Agents for Analysis & Simulation

Analysis work has three phases, and an agent can help - or fully drive - each. Setting up is the phase that used to demand a specialist: preparing the model for analysis (simplifying geometry, assigning material and thermal properties, defining zones), choosing an appropriate method and tool, and setting the boundary conditions and assumptions (climate data, occupancy and schedules, loads, restraints, cost rates). An agent can do a great deal of this - translating a design model into an analysis-ready one, picking sensible defaults, and, crucially, stating the assumptions it has made so you can check them. Running is the phase agents make almost free: meshing, solving, iterating, sweeping parameters, running the study across many variants. What was an overnight job becomes a coffee-break one, which is what lets analysis rejoin the design loop.

Interpreting is where agents add a newer kind of value: turning raw output - contour plots, tables, load diagrams, energy breakdowns - into plain-language insight a designer can act on. 'This wing overheats in the afternoon because of the west glazing; reducing it or adding shading would cut the peak.' 'The structure is fine except this transfer beam, which is governing.' 'Most of the cost is in these three line items.' This translation from numbers to design implications is what makes analysis usable by a generalist designer rather than only a specialist, and it is genuinely powerful.

Across the common studies - daylight and glare, energy and thermal comfort, structural adequacy, cost and quantities, and increasingly acoustics, airflow and embodied carbon - the pattern is the same: the agent compresses the time from question to answer, and lowers the expertise needed to ask. That is real and valuable. But notice what the agent has taken over: the setting of assumptions and the interpretation of results - which are exactly the two places where analysis goes wrong. Speed and accessibility are the gift; they are also why the verification burden falls harder on you, not lighter. The faster and easier it is to get an answer, the more disciplined you must be about whether the answer is true.

Every simulation is a model of reality, and a model is defined by its assumptions - the climate file, the occupancy and schedules, the material and thermal properties, the loads and restraints, the unit rates and quantities. Get those right and the answer can be trustworthy; get one wrong and the answer is confidently, precisely false, because the solver does not know your assumption was wrong - it computes faithfully on whatever you gave it. This is the oldest law in simulation, garbage in, garbage out, and agents raise the stakes on it in two ways. First, an agent may choose or default many assumptions for you, quietly, so the error can enter without your noticing. Second, the agent's fluent, confident narration of the results makes a wrong answer feel authoritative rather than suspect.

The failure is insidious because the output looks identical whether the assumptions were right or wrong: the same precise numbers, the same polished charts, the same calm explanation. A daylight study run on the wrong latitude, an energy model with a phantom occupancy schedule, a structural check missing a load case, a cost estimate on stale rates - each produces a clean, plausible result that can steer a real design decision the wrong way. And because the agent lowered the expertise needed to run the study, the person reading the answer may be least equipped to notice that a key assumption is off. That is a dangerous combination.

So the first and most important verification is not of the result but of the inputs: what assumptions did the agent make or use, and are they right for this project? Insist that the agent surfaces its assumptions explicitly - climate source, occupancy, properties, loads, rates, method - and check the ones that matter against reality before you look at a single output number. A result you cannot trace to its assumptions is not a result; it is a guess with a chart. Verifying the frame, exactly as in generative modelling (Module 5.2), is the discipline that separates analysis that improves a design from analysis that confidently misleads it.

The highest value an agent adds in analysis is turning results into insight - and it is also where a subtle trap lies. Done well, this is transformative: an agent can read a mass of output, find the pattern that matters, connect it to a design cause, and propose a design response - 'the overheating is driven by west glazing; here are three moves that would help and their rough trade-offs'. That closes the loop from analysis to design in a way that used to require a specialist sitting beside you, and it lets a designer reason about performance early and often. Used to inform judgement, it is one of the best things agents do in this whole course.

The trap is treating the interpretation as the truth rather than as a reading to check. An agent's narrative is a hypothesis about what the numbers mean, and it can be wrong in ways that are hard to catch: it can over-claim certainty the data does not support, miss a second-order effect, mistake correlation for cause, or smooth a messy, ambiguous result into a tidy story that is more confident than the underlying analysis warrants. The more fluent and helpful the explanation, the easier it is to accept without asking whether it is actually supported by the output - and whether the output itself is sound.

The discipline, then, is to use the agent's interpretation as a fast, valuable first reading, and to keep your own judgement in the loop: does the explanation match the raw output; is the confidence warranted; what is it not saying; would a specialist agree. For anything consequential, look at the underlying results yourself, not just the summary. And hold the distinction between decision support and decision: the analysis, however good, informs a design decision that remains yours - the agent tells you the room overheats, but whether to change the glazing, add shading, or accept it is a design judgement you make and own. Insight from analysis is a powerful input to your thinking; it is not a substitute for it.

Because analysis can be confidently wrong and can drive real decisions, it demands a specific verification discipline, and the higher the consequence, the tighter it gets. Sanity-check every result against expectation and order of magnitude. Before trusting a number, ask whether it is even plausible - does this energy figure sit where similar buildings do, is this deflection remotely reasonable, does this cost land near a benchmark. An answer that is an order of magnitude off is usually an assumptions or units error, and a rough hand-check or a known benchmark catches most gross mistakes. Check against ground truth where you can: rules of thumb, published benchmarks, a simpler independent calculation, or measured data from comparable buildings. Cross-validation is your cheapest defence against a plausible-but-wrong result.

The stakes rise sharply for safety-critical and cost-critical analysis. A structural check that says the beam is fine, a fire or egress analysis, a cost estimate that becomes a client commitment - these are not decision-support you can accept on the agent's word. Structural adequacy must be verified and, where the jurisdiction requires it, signed off by a qualified structural engineer; an agent-run analysis is at most a preliminary check to inform the design, never the professional certification (Module 8.2, and the Electrical and Structural safety principle: specify and verify, never let a tool own a life-safety decision). A cost figure that goes to a client is your commitment; verify the quantities and rates before it does. Treating an agent's structural or cost output as authoritative is exactly the kind of confident-wrong liability this course warns against, with the highest possible stakes.

The honest frame is this: agents make analysis faster, cheaper and more accessible, which is a real advance for design quality, letting you test ideas you could not have afforded to test before. But they do not make you a structural engineer, an energy specialist or a quantity surveyor, and they do not carry the responsibility that those roles hold. Use agent-driven analysis boldly to inform and explore; verify the assumptions and the results; and keep every professional sign-off that touches safety or a binding commitment exactly where it belongs - with a qualified human who answers for it.