays off enormously. A vague goal ('research sustainable facades') yields a vague, generic briefing; a sharp goal yields something you can actually use. Good direction has three parts: a precise goal, guidance on sources, and a format that makes verification easy.

Frame a precise goal. State the building type, location, constraints and the decision the research serves. 'Compare double-skin, terracotta rainscreen and high-performance glazing facades for a naturally-ventilated office in Bengaluru's climate, on cost, thermal performance, maintenance and local availability, to inform a cladding decision' will produce something ten times more useful than 'facade options'. Tell the agent what you already know so it does not waste effort, and tell it what a good answer looks like. Guide the sources. Where you can, point the agent at authoritative material - a standards library, specific manufacturer data, reputable references - and tell it to prefer primary and recent sources and to say when it is relying on weak or dated ones. Grounding an agent in trustworthy documents you provide (retrieval over your own reference set) dramatically reduces fabrication compared with letting it roam the open web unguided. Demand a verifiable format. Ask for claims with inline citations, a clear separation of fact from inference, an explicit list of assumptions, and a flag on anything the agent is unsure about or could not confirm. A comparison table with a source column is far easier to check than a flowing essay.

Then iterate. Read the first pass critically, push back where it is thin or suspicious ('this cost figure seems low - source it or correct it'), ask it to dig deeper on the parts that matter and to drop the padding. Treat the exchange as directing a capable but over-confident junior researcher: clear brief, sceptical review, specific corrections. And keep the confidential-data discipline in mind from the start (Module 8.3) - be deliberate about what project or client information you put into a tool, because research often tempts you to paste in a brief or a site's details. Directed this way, a research agent becomes a genuine force multiplier; directed carelessly, it becomes a fast generator of plausible material you then have to unpick.

Agents do more than gather; they can also analyse - and the line between the two is where your judgement must stay sharpest. Analysis means finding patterns, drawing comparisons, ranking options against criteria, and surfacing tensions and trade-offs. An agent can take fifty precedents and cluster them by strategy; take a dozen products and rank them against your weighted criteria; take a pile of user research and pull out recurring needs. This is real analytical value, and it can genuinely sharpen your thinking by showing you structure you would have missed.

But analysis carries a subtler risk than gathering, because its conclusions feel like insight rather than mere data, and are therefore more tempting to accept whole. Two cautions matter. First, an agent's analysis embeds the agent's framing and any bias in its sources - if it ranked facades and quietly over-weighted cost, or drew its precedents from a narrow, unrepresentative set, the tidy conclusion inherits that skew invisibly. Interrogate the criteria and the inputs, not just the output: ask what it weighted, what it left out, what a contrary view would say. Second, the leap from analysis to a design decision is yours and is not analytical - the agent can tell you that option B scores highest on your stated criteria, but whether option B is right for this place, these people and this project is a judgement that includes things no criteria captured, and that judgement is the designer's, not the tool's. The agent proposes the analysis; you dispose on the decision.

So the healthy pattern is a partnership with a clear division. Let the agent do the tireless work of gathering wide and structuring the material - the part that used to eat your days. Bring your own eyes to whether the sources are sound and the framing is fair, verifying everything that will become a commitment. And keep firmly for yourself the interpretation that turns sound research into a good decision, because that is design, not research. Used this way, agents do not make you a lazier researcher - they let you be a more thorough one, covering more ground and then spending your freed attention on the judgement that research exists to serve. That is the shift this whole module is about: agents on the surrounding work, the designer on the decision, verification joining the two.