Introducing Grounding into the Agentic Design Loop: An Exploratory Review of an Unclaimed Problem
This exploratory review identifies and consolidates the unclaimed problem of "Design Grounding"—the critical lack of mechanisms to verify that LLM-generated designs correspond to real-world constraints before fabrication—by screening thousands of papers to reveal that only a handful address this structural challenge, which remains fragmented across adjacent fields.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where computers do not just follow instructions but begin to draw up the blueprints for new things. They are moving beyond simple tasks, like sorting emails or playing games, and are now designing complex systems: the architecture of computer chips, the workflows for scientific experiments, and the structures of software. This shift is called "agentic design." In this new role, the computer acts as an architect, deciding not just how to build something, but what to build in the first place. However, there is a dangerous gap in this process. When a human architect designs a house, they visit the site, measure the ground, and check the local weather. They ensure their plan fits the real world before a single brick is laid. When a computer designs a system, it often works in a vacuum. It creates a plan that looks perfect on the screen, full of internal logic, but it has never touched the reality the system must survive. If the computer's plan is flawed, the error might not be discovered until the system is built, at which point fixing it could be incredibly expensive or even impossible.
A team of researchers at Tongji University set out to investigate this specific blind spot. They wanted to know if these computer agents were checking their designs against the real world, or if they were just building castles in the air. They coined a term for the missing link: "design grounding." This is the process of connecting a computer's internal plan to the physical or operational reality it must satisfy. The researchers asked a simple but profound question: when an agent designs something, does it have a way to know if that thing will actually work in the real world before it is built? To find the answer, they did not just look at one field; they scanned thousands of research papers across computer science, engineering, and robotics to see how often this connection actually happens.
The researchers built a massive search to find papers where an agent designs a system and then tests it against the real world. They looked at nearly three thousand research papers published between 2020 and mid-2026. They sorted these papers into four groups: those where agents only designed without testing, those where agents only tested without designing, those where neither happened, and the rare group where agents both designed and tested their work against the real world. The results were stark. The group where agents successfully closed the loop—where they designed something and then checked it against reality—was the smallest of all, representing only about one in fifteen of the papers they reviewed. Even more telling, when they looked closely at the twenty papers that did manage to close this loop, they found that nineteen of them were still testing their designs against simulations. They were checking their blueprints against a computer model of the world, not the world itself. Only one paper in the entire group actually used real-world measurements to verify the design.
The study suggests that the difficulty is not just a lack of effort, but a structural problem. The object an agent is designing does not exist yet. You cannot measure a bridge before it is built, and you cannot test a new computer chip before it is manufactured. Because the real object is missing, the agent is forced to rely on predictions, simulations, or models. These tools are useful, but they are not the same as reality. The researchers found that when agents do try to test their designs, they often use tools that they or their creators built specifically for the test. This creates a trap where the design is checked against a world that was made to be easy to pass, rather than a world that is independent and unpredictable. It is like a student taking a practice exam that they wrote themselves; they might get a perfect score, but that does not mean they are ready for the real test.
The consequences of this gap are severe. In engineering, decisions made early in the design phase determine the vast majority of a project's final cost. If a flaw is found after the system is built, fixing it requires undoing work that has already been done, which can cost ten times more than fixing it during the design stage. In some cases, like building a physical chip or a medical device, a wrong design cannot be fixed at all; it must be discarded, wasting time and resources. The researchers found that the field of "agentic design" is growing rapidly, with the number of papers doubling every year, but the ability to ground these designs in reality has not kept pace. The tools that exist in other fields, such as robotics or software testing, are not being combined into a single process that allows an agent to design, test, and revise in a continuous cycle.
The paper concludes that the solution is not just better software, but a new way of assembling the tools we already have. We need to build a system where the agent can observe the world, validate its ideas against independent facts, revise its plans based on real feedback, and measure the results, all within a single loop. Currently, these pieces exist in separate fields: robotics has ways to move from simulation to reality, and hardware engineers have tools to check designs against physical laws, but no one has put them together for the new generation of design agents. The researchers argue that until this connection is made, the promise of agents designing the future will remain limited by the risk of building things that look good on paper but fail in the real world. The path forward requires building a bridge between the digital mind of the agent and the physical constraints of the world it is trying to shape.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.