In the modern world of artificial intelligence, machines are increasingly asked to solve problems by reading vast amounts of information. Imagine a librarian who must answer a specific question about a library containing millions of books. One approach is to read every single book cover to cover before attempting an answer. Another approach is to give the librarian a set of simple tools—a way to look up a card catalog, walk to a specific shelf, and read just the relevant pages. This second method relies on the idea that a machine can learn to navigate a complex structure without needing to memorize everything in advance. This is the core challenge of building agents that can reason over knowledge graphs, which are digital maps connecting facts, concepts, and data tables. The goal is to create a system that can find the right information in a massive, changing network of data, much like a human expert would, rather than trying to force the entire network into a single, overwhelming memory slot.

Researchers at Oplit have developed a new system called GRA, a Graph Reasoning Agent, to test this navigation approach against a more traditional method. They built a simulated factory environment, complete with a digital twin of its operations, including rules about how products are made, data about orders, and records of machine performance. This environment, known as the Unified Factory Knowledge Model, contains hundreds of analytical questions that require the system to find specific answers by connecting different pieces of information. Some questions ask for simple numbers, while others require the system to understand complex rules, such as whether a new production schedule is possible given the current state of the factory. The researchers wanted to see if an agent that explores this digital factory step-by-step, using a small set of generic tools to look up information, could perform better than an agent that is simply fed the entire factory manual and database schema all at once.

The team tested three different ways of handling this task. The first method, which they call the full-context approach, dumps the entire description of the factory, including every table and rule, into the artificial intelligence's memory before it starts. This is like handing the librarian the entire library's catalog and every book in the building before asking a single question. The second method removes the map entirely, giving the agent only a list of text documents and tables to search through without any connections between them. The third method, GRA, gives the agent a set of seven simple tools: it can list what is nearby, read the details of a specific item, search for words, search for meanings, run a database query, think through a plan, and finally give an answer. Crucially, the agent does not know the names of the tables or the rules in advance; it must discover them by using these tools to explore the graph.

The results showed that the agent that navigated the graph step-by-step was more accurate than the one that tried to read everything at once. On a test set of 258 questions, the navigating agent answered correctly 88.4 percent of the time, while the agent that read the full context answered correctly 83.3 percent of the time. More importantly, the navigating agent achieved this higher accuracy while reading less than one-third of the information that the full-context agent had to process. This suggests that for complex tasks, it is better to see less but see it more selectively. The agent that read everything was often overwhelmed by the sheer volume of text, whereas the navigating agent focused only on the specific pieces of data needed to solve the problem at hand.

The researchers also discovered that the success of this navigation method depended heavily on the reliability of the tools the agent used. When the artificial intelligence model was good at following instructions and using its tools without making mistakes, the navigation approach worked very well. However, when the model struggled to use the tools correctly, the advantage disappeared, and the method that read everything at once performed better. This indicates that the benefit of navigating a graph comes from the ability to fetch exactly what is needed, but this only works if the agent can be trusted to fetch it correctly. The study found that the structure of the graph itself was not the main reason for the improvement; rather, the improvement came from the act of selectively accessing information. Even when the graph was removed and the agent searched through flat text, it still outperformed the full-context method, proving that the key was the strategy of looking for information rather than reading everything.

To demonstrate how this works in a real-world scenario, the researchers showed how the agent handled a request from a factory operator. The operator asked if a specific rule could be followed: that aluminum frames should be welded at one of two specific stations on a Monday. The agent did not know the answer immediately. It started by looking up the welding stations in the digital map. It found that one station was actually restricted to welding carbon frames only, a rule that was not obvious from the operator's request but was stored in the graph. This immediately ruled out one option. The agent then checked the other station and calculated whether it had enough time to handle the workload. It compared standard time estimates with actual historical data and found that the real work took longer than expected, making the request impossible to fulfill on that day. The agent returned a clear refusal, explaining the two specific reasons why the rule could not be followed, and even suggested two ways to fix the schedule.

In a second example, an operator asked if it was possible to limit the number of color changes on an assembly line to three per shift. The agent again navigated the graph to find the definition of a shift and the records of past production. It replayed ten months of history to see how often this limit would have been broken. It found that on a small number of days, the limit would have been exceeded, but not so often that the rule was impossible to adopt. The agent concluded that the rule was feasible but carried a small seasonal risk. Once the operator confirmed this, the system automatically translated the rule into a mathematical format that a computer solver could use to manage the factory schedule, and then recorded this new rule back into the digital map for future use.

This work demonstrates that artificial intelligence can effectively manage complex industrial knowledge by acting like a skilled investigator rather than a passive reader. By using a few simple tools to explore a structured network of facts, the system can answer difficult questions more accurately and efficiently than by trying to process everything at once. The study shows that seeing less, but seeing it with purpose, allows the agent to answer better. This approach is particularly valuable as data sets grow larger and more complex, eventually becoming too big to fit into a single memory prompt. In these situations, the ability to navigate and retrieve specific information becomes the only viable path to intelligent decision-making. The researchers conclude that the next step is to refine the system so it can not only judge whether a rule is possible but also automatically write the computer code needed to enforce it, closing the loop between human instruction and machine execution.