A few typos by a general-purpose agent in an office scenario are inconsequential, but in an industrial setting, errors mean irreversible losses. A slight deviation in the tightening torque of a single screw or a 5% over-adjustment in pump pressure could halt an entire production line or even scrap the equipment.
Large language models (LLMs) excel at probabilistic generation, whereas industry requires deterministic control. This is the most core contradiction in the deployment of industrial agents.
Based on the long-term exploration of Yang Chengbiao, CTO of KGData, in the practical deployment of industrial agents, this article focuses on the ontology-driven technical paradigm and outlines a technical path to enable agents to achieve deterministic execution in low-tolerance scenarios.
Guest Speaker: Yang Chengbiao, CTO of KGData.
The content has been streamlined. To obtain the expert's complete courseware and video, please scan the QR code at the end of the article.
01 - Why Are Industrial Agents Prone to Deployment Failures? Three Major Dilemmas Lie Ahead
Whether your company is still stuck with the private deployment of large language models, has initially experimented with agents, or is using agent platforms like OpenClaw and Claude Code for deployment, you will encounter three highly significant dilemmas in industrial scenarios.
First, the fracture of global memory.
Within a single session, memory accuracy is relatively high, but once the current session is closed and a new one is initiated, the original memory is completely lost. Taking OpenClaw as an example, although its memory function is relatively comprehensive, precise memory for specific knowledge still requires iterative upgrades or targeted modifications to be realized.
However, the vast majority of AI software driven by large models have relatively weak capabilities in continuous memory.
Second, the collapse of multi-step execution.
General-purpose agents can invoke tools and perform task planning to complete complex tasks. However, during execution, due to the long task chain, agents often cause the final result to significantly deviate from expectations due to a mistake in one step. Therefore, implementing strict control over task planning is critical.
In office scenarios, the importance of this control is relatively low because the error tolerance is high; whether it is document writing, PPT creation, or image generation, a few errors usually do not have a significant impact on the final result. However, industrial production is not like this: a deviation in the installation torque of a single screw could lead to irreversible losses for the entire production line.
Therefore, agents must follow specific paradigms in every step of the reasoning process, including the experience of senior technicians, the specifications in equipment operation manuals, and relevant physical common sense, and enforce them during task planning. This is precisely the capability that is universally lacking in current agent applications.
Third, errors in system-level interactions.
Industrial informatization has evolved over the years, and especially in leading enterprises, the number of informatization systems is massive. If agents are used to schedule these systems, the number of interfaces to be docked is extremely considerable, with the interface scale in some leading factories even reaching the thousands. These interfaces can be used to invoke and deploy various systems, and even directly drive machinery and equipment.
When agents invoke these interfaces, especially in scenarios with frequent interactions among multiple systems, parameter transmission errors or interface invocation errors may ultimately lead to irretrievable losses.
The superposition of the above three types of problems leads to a relatively high failure rate of industrial agents in the deployment process.
02 - Ontology-Driven: Teaching Agents to Follow the Rules
To solve these problems, we have tried various solutions and ultimately confirmed that ontology is a suitable technical choice. Ontology can be viewed both as a technology and as a carrier of data and knowledge to constrain agent behavior, thereby improving its performance in industrial scenarios. First, reshaping the agent's memory.
Recently, knowledge graphs and ontologies have regained attention, which is closely related to the memory problem. Currently, a large number of open-source projects adopt the form of knowledge graphs to organize and store agent memory.
Why choose a knowledge graph? This is because it is well-suited for describing semantic relationships between knowledge points. As a network structure, a knowledge graph can connect different forms of data modalities and various types of data in different links, realizing the association of multi-dimensional data. Therefore, using ontology to model memory is an effective technical means.
In conventional large model memory solutions, when the data volume is too large, vectorization is usually adopted to convert various modal information such as images, text, sensor data, and interaction records into vectors and store them in a vector database. During task execution, the agent retrieves relevant memories from the vector database. However, this retrieval method based on semantic similarity is difficult to completely restore the context of events. Through ontology modeling, all memories can be connected into a complete knowledge network, where temporal information is particularly critical, and the knowledge graph happens to effectively express such temporal relationships. From the various data collected and situations occurring on a certain machine at the T1 time node, to the changes that appear at the T2 node, all can be completely recorded in the knowledge graph. Recording memories in this form can more accurately restore the real situation of the physical world. Second, setting constraints. In addition to reshaping memory, it is also necessary to further standardize the agent's adherence to constraints during the task planning process.
When traditional models perform planning, they mainly rely on their internalized knowledge. This knowledge originates from the massive data obtained from the Internet during model training and multiple rounds of pre-training and post-training. Fine-tuning is usually not performed because the fine-tuning cost is high and it is difficult to obtain sufficient data locally. Therefore, the physical world rules learned by such models on the Internet, when applied within enterprises, especially in industrial manufacturing scenarios, often exceed their capability scope.
A more ideal approach is to use ontology for constraints. When an agent generates a task plan based on the current situation, each planning step needs to be queried and verified in the knowledge graph.
The key here is to store machine-executable actions and physical knowledge in the knowledge graph.
When the agent executes a certain action, it can query the preconditions, post-requirements, and relevant physical constraints of the action in the graph. This information needs to be preset into the system through product manuals, expert experience, and other channels before the agent is put into the production line.
During task planning and execution, the agent can dynamically trigger these physical constraints and iteratively optimize the planning scheme based on them to make it more aligned with the actual needs of the enterprise.
As the constraints are continuously enriched and improved, the execution capability and final success rate of the agent will also be significantly enhanced.
Third, tool invocation. We recommend incorporating tools and all interfaces into the graph and uniformly modeling them through ontology.
The current standardized methods for agent tool invocation mainly include MCP, CLI, and the widely used Skill. However, regardless of the method adopted, when the number of interfaces reaches hundreds or even thousands, it is difficult for the agent to achieve precise selection, and problems are also prone to occur in interface assembly and parameter invocation.
If the interface capability is described in only one sentence, the depth of the description is insufficient; when the number of interfaces increases, the description accuracy decreases, and the large model is prone to errors during the selection process; a better approach is to put it into the graph.
All three aspects mentioned above achieve knowledge normalization through ontology: machines, processes, constraints at each node, and callable interfaces within the enterprise are all incorporated into the graph.
When the agent searches for tools along the graph, the accuracy is higher. The knowledge graph is equivalent to a map, providing navigation guidance; compared with global search, it has significant improvements in both efficiency and accuracy.
03 - Construction of Knowledge Agents Based on Knowledge Graphs
When building industrial-grade agents, it needs to be emphasized: the knowledge base and the agent are inseparable, and neither can fully exert its efficacy if separated. Based on this, we call it a knowledge agent.
In the deployment process of agents, a large amount of work is not writing specific business logic, but sorting out the internal knowledge of the enterprise.
Enterprise knowledge can be divided into three categories.
The first category is explicit knowledge, which comes from various informatization systems within the enterprise, including accumulated documents, pictures, videos, and other unstructured data.
The second category is tacit knowledge. This type of knowledge is often ignored but is crucial in industrial manufacturing enterprises. A large amount of production experience exists in the practical experience of senior technicians, which belongs to the category of tacit knowledge. It needs to be recognized and precipitated by agents through various methods, and can be obtained through meeting minutes, technical forums, training lectures, personal notes, and even chat records.
The third category is external knowledge. In specific business scenarios (such as law, scientific research, etc.), data may need to be obtained from the Internet.
The effective integration and utilization of the above three types of knowledge are critical for agents. To introduce the three types of knowledge into the system, an effective data integration mechanism needs to be established, including business system integration, online document integration, and the precipitation of expert experience, relying on a complete system to achieve comprehensive data collection.
After data collection is completed, knowledge processing is required to make the knowledge callable by agents. The core idea is to guide the construction of an enterprise-level knowledge graph through ontology modeling, thereby providing multi-dimensional knowledge support for agents. This process involves a series of technical means, including OCR, translation, annotation, vectorization, etc., extracting effective knowledge from the three types of data through multiple models to build the knowledge graph. The construction of the knowledge graph also requires the aid of specialized tools. There are some open-source tools on the market, but their design focuses on serving agents or machines, with insufficient support for manual operation and maintenance.
However, at this stage, to achieve precise and controllable knowledge graphs within the enterprise, the mode of human-machine co-construction still needs to be adopted. Human experts need to have sufficient control over the knowledge graph, so perfect tool support is needed, including modeling tools, human-machine collaborative tools for knowledge extraction, fusion tools, etc. These tools are not only for machine use; human experts also need to be deeply involved to build an accurate and practical knowledge graph.
We recommend that manufacturing enterprises build a knowledge graph centered on their own products, first sorting out knowledge around the products, and then gradually supplementing knowledge from various links such as R&D design, manufacturing to sales and services, to make the graph increasingly perfect. However, it is not necessary to build a large and comprehensive graph before it can be put into use; it is advisable to adopt a progressive construction strategy, building and using simultaneously.
Two methods are recommended for knowledge management: explicit knowledge is realized by importing internal enterprise documents, pictures, videos, and informatization system data; tacit knowledge is recommended to be precipitated in the form of a Wiki, forming experience entries, and continuously collecting tacit knowledge within the enterprise.
Relying on this technical architecture, search capabilities can achieve cross-modal penetration, covering design files, CAD files, 3D files, pictures, videos, and other content, which has a significant improvement compared to traditional name-based retrieval methods.
The question-and-answer function is also indispensable. Through simple questioning, relevant materials and answers can be retrieved in the knowledge base.
In addition, agents also have application value in the following scenarios. For example, process document review: a large number of personnel in manufacturing enterprises are engaged in repetitive work, requiring frequent adjustments and modifications when new demands arrive. The quality of process guidebooks written by designers is uneven; although the upper limit is high, the lower limit is low, and low-quality guidebooks are prone to causing production problems.
Agents can integrate enterprise standards, product knowledge, and the experience of senior experts to review existing guidebooks, avoiding the repeated occurrence of similar errors; they can also directly write process guidebooks, after-sales operation and maintenance plans, quality inspection specifications, and other documents.
The reason why agents can directly write documents stems from the systematic accumulation of knowledge in the early stage. After the process guidebook is completed, a quality inspection plan can be generated based on it, clarifying precautions and mandatory indicators to achieve systematic output of documents.
In addition, agents can be used for pre-sales support. When customer demands arise, product features can be mined from the knowledge base, and multi-dimensional evaluation and scoring can be conducted for different products that meet the demands, outputting a complete report for sales personnel to refer to when communicating with customers.
In manufacturing enterprises, the workload of R&D and design personnel is usually heavy. They need to provide technical guidance to production line workers and also support the sales and marketing teams. However, senior expert resources are limited, and it is a waste of resources for them to engage in repetitive work.
On the other hand, agents also have application value at the production line end. When production line workers encounter problems, they usually contact designers via telephone or instant messaging tools, but designers are busy with work, and the response is often not timely enough.
With the help of knowledge precipitation, basic agents can be built, supporting functions such as voice input, on-site picture and video capture. If difficult problems are encountered and cannot be solved, the complete chat records and relevant data can be submitted to human experts to apply for support, and the experts will make judgments and provide key guidance.
If the problem is ultimately resolved, the process is transformed into an experience. In the past, such experiences were scattered in instant messaging records or calls, making it difficult to precipitate them into the knowledge base; however, through agents, efficient accumulation of experiences can be realized.
04 - Three Practical Cases in Industrial Manufacturing Scenarios
Case 1: Cross-shift troubleshooting of large rotating equipment.
Production personnel found that the No. 2 compressor tripped, and the rotor vibration value suddenly increased to 85 mm. What are the possible causes?
Without the memory function, if this question is directly asked to the agent, the agent may only retrieve similar problems in the manual and fault diagnosis experience library and conduct a simple analysis. Due to the lack of combination with historical memory, its answer may have a large deviation.
If the memory function is available, the system has recorded the recent events and operation history of the No. 2 compressor. The ontology records the hierarchical relationship from equipment to components, component abnormal events, and corresponding manual operations. By inputting the complete historical record as memory into the agent, a more accurate judgment can be obtained.
It can be seen from this case that the system records show that intervention was carried out and lubricating oil was added the previous day due to a sudden drop in pressure, from which it can be inferred that the current fault is related to the lubricating oil. Based on this chain of evidence, a more accurate conclusion can be drawn.
Ultimately, the repair time is significantly shortened, and the risk of scrapping precision equipment is also significantly reduced.
Case 2: Planning enhancement application in dynamic process scheduling of electrode coating lines
The process engineer found that the viscosity of the current batch of slurry is slightly higher than the previous one, but still within the allowable range, so the equipment parameters need to be adjusted.
If this task is directly assigned to the agent, due to the lack of precise basis, the agent may directly adjust the pump pressure to 10%, but this operation is actually infeasible, and the agent is unaware of this constraint.
Now we incorporate all constraints and operations into the knowledge graph. For example, the pump speed constraint in the SOP: any pump speed adjustment operation corresponds to a constraint node, limiting the single adjustment amplitude to not exceed 5%, and the thickness of the first piece must be measured after adjustment, which is a mandatory post-process.
Under the guidance of graph constraints, the agent can generate a more scientific plan: first increase the pump pressure by 5%, then measure the thickness of the first piece, and then adjust the air temperature, thereby ensuring that the process advances in a more controllable direction.
In this case, the probability of batch scrapping under similar conditions is reduced to nearly zero, and the planning generation time is significantly shortened.
The accumulation of such empirical common sense is crucial; the more, the better. There is no need to wait for the system to be complete; they can be entered one by one. When the agent performs task planning, entering the corresponding node will forcibly trigger constraints and pre-processing and post-processing actions, thereby ensuring the scientific nature of the execution.
Case 3: Landing deduction of the global knowledge management system based on precise tool invocation
We integrate the APIs of systems such as PLM, MES, and ERP into the knowledge graph. When a user issues a deductive command, such as a typhoon warning, the system can locate and invoke precise APIs by penetrating multiple systems, and ultimately output the links and content that may be affected.
With the help of the precise control of the graph, the success rate of long-chain continuous invocation of 4 to 5 interfaces is increased from 50% to 99%, and the cross-system scheduling time is shortened from half a day to 30 seconds.
05 - Crossing the Valley of Death: A Three-Step Approach to the Deployment of Industrial Agents
The deployment process of industrial agents can be called "crossing the valley of death". Looking back at 2023 and 2024, the industry focused on the private deployment of large models; from 2024 to 2025, it gradually shifted to knowledge base construction and application development supported by large models; starting from 2026, it began to try agent applications, but the process will inevitably experience many challenges.
Based on our practical experience, it is recommended to advance in three stages.
Stage 1: Sorting out knowledge assets. In the preparation stage, it is necessary to systematically sort out the internal knowledge assets of the enterprise, which are the foundation for constraining agent behavior and improving its performance. From the perspective of implementation, accumulating in the form of ontology or knowledge graph is highly targeted, but the specific form can be flexibly selected. The key is to clearly sort out the dictionary, scenarios, and core physical constraints.
Stage 2: Building the framework and running through a single scenario. Build an agent framework (either open-source or commercial), connect the memory module with the common sense sorted out in the first stage, select specific scenarios (production, after-sales, design, etc.) for verification, and ensure that a single scenario runs through.
Stage 3: Horizontal replication and vertical evolution. In the horizontal dimension, achieve batch replication and expand the scenarios that have run through to more fields. In the vertical dimension, promote the agent to gradually transition to the stage of automatic evolution: the experience of human experts is continuously precipitated in the system, and the precipitated experience feeds back into the agent, forming a virtuous cycle. After entering this cycle, the probability of project success will be significantly increased.
That is all for this sharing. If you need to obtain the complete courseware and video playback, please scan the QR code to get it.
Ph.D. from Southeast University, with main research directions in knowledge graphs and natural language processing. He has published more than ten high-level papers and obtained more than ten invention patents related to knowledge graphs and large models. He has served as the head of NLP at NetEase Hangzhou Institute of Engineering and the Technical Director of Shexing Intelligence, successfully developing multiple phenomenal AI products and winning awards such as the "Gold Award for Outstanding Artificial Intelligence Products in Jiangsu Province" and the "Top Ten Outstanding Artificial Intelligence Software Products". The developed knowledge graph-enhanced large model products have currently been successfully deployed in more than twenty Fortune Global 500 enterprises.