The Rise of AI Agents in Design, Engineering & Architecture

Across four years of CDFAM Computational Design Symposium talks, AI agents have gone from a research curiosity to something structural engineers run in Slack. The record of how fast that happened is sitting in the transcripts.

Washington DC recordings are still incomplete. Four of the first six talks released mention agents.

The word arrived before the capability. 

In 2023, when Chris McComb of Carnegie Mellon described agents to a CDFAM audience in New York, he was reporting an experiment rather than a product. His group had run human design teams against hybrid ones, and the hybrid teams were missing half their people.

We had some human teams, we had some hybrid teams where we ripped out half of the people and replaced them with AI agents.

Chris McComb, Carnegie Mellon University, CDFAM NYC 2023

The finding was that the hybrid teams held up, and that they held up because communication went up rather than down. 

People took on what McComb called the role of an AI handler, translating between the model and the task so the agents could stay useful members of the team. Two years later that description reads less like a laboratory result and more like a job.

The Turn

What changed between 2023 and now was not the idea but the plumbing. By the Amsterdam symposium in 2025, Andrew Sartorelli of Synera was describing an agentic layer sitting on top of automated workflows that already existed, wired into the systems a manufacturer actually keeps its knowledge in: product lifecycle management, simulation data management, requirements tracking. 

The agents were not being asked to invent engineering. They were being given access to it.

That is the pattern that repeats across every discipline in the year since. The interesting work is rarely the model. It is the harness.

Aerospace

Ryan McClelland's generatively designed spacecraft parts at NASA Goddard were already familiar to the CDFAM audience by 2024, but the agent question he raised was about assembly rather than geometry. 

Once individual evolved components exist, they can be combined into larger systems either by code or by agents, until an automated electronic box is something you specify rather than something you draw. The unit of design moves up a level.

Automotive

The most concrete automotive account came from SEAT in Barcelona in 2026. A year of automation work on body-in-white development had already cut development time from two hours to five minutes in one case, and from forty-two hours to one in another, before agents entered the picture at all.

What they showed next was a multi-agent system for sheet metal manufacturability. An engineer uploads a CAD component into a chat window and a supervisor orchestrates the agents beneath it: one extracting geometry such as wall thicknesses and blank size, another running forming simulation, others handling cost and carbon once sourcing data is supplied.

There's a supervisor which is calling and orchestrating all the agents underneath.

Tilman Steininger, Synera, CDFAM Barcelona 2026

The output is not a verdict but a location. If a component cannot be made, the system returns images showing where the problem is, and the engineer goes back into CATIA to resolve it. That loop, rather than any individual model, is the product.

Architecture

Architecture and structural engineering produced the year's most concrete demonstrations, possibly because the work is so heavily gated by specialist availability. At CD/DC in Washington, Sergey Pigach of CORE studio showed an agentic system his firm runs on AWS and interacts with through Slack. Ask it, in a sentence, for a concrete column stack and footing for a five-storey residential building in New York and let it make the remaining assumptions, and it calls the same tools the firm's engineers use, renders the column and writes a design summary.

His framing of why it works is the sharpest line anyone offered on the subject.

It's not a domain expertise problem. It's really an unhobbling problem.

Sergey Pigach, CORE studio | Thornton Tomasetti, CDFAM Washington DC 2026

Frontier models, on his evidence, are already competent at engineering work. What they lack is a way to reach the tools. Pigach also noted something easy to overlook: the agent lives in Slack deliberately, because putting it in a shared space where people already work changes the relationship compared with sitting down in front of a chatbot.

Matthew Goldsberry and Junling Zhuang of HDR gave the practitioner's counterweight at NYC 2025, opening by locating themselves precisely: from the trenches of trying to use these tools in an architecture firm. Their term for it, form jamming, describes assembling AI coding tools and image generation into something usable per project, and their conclusion was that the framework matters more than any finished pipeline.

At Barcelona, Augmenta described full automation of mechanical, electrical and plumbing design as fully agentic, compressing MEP work from weeks to days, while noting plainly that there is a long way still to go.

Industrial design

The consumer-facing end of the field has been slower to use the word and quicker to build the constraints agents will eventually need. Ruben Verhack of Datameister put the problem precisely at Amsterdam in 2025. Industrial design is governed by hard points: fixed locations, dimensions and rules that cannot move without major consequence, whether that is where a car's engine and occupants sit or where the batteries and cooling zones fall in a consumer device.

You don't want to be creative on these things.

Ruben Verhack, Datameister, CDFAM Amsterdam 2025

General-purpose visual AI generates from the surface inward, he argued, while design works from the inside out, and the last twenty percent is always where it breaks. His team's answer was to constrain generation to geometry already known to be manufacturable, so an exterior can be explored freely around components that are fixed. Laurence Cook of Generative Engineering made a parallel case at Barcelona for splitting a simulation pipeline apart, so stress, impact displacement and manufacturing cost can be explored alongside a designer working in the language of form.

What has to exist underneath

The most structural claim came from Intact Solutions at CD/DC. Today's tools, they argued, give you a co-pilot sitting inside an application. What replaces it is a set of agents standing in for domain and subject-matter experts, talking to each other, over what has to be a robust callable physics infrastructure underneath. Without that layer there is nothing for an agent to call, and the demonstration collapses into text generation about engineering rather than engineering.

That is also why the sceptical voices this year have been so precise. SimScale's Andrew Acuff opened his CD/DC talk by renaming it: the programme said agentic workflows were reshaping simulation-driven product development, and he told the room he preferred accelerating, because the iterative loop between design and simulation is worth keeping. What he wanted gone was the waiting.

Read across four years, that is the actual shape of the change. Not the arrival of autonomous design, which nobody credible is claiming, but the steady disappearance of the queue.

Hear it from the people building it

Every talk quoted here is in the CDFAM presentation index, along with transcripts of every other recorded presentation from the symposium series. Search by speaker, company or topic, or search what was said to find the moment itself. Every result links to the exact second of the recording.

Explore the CDFAM presentation index