Simulation belongs inside the loop, now that answers come faster.

There are two ways simulation shows up in product programs. In the first, geometry gets designed, then sent away for analysis, then a report returns weeks later confirming decisions that have already hardened, or condemning them, at which point nobody wants to hear it. In the second, simulation runs inside the design loop: fast, iterative, answering the question of the week while the geometry is still soft. Same tools. Completely different value. Cheaper compute makes the second way easier to run than ever. It doesn't stop a team from working the first way.

Decide, don't illustrate

The test for any analysis is brutal: which decision does this change? A stress plot that decorates a review deck changes nothing. A tolerance stack that decides between two architectures, a thermal study that kills a fanless fantasy early, a drop simulation that moves a boss before the tool is cut: those earn their cost many times over. We scope simulation backwards from decisions: name the question, pick the cheapest analysis that answers it, run it while the answer can still steer. That test gets more important as analysis gets cheaper. When a parameter sweep costs less than the meeting that reviews it, the temptation is to run everything and decide nothing.

The loop in practice

Inside a design loop, simulation compresses iteration from weeks to days: geometry improves between reviews instead of after failures. It also changes team behaviour: designers propose bolder structures when the physics feedback is fast, because being wrong costs an afternoon, not a milestone. That's the quiet compounding benefit: not better analysis, but wider exploration made safe. Push the idea further and you get simulation-in-the-loop optimisation: geometry evolved directly against the target, with manufacturing constraints enforced.

When to build the rig anyway

Simulation earns trust it must also keep. Models carry assumptions (material data, boundary conditions, contact behaviour), and every consequential result gets its assumptions stated where reviewers can attack them. Some questions still belong to physical rigs: seals, snap-fit feel, anything where surface finish or human perception dominates. The discipline isn't sim versus test; it's knowing which question belongs to which, and never shipping on a model nobody validated. That rule matters more as AI-trained surrogate models start returning near-instant estimates. They predict from the cases they were trained on, and outside that range they can be confidently wrong, so a fast answer still gets checked against a full solve or a rig before it moves a decision. That's how our engineering practice runs analysis: margin where it matters, reports your engineers can audit, and a rig where the model would be guessing.

The agent-native lens: ratings that say how they were earned

When the buyer sends an assistant. Buyers increasingly ask an assistant whether a product will cope with their conditions: a sealed cabinet in summer, a drop onto concrete, a cold start at dawn. It can only answer from what you publish, and a bare rating rarely says which conditions it holds under, or whether it came from a rig or a model. Publish the conditions with the claim: ambient temperature, mounting, duty cycle, the standard tested to where there is one. Say plainly which figures were measured and which were predicted by analysis. An assistant matching your specification to someone's use case can then tell whether the rating applies, instead of guessing, in your favour or against you.

When the business runs on agents. Much of a simulation loop is set-up and bookkeeping, and agents can carry a growing share of it: re-meshing the revised geometry, re-running last week's load cases against this week's CAD, plotting results the same way every time so a change stands out, and flagging when a design change has pushed a study outside the assumptions it was built on. They can keep the assumption register (material data, boundary conditions, contact settings) attached to every result. The one thing they don't do is close a question. An engineer reviews the assumptions behind any result that changes a decision and signs it; until then, an agent's run stays labelled as a draft.

Faster answers, the same test

Simulation is likely to keep getting quicker to run, with more of it set up by software. None of that changes the discipline. Name the decision, run the cheapest analysis that settles it while the geometry is still soft, state the assumptions where reviewers can attack them, and build the rig where the model would be guessing. Analysis that steers keeps its value. Analysis that decorates just gets cheaper to make.