Imagine a world where complex metal parts are not carved from solid blocks or cast in molds, but are instead built layer by layer, like a three-dimensional printer stacking molten metal. This is additive manufacturing, a technology that promises to create shapes previously impossible to make. However, building these parts is a delicate balancing act. If a wall is too thin, it might collapse; if a section hangs too far out without support, it could warp under heat. For decades, engineers have had to manually check their designs against a long list of rules, run simulations to predict how the metal will behave, and then redraw the part if something looks wrong. It is a slow, repetitive process where human judgment bridges the gap between a computer's raw numbers and the physical reality of the factory floor.

Now, a new system called AgenticDfAM is changing how this work gets done. Instead of just checking a design against a static list of rules, this system uses a team of intelligent software agents that can reason, measure, and even rewrite the rules they are following. These agents do not just look at a picture of a part; they interact with specialized tools that measure the geometry with extreme precision. If a design fails a check, the system doesn't just flag an error; it proposes a fix, runs a simulation to see how that fix affects the part's strength and temperature, and then decides if the change is worth making. Crucially, the system is designed to be transparent. Every decision it makes is backed by a specific measurement or a calculated number that a human engineer can inspect and verify.

The researchers behind this work, led by David Guirguis, wanted to see if they could automate this entire cycle of design, assessment, and repair without losing the rigorous standards required for engineering. They built a workflow where a large language model acts as a manager, coordinating different software tools to handle specific tasks. One tool recognizes the shapes in a digital model, another measures distances and angles, and a third runs complex physics simulations to predict how the metal will heat up and cool down. The language model agent reads the results from these tools, interprets what they mean, and decides what to do next. If a part has a section that is too thin, the agent might suggest thickening the wall. If that makes the part too heavy, it might suggest a different shape. The system keeps a running log of every measurement and every rule it applied, ensuring that the final design is not just a guess, but a solution grounded in hard data.

To test this system, the researchers started with a digital model of a complex metal bracket. They fed it into the workflow along with a set of manufacturing guidelines for a specific type of nickel alloy. The system immediately began its work, breaking the design down into its geometric features and measuring them against the rules. It found several violations, such as walls that were too thin and overhangs that were too steep. The agent then generated three different ways to fix these problems. One strategy kept the overall shape of the part exactly the same but added material to strengthen weak spots. Another strategy removed some material to make the part lighter. A third approach tried to balance both. The system ran simulations on each of these new versions to see how they would behave during the actual printing process.

The results showed that simply fixing the geometric errors was not enough. The version that added the most material to fix the weak spots ended up having the least amount of warping when the part was removed from the printer. However, the version that removed material to save weight actually warped more in certain directions, even though it satisfied all the basic geometric rules. This finding highlights a key limitation of traditional design tools: they often look at one problem at a time. AgenticDfAM, by contrast, could see the trade-offs. It showed that making a part stronger in one way could make it weaker in another, and it allowed the engineer to see these consequences before the metal was ever melted.

The system also proved capable of handling problems that were not in its original list of rules. In one experiment, the researchers asked the system to design support structures for a clamp. These supports are temporary scaffolds that hold up the part while it is being printed. The system generated a design, but when a human reviewer looked at it, they noticed something odd: the supports were sitting on top of the part's surface where they shouldn't be, rather than underneath it. Instead of just flagging this as a human error, the agent investigated. It asked the software to measure the distance between the support and the part surface and realized the code used to generate the supports was making a mistake with curved surfaces. The agent then rewrote the code to fix the error, verified the new code on a flat surface and a curved surface, and ran the entire process again. This time, the supports were placed correctly, and the amount of wasted support material dropped significantly.

Perhaps the most striking demonstration of the system's adaptability came when it was asked to redesign a part to handle heat. The simulation showed that heat was building up in a specific area of the bracket, which could weaken the metal. The agent proposed a series of changes to the shape of the part to improve how heat flowed through it. It suggested rounding off sharp corners and filling in narrow gaps. After running the simulations again, the system found that one of these changes reduced the peak temperature by 160 degrees Celsius. This was a massive improvement, but it came with a cost: the part became slightly heavier and shifted a tiny bit more when the supports were removed. The system presented this trade-off clearly, showing that the engineer could choose between a cooler part or a lighter one, with full knowledge of the consequences.

Throughout these tests, the system maintained a strict separation between the reasoning and the measurement. The language model agent could suggest a change, but it could not invent a number. If it needed to know the thickness of a wall, it had to ask a tool to measure it. If it needed to know the temperature, it had to ask a simulation to calculate it. This ensured that the system could not hallucinate facts or make up data to fit a story. When the system made a mistake, such as the initial error with the support structures, it was able to diagnose the problem, correct the underlying code, and verify the fix before moving on. This ability to self-correct and adapt the very procedures it uses to build a design is what sets AgenticDfAM apart from previous attempts at automation.

The researchers found that the system could successfully navigate complex design challenges that usually require a team of engineers and hours of manual work. It could identify violations, propose multiple solutions, simulate their outcomes, and even fix the code it uses to generate those solutions. In the case of the clamp, the system reduced the number of times it had to go back and fix the support design from ten attempts down to five. It also managed to reduce the amount of wasted support material by nearly 30 percent in one orientation. These improvements were not just about speed; they were about finding better solutions that a human might have missed because they were too focused on a single rule or a single constraint.

The work demonstrates that artificial intelligence can be a powerful partner in engineering, not by replacing the engineer, but by handling the heavy lifting of measurement and calculation. The system acts as a tireless assistant that never gets tired of checking numbers or running simulations. It allows the human engineer to focus on the bigger picture: deciding which trade-offs are acceptable and what the final part needs to do. By keeping the measurements and the reasoning separate, the system ensures that every decision can be traced back to a specific piece of evidence. This transparency is crucial for industries like aerospace and medicine, where a mistake can have serious consequences.

In the end, AgenticDfAM shows a path forward for how we design the things we build. It moves beyond the idea of a computer simply checking a list of rules and toward a system that can understand the context of a design, predict its behavior, and adapt its own methods to solve problems. The system does not claim to have all the answers, but it provides a way to find them faster and with greater confidence. It turns the design process into a conversation between human intent and machine precision, where every suggestion is backed by a number, and every number is open to inspection. This is not just a new tool for engineers; it is a new way of thinking about how we create the complex objects that make up our modern world.