A multi-agent operating system for enterprise computing

Abstract

This invention discloses a multi-agent operating system for enterprise computing, belonging to the fields of artificial intelligence and enterprise information technology. The system adopts a five-layer architecture design: infrastructure layer, basic capability layer, agent runtime layer, multi-agent collaboration layer, and enterprise application layer. It also provides methods for multi-agent communication, knowledge graph construction and multi-path reasoning, and enterprise scenario adaptation, enabling standardized lifecycle management of agents, efficient collaborative communication between agents, multi-path reasoning decision-making based on enterprise knowledge, and rapid adaptation to cross-industry scenarios. This invention achieves unified management and continuous self-evolution of enterprise-level agents, improves resource utilization and decision accuracy, shortens scenario deployment cycles, and ensures system security and stability. It can be widely applied to intelligent enterprise scenarios in various industries such as healthcare, industrial manufacturing, and financial services.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and enterprise information technology, and in particular to a multi-agent operating system for enterprise computing. Background Technology

[0002] With the rapid development of artificial intelligence technology, Large Language Models (LLM) and multi-agent systems have become important technological drivers for enterprise digital transformation. In enterprise application scenarios, agent technology is widely used in various fields such as intelligent customer service, data analysis, document processing, and business process automation. However, existing enterprise intelligent solutions face the following technical challenges: The fragmentation of intelligent agents: Various AI applications within enterprises (such as customer service robots, data analysis tools, document processing systems, supply chain management, etc.) operate independently, lacking a unified operating system for management, scheduling, and collaboration. Each intelligent agent uses an independent operating environment and communication mechanism, resulting in capability silos, severe resource waste, and an inability to achieve complex task collaboration across intelligent agents.

[0003] Low efficiency in collaborative communication: Existing communication mechanisms between agents are mostly point-to-point direct calls (such as HTTP REST or gRPC), lacking a unified message bus and standardized communication protocols. When a task involves collaboration among multiple agents, the communication topology is complex, message passing efficiency is low, and advanced communication capabilities such as task coordination, resource negotiation, and conflict resolution are lacking.

[0004] Difficulty in scenario adaptation: Different industries and scenarios have significantly different needs for enterprise intelligence. For example, medical scenarios require strict compliance review and multimodal perception capabilities, manufacturing scenarios require real-time scheduling and equipment collaboration capabilities, and financial scenarios require high security and risk reasoning capabilities. Existing systems lack flexible scenario adaptation mechanisms, requiring extensive customized development each time they enter a new scenario, resulting in high costs and long cycles.

[0005] Lack of knowledge management: Enterprise knowledge is scattered across multiple sources such as documents, databases, and personnel experience, lacking automated knowledge graph construction and multi-path reasoning capabilities. Existing intelligent agents mainly rely on general knowledge from LLM (Local Management Model), lacking efficient utilization of enterprise-specific knowledge, resulting in limited decision-making quality and poor performance in specialized domains.

[0006] Inadequate lifecycle management: Existing systems lack systematic management of the lifecycle stages of agents, including creation, registration, scheduling, operation, self-evolution, and recycling. Agents cannot automatically optimize based on operational feedback, and their knowledge and experience cannot be persisted and reused.

[0007] In summary, existing technologies cannot build a multi-agent operating system that meets the intelligent needs of enterprises, and there is an urgent need for an integrated solution that can achieve "unified management, efficient collaboration, intelligent decision-making, and scenario adaptation". Summary of the Invention

[0008] The purpose of this invention is to provide a multi-agent operating system for enterprise computing, which solves the core problems of fragmentation of enterprise agents, low collaboration efficiency, insufficient knowledge utilization, and difficulty in scenario adaptation.

[0009] To achieve the above objectives, this invention provides a multi-agent operating system for enterprise computing, employing a five-layer architecture design, from bottom to top: infrastructure layer, basic capability layer, agent runtime layer, multi-agent collaboration layer, and enterprise application layer; wherein: The infrastructure layer provides computing resources, container orchestration, network communication, persistent storage, and monitoring and alerting. It adopts a microservice architecture and containerized deployment, and supports elastic scaling and high availability deployment. The basic capability layer includes an LLM inference engine, a knowledge graph module, a vector database, a message queue, and a security sandbox. The LLM inference engine integrates several large language models and provides a unified inference interface. The intelligent agent runtime layer provides a standardized runtime environment for each intelligent agent, including a perception module, a cognitive engine, a decision-making module, an execution module, and a learning module. The multi-agent collaboration layer includes a task decomposition engine, an agent registration center, a collaborative scheduler, a resource negotiation module, and a conflict resolver. The task decomposition engine uses an LLM-based task planning algorithm to decompose complex tasks into a sub-task network represented by a directed acyclic graph. The enterprise application layer provides standardized application interfaces for specific enterprise business scenarios.

[0010] Preferably, the LLM inference engine supports mainstream large language models through a unified model adapter, and supports automatic model selection and load balancing based on task type. The mainstream large language models include the GPT series, GLM series, and Llama series. The intelligent agent runtime layer adopts the Actor model design, with each intelligent agent running as an independent Actor in a secure sandbox, and is configured with standardized lifecycle management interfaces such as init(), run(), learn(), and shutdown() for the intelligent agents; The agent registration center maintains a global capability directory and uses capability profile vectors to represent the capability features of agents, supporting agent capability matching and dynamic discovery based on semantic similarity.

[0011] Preferably, the knowledge graph module is built on a graph database and supports efficient querying of hundreds of millions of nodes; the vector database is implemented using either Milvus or FAISS and supports millisecond-level similarity retrieval of billions of vectors. The security sandbox provides each agent with an independent, isolated operating environment, supporting resource quota limits and process-level isolation.

[0012] Based on the aforementioned operating system, this invention also provides a multi-agent communication method, comprising the following steps: (1) Establish a multi-agent message bus as the core hub for communication between agents, and adopt a hybrid communication mode that combines publish / subscribe and point-to-point communication; (2) Five standardized communication protocols are implemented on the multi-agent message bus, namely, task protocol, negotiation protocol, notification protocol, state synchronization protocol, and security authentication protocol. Among them, the task protocol defines the standard message format for task publication, acceptance, execution, and reporting; the negotiation protocol adopts a negotiation mechanism based on contract network to support the complete negotiation process of bidding-tendering-winning-signing; the notification protocol realizes event notification, state change broadcasting, and alarm push; the state synchronization protocol adopts an incremental synchronization mechanism to periodically synchronize the agent's running status; and the security authentication protocol adopts JWT-based token authentication and TLS1.3-based communication encryption.

[0013] Preferably, the multi-agent message bus adopts a layered design, including a transport layer and a message layer; the transport layer supports two transport protocols, TCP and WebSocket, and automatically selects the optimal protocol according to the communication scenario, using WebSocket for long connection scenarios and TCP for batch transmission scenarios.

[0014] Preferably, the message layer defines a unified message format, and each message includes a message header and a message body. The message header includes a message ID, sender ID, receiver ID, timestamp, and priority, and the message body includes a task type, message content, and additional parameters.

[0015] Based on the aforementioned operating system, this invention also provides a knowledge graph construction and multi-path reasoning method, comprising the following steps: S1. Multi-source data acquisition: Raw data is acquired in parallel from enterprise documents, business databases, external knowledge bases, and interaction logs; S2, NLP Preprocessing and Entity Extraction: The raw data is segmented, named entity recognition is performed, relation extraction is performed, and attribute extraction is performed. The NER model is fine-tuned based on the BERT architecture. S3. Knowledge Graph Construction: The extracted entities, relations and attributes are stored in the graph database in the form of RDF triples; S4. Vectorized storage: The text content is converted into a vector representation through an embedding model and stored in a vector database; S5. Multi-path reasoning: It integrates three paths—rule-based forward chain reasoning, graph neural network-based graph reasoning, and LLM-based comprehensive reasoning—to be executed in parallel. The reasoning results are integrated through a weighted fusion mechanism, and the fusion weights are dynamically adjusted based on the reasoning confidence and historical accuracy. S6. Intelligent Decision Output: Generates recommendation suggestions, risk warnings, and intelligent question-and-answer responses based on the fused reasoning results; S7. Feedback Learning and Knowledge Update: Based on decision-making effect feedback data, automatically update the knowledge graph and optimize the inference model parameters.

[0016] Preferably, the entity relationships in the knowledge graph include four categories: hierarchical relationships, association relationships, temporal relationships, and causal relationships, and a reasoning update module is configured to periodically clean up outdated and redundant knowledge in the knowledge graph.

[0017] Based on the aforementioned operating system, this invention also provides an enterprise scenario adaptation method, comprising the following steps: T1. Scene Feature Extraction and Analysis: Perform multi-dimensional analysis of business features, data features, and interaction features of the enterprise scenarios accessed; T2. Agent Template Matching: Select the combination of agent templates that best matches the enterprise scenario from the pre-set agent template library. The template library covers five types of agent templates: management, analysis, execution, knowledge, and perception. T3. Dynamic assembly of capability components: Based on the needs of enterprise scenarios, the five modules of perception, cognition, decision-making, execution and learning of the intelligent agent are dynamically selected and their parameters are configured. T4. Collaboration Strategy Generation: Based on enterprise scenario characteristics and template combinations, automatically generate the optimal task allocation strategy, communication topology, and conflict resolution rules; T5. Deployment Configuration and Initialization: Automatically complete container orchestration, resource configuration, knowledge base initialization, and API interface exposure; T6. Operational Data Acquisition and Iterative Optimization: Continuously collect system operation data and optimize collaborative strategies and agent parameters through A / B testing and reinforcement learning algorithms.

[0018] Preferably, the intelligent agent template library contains pre-set intelligent agent templates and best practice configurations for seven major industry scenarios: healthcare, industrial manufacturing, financial services, education and research, retail e-commerce, smart cities, and supply chain management.

[0019] Therefore, this invention employs the aforementioned multi-agent operating system for enterprise computing, effectively addressing industry pain points such as fragmented enterprise agents and low collaborative efficiency. It achieves unified management of enterprise-level agents through a five-layer architecture, eliminating capability silos and improving resource utilization. Relying on standardized communication protocols and message buses, it significantly improves collaborative communication efficiency and reduces communication latency. A scenario adaptation mechanism enables rapid cross-industry deployment, reducing customized development workload and shortening the deployment cycle from months to days. Multi-path reasoning and knowledge graph technologies enhance the accuracy of agent decision-making. Combined with full lifecycle management, it enables continuous system self-evolution, while designs such as a security sandbox ensure enterprise data security and stable system operation.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a diagram illustrating the overall architecture of the multi-agent operating system according to an embodiment of the present invention. Figure 2 This is a flowchart of the intelligent agent lifecycle management according to an embodiment of the present invention; Figure 3 This is a diagram of a multi-agent collaborative communication architecture according to an embodiment of the present invention; Figure 4 This is a flowchart of the knowledge graph construction and reasoning process according to an embodiment of the present invention; Figure 5 This is a flowchart of the enterprise scenario adaptation engine workflow according to an embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0024] The following is in conjunction with the appendix Figure 1-5The specific implementation of the present invention will be described in detail. This embodiment takes two typical enterprise scenarios, namely medical and health care and industrial manufacturing, as the core, and combines technical solutions such as multi-agent communication, knowledge graph construction and reasoning, enterprise scenario adaptation and agent lifecycle management.

[0025] Example 1: A Multi-Agent Operating System for Enterprise Computing like Figure 1 As shown, the system adopts a five-layer architecture design, from bottom to top as follows: 1. Infrastructure layer Deployed on a cloud computing platform (supporting public, private, and hybrid clouds), it employs the Kubernetes container orchestration system for service governance, supporting automatic elastic scaling and high-availability deployment. Persistent storage utilizes a hybrid approach combining a distributed file system and a relational database. The monitoring and alerting module leverages Prometheus and Grafana for end-to-end monitoring, supporting agent-level resource utilization monitoring and anomaly alerts.

[0026] 2. Basic Capability Level LLM Inference Engine: Supports multiple large language models (including GPT, GLM, and Llama series) through a unified model adapter, supports automatic model selection and load balancing based on task type, and supports hot model switching. Knowledge Graph Module: Built on the Neo4j graph database, supports efficient querying of hundreds of millions of nodes, storing enterprise knowledge entities, relationships between entities, and inference results. Vector Database: Implemented using Milvus or FAISS, supports millisecond-level similarity retrieval of billions of vectors, with vector dimensions adaptable to the embedding output specifications of different large models. Message Queue: Based on a publish / subscribe pattern using Kafka or RabbitMQ, providing asynchronous message communication infrastructure. Security Sandbox: Based on container-level isolation technology, providing an independent running environment for each agent, supporting resource quota limits and process-level isolation.

[0027] 3. Agent runtime layer The system employs an Actor model, where each agent runs as an independent actor within a secure sandbox, providing standardized lifecycle management interfaces (init(), run(), learn(), shutdown()). At runtime, each agent receives five standardized modules: a perception module receives external multimodal input (text, speech, images, structured data); a cognition engine performs information understanding and semantic analysis based on LLM and knowledge graphs; a decision-making module generates action strategies based on cognitive results and task objectives; an execution module translates decisions into concrete actions such as API calls, data manipulation, or message sending; and a learning module fine-tunes the model, updates knowledge, and optimizes collaborative strategies based on runtime feedback.

[0028] 4. Multi-agent collaboration layer Task Decomposition Engine: Employs an LLM-based task planning algorithm to automatically decompose complex enterprise tasks into a network of subtasks represented by a Directed Acyclic Graph (DAG), determining the dependencies and execution order between subtasks. Agent Registry: Maintains a global agent registry, using capability profile vectors to represent the capability characteristics of each agent (including functional tags, domain knowledge, performance metrics, and other multi-dimensional features), supporting capability matching and dynamic discovery based on semantic similarity. Cooperative Scheduler: Assigns subtasks to the most suitable agent for execution based on capability matching and resource constraints. Resource Negotiation Module: Negotiates and dynamically allocates computing resources (CPU, GPU, memory) among agents. Conflict Resolution Module: Resolves conflicts based on priority rules and negotiation mechanisms when multiple agents experience resource contention or task conflicts.

[0029] 5. Enterprise Application Layer It provides standardized application interfaces (RESTful API and WebSocket) for specific business scenarios, supporting various enterprise applications such as intelligent hospital management, supply chain optimization, customer service, production scheduling, and knowledge management, and supports access from various front-end clients (Web, mobile, and IoT devices).

[0030] Example 2: Overall Deployment and Application of Multi-Agent Operating System in Hospital Intelligent Management Scenarios This embodiment is based on the five-layer architecture operating system described in Embodiment 1, and combines it with the intelligent hospital management scenario to achieve end-to-end deployment, adapting to the business needs of the entire hospital treatment process, multidisciplinary consultation, medication safety review, etc.

[0031] 1. Infrastructure Layer Deployment The system adopts a hybrid cloud deployment model for hospitals, with core services deployed in the hospital's local data center (4 servers with 16 cores and 64GB of RAM, configured with NVIDIA T4 GPUs), and elastically expandable nodes connected to the public cloud. Container orchestration and service governance are achieved through Kubernetes, supporting elastic scaling and high-availability deployment of agent services. Persistent storage adopts a hybrid solution of distributed file system and relational database to store hospital business data, agent operation logs, and knowledge graph data. The monitoring and alarm module is based on Prometheus + Grafana to achieve full-link monitoring, supporting agent-level CPU / GPU / memory usage monitoring and anomaly alarms, meeting the hospital's 24 / 7 business operation requirements.

[0032] 2. Basic Capability Layer Configuration LLM Inference Engine: Integrates the GLM-4 language model through a unified model adapter, and achieves automatic model selection and load balancing based on task types such as consultation, prescription review, etc. Local deployment of GLM-4 ensures medical data privacy. Knowledge Graph Module: Based on the Neo4j graph database, a medical knowledge graph is built, storing hundreds of millions of nodes and their relationships, including diseases, symptoms, examination items, and drugs, and supporting millisecond-level efficient queries; Vector database: Implemented using Milvus, with dimensionality adapted to 1024-dimensional embedding model output, storing billions of vectors such as medical literature and prescription records, and supporting millisecond-level semantic similarity retrieval; Message Queues and Security Sandboxes: Implement asynchronous message communication using a publish / subscribe pattern based on Kafka; employ a security sandbox with container-level isolation technology to allocate an independent running environment to each agent, limiting resource quotas to 4 CPU cores and 16GB of memory, and achieving process-level isolation.

[0033] 3. Intelligent agent runtime layer configuration The Actor model is adopted, configuring five types of intelligent agents for the hospital scenario: outpatient management, laboratory analysis, medication safety, patient follow-up, and resource scheduling. Each type of intelligent agent runs as an independent Actor in a secure sandbox, configured with standardized lifecycle management interfaces such as init(), run(), learn(), and shutdown(). Specifically, in this embodiment: Outpatient Management Intelligent Agent (Capability Tags: Triage, Registration, Waiting Room Management; Model: GLM-4); Laboratory Analysis Intelligent Agent (Capability Tags: Laboratory Report Interpretation, Anomaly Warning; Integrated Medical Knowledge Graph); Medication Safety Intelligent Agent (Capability Tags: Prescription Review, Drug Interaction Detection; Integrated Drug Knowledge Graph); Patient Follow-up Intelligent Agent (Capability Tags: Remote Follow-up, Health Management; Supports Multimodal Interaction via Text and Voice); Resource Scheduling Intelligent Agent (Capability Tags: Bed Management, Operating Room Scheduling, Equipment Allocation; Optimized Algorithm Based on Real-Time Data and Historical Statistics).

[0034] 4. Multi-agent collaboration layer workflow Task Decomposition Engine: This engine decomposes the complex patient visit process into a directed acyclic graph (DAG) sub-task network using an LLM-based task planning algorithm. The sub-tasks are symptom description, initial assessment, examination scheduling, prescription review, bed allocation, and postoperative follow-up. The dependencies between these sub-tasks are clearly defined. Specifically: The patient describes symptoms through the outpatient management agent → the outpatient management agent invokes the laboratory analysis agent for preliminary assessment via the task protocol of the message bus → the laboratory analysis agent recommends examinations based on knowledge graph reasoning → the outpatient management agent arranges the examinations → after the test results are returned, the medication safety agent collaborates with the laboratory analysis agent to review the prescription via a negotiation protocol → the resource scheduling agent receives the hospitalization arrangement task and allocates beds via a notification protocol → after the patient is discharged, the patient follow-up agent takes over subsequent management. The entire process is completed collaboratively through five protocols of the message bus.

[0035] 5. Enterprise application layer integration It provides standardized RESTful API + WebSocket application interfaces for various business departments of the hospital, connects to the hospital's core systems such as HIS, LIS, and PACS, and supports access from front-end clients such as doctor workstations, hospital mobile terminals, and self-service machines, realizing the full-scenario implementation of intelligent hospital management, covering core businesses such as outpatient management, inpatient scheduling, medication safety, and patient follow-up.

[0036] Example 3: Agent Lifecycle Management like Figure 2 As shown, taking the creation of a new "Medical Quality Control Analysis Intelligent Agent" as an example, the complete lifecycle management process is demonstrated: S1. Requirements Definition and Modeling: Based on the requirements of the hospital's quality control department, define the functional scope of the intelligent agent (medical quality data analysis, abnormal indicator early warning, and quality control report generation).

[0037] S2. Capability Profile Generation: The system automatically generates capability profile vectors, including functional tags (quality control analysis, data statistics, report generation), domain knowledge (medical quality management standards, JCI accreditation standards), and performance indicators.

[0038] S3. Register to the agent registry: Register the capability profile to the agent registry so that it can be discovered and invoked by the coordinator.

[0039] S4. Resource Allocation and Initialization: The co-scheduler allocates computing resources of 4 CPU cores, 16GB of memory, and 1 GPU, and initializes the runtime environment in a secure sandbox.

[0040] S5. Join the task pool and wait for scheduling: The agent joins the global task pool and begins to receive quality control data analysis tasks.

[0041] S6. Task Execution and Collaborative Interaction: When the agent executes the quality control analysis task, it calls the inspection analysis agent through the message bus to obtain inspection data and calls the knowledge graph to query the quality control standards.

[0042] S7. Self-evolution judgment and execution: After running for 30 days, the system evaluation finds that the analysis accuracy of some rare indicators is lower than the threshold, triggering the self-evolution process to fine-tune the model based on the analysis data of the most recent 30 days.

[0043] S8. Knowledge Update and Performance Optimization: After self-evolution is completed, the system will update the optimized quality control-related rules and model parameters to the knowledge graph and basic capability layer to achieve knowledge persistence and reuse.

[0044] Example 4: Multi-agent cooperative communication like Figure 3 As shown, taking the "emergency multidisciplinary consultation" scenario as an example, the specific process of collaborative communication is demonstrated.

[0045] A multi-agent message bus is established as the core hub for communication between agents, adopting a hybrid communication mode that combines publish / subscribe and point-to-point communication. The message bus adopts a layered design of transport layer and message layer. The transport layer supports two transport protocols, TCP and WebSocket, and can automatically select the optimal protocol according to the communication scenario. The message layer defines a unified message format. Each message contains a message header and a message body. The message header contains message ID, sender ID, receiver ID, timestamp, and priority. The message body contains task type, message content, and additional parameters.

[0046] After the outpatient management intelligent agent identifies a case as high-risk, it triggers a multidisciplinary consultation process and issues an emergency consultation task via the message bus. (1) Task Agreement: The outpatient management intelligent agent publishes multidisciplinary consultation tasks through the message bus, defines the relevant content of task publication, claiming, execution and reporting according to the standard message format, and clarifies the core requirements of the task; for example, the message header includes priority = urgent and deadline = 15 minutes; (2) Notification Protocol: The system sends event notifications to the corresponding intelligent agents in the relevant departments through broadcast mode to complete the broadcast of status changes and the push of alarms; for example, it broadcasts to notify the relevant intelligent agents in the cardiology, thoracic surgery and radiology departments. (3) Negotiation Agreement: Based on the negotiation mechanism of the contract network, the relevant intelligent agents complete the complete negotiation process of the execution order and resource allocation of the corresponding medical operations; for example, the radiology intelligent agent and the laboratory analysis intelligent agent negotiate the execution order and resource allocation of image analysis and blood test through the contract network mechanism; (4) Status synchronization protocol: An incremental synchronization mechanism is adopted, and each participating intelligent agent periodically synchronizes the task progress and patient-related status to the message bus; (5) Security authentication protocol: All communications are authenticated with JWT tokens and encrypted with TLS 1.3 to ensure patient data security.

[0047] Throughout the consultation process, multiple intelligent agents achieve efficient collaborative communication through a message bus, which meets the efficient business needs of hospital emergency scenarios compared to traditional communication methods.

[0048] Example 5: Knowledge Graph Construction and Multi-Path Reasoning This embodiment, based on knowledge graph construction and multi-path reasoning methods, uses hospital medication safety review as a specific scenario to achieve efficient utilization of enterprise private knowledge and intelligent decision-making. (Attached) Figure 4 A flowchart for knowledge graph construction and reasoning is provided. The specific implementation steps are as follows: S1. Multi-source data acquisition: Raw data related to hospital medication is collected in parallel from enterprise documents, business databases, external knowledge bases, and interaction logs; this embodiment is based on drug instructions (OCR parsing of 100,000+ PDF documents), hospital prescription database (ETL daily synchronization of 5 million+ historical prescriptions), National Medical Products Administration knowledge base (API interface to obtain drug approval information in real time), and adverse drug event reports (collected in real time).

[0049] S2. NLP Preprocessing and Entity Extraction: The raw data is segmented, named entity recognition, relation extraction, and attribute extraction are performed. The NER model is based on the BERT architecture and fine-tuned to accurately identify various entities and relationships related to medication. In this embodiment, the NLP pipeline automatically identifies entities and relationships such as drug name (e.g., "amoxicillin"), indication (e.g., "bacterial infection"), contraindications (e.g., "penicillin allergy"), and drug interactions (e.g., "increased bleeding risk when used in combination with warfarin").

[0050] S3. Knowledge Graph Construction: The extracted entities, relations, and attributes are stored in the graph database in the form of RDF triples to construct a medical knowledge graph, which includes four types of entity relations: hierarchical relations, association relations, temporal relations, and causal relations. S4. Vectorized storage: The text content related to medication is converted into a vector representation through an embedding model and stored in a vector database.

[0051] S5. Multi-path Reasoning: This system integrates three parallel reasoning paths—rule-based forward chaining, graph neural network-based graph reasoning, and LLM-based comprehensive reasoning—to comprehensively review prescriptions. The results of these three reasoning paths are integrated through a weighted fusion mechanism, with the fusion weights dynamically adjusted based on reasoning confidence and historical accuracy. In this embodiment, the three reasoning paths execute in parallel: the rule engine checks for drug dosage exceeding limits, graph reasoning checks for contraindications between drugs, and LLM reasoning comprehensively assesses individualized patient risks (such as liver and kidney function). The weighted fusion of the three results generates a review report, automatically marking abnormal prescriptions and notifying the doctor. Review records are fed back to update the knowledge graph.

[0052] S6. Intelligent Decision Output: Generates risk warnings and recommendations related to medication safety review based on the fused reasoning results, and completes the intelligent decision output. S7. Feedback Learning and Knowledge Update: Based on the feedback data of the decision-making effect of medication safety review, the medical knowledge graph is automatically updated and the parameters of the inference model are optimized; at the same time, through the inference update module, outdated and redundant knowledge in the knowledge graph is cleaned up regularly.

[0053] Example 6: Application of Enterprise Scenario Adaptation Method in Industrial Manufacturing Scenarios This embodiment, based on the aforementioned enterprise scenario adaptation method, rapidly adapts the multi-agent operating system of the present invention from the healthcare scenario to the industrial manufacturing enterprise scenario. Figure 5 Adapt the engine workflow diagram for enterprise scenarios. The specific implementation steps are as follows: T1. Scene Feature Extraction and Analysis: The system connects with relevant systems of manufacturing enterprises and performs multi-dimensional analysis on the connected manufacturing scenarios, including business characteristics (production scheduling process, supply chain management requirements), data characteristics (MES system data, ERP data, IoT sensor data format and volume), and interaction characteristics (workshop operators using mobile terminals, response time requirements in seconds), to clarify the core needs of the manufacturing scenarios.

[0054] T2. Agent Template Matching: Select the combination of agent templates that best matches the manufacturing scenario from the pre-set agent template library; such as production scheduling agent, quality inspection agent, supply chain management agent, equipment maintenance agent, and data analysis agent.

[0055] T3. Dynamic Assembly of Capability Components: Based on the actual needs of the manufacturing scenario, the five modules of perception, cognition, decision-making, execution and learning of the intelligent agent are dynamically selected and their parameters are configured; for example, an IoT sensor data perception module and a fault prediction and decision-making module are added to the equipment maintenance intelligent agent.

[0056] T4. Collaborative Strategy Generation: Based on the characteristics of the manufacturing scenario and template combinations, the system automatically generates the optimal task allocation strategy, communication topology, and conflict resolution rules; that is, the production scheduling agent acts as the management node, and other agents work collaboratively through task protocols and notification protocols.

[0057] T5. Deployment, Configuration and Initialization: The system automatically completes container orchestration and resource configuration, initializes the relevant knowledge base of the manufacturing enterprise and exposes API interfaces, and connects with the enterprise's existing business systems; T6. Operational Data Acquisition and Iterative Optimization: The system continuously collects operational data in manufacturing scenarios, optimizes collaborative strategies and agent parameters through A / B testing and reinforcement learning algorithms, and achieves continuous improvement in system performance.

[0058] Therefore, this invention employs the aforementioned multi-agent operating system for enterprise computing, effectively addressing industry pain points such as fragmented enterprise agents and low collaborative efficiency. It achieves unified management of enterprise-level agents through a five-layer architecture, eliminating capability silos and improving resource utilization. Relying on standardized communication protocols and message buses, it significantly improves collaborative communication efficiency and reduces communication latency. A scenario adaptation mechanism enables rapid cross-industry deployment, reducing customized development workload and shortening the deployment cycle from months to days. Multi-path reasoning and knowledge graph technologies enhance the accuracy of agent decision-making. Combined with full lifecycle management, it enables continuous system self-evolution, while designs such as a security sandbox ensure enterprise data security and stable system operation.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-agent operating system for enterprise computing, characterized in that, The architecture adopts a five-layer design, from bottom to top: infrastructure layer, basic capability layer, intelligent agent runtime layer, multi-agent collaboration layer, and enterprise application layer; among which: The infrastructure layer provides computing resources, container orchestration, network communication, persistent storage, and monitoring and alerting. It adopts a microservice architecture and containerized deployment, and supports elastic scaling and high availability deployment. The basic capability layer includes an LLM inference engine, a knowledge graph module, a vector database, a message queue, and a security sandbox. The LLM inference engine integrates several large language models and provides a unified inference interface. The intelligent agent runtime layer provides a standardized runtime environment for each intelligent agent, including a perception module, a cognitive engine, a decision-making module, an execution module, and a learning module. The multi-agent collaboration layer includes a task decomposition engine, an agent registration center, a collaborative scheduler, a resource negotiation module, and a conflict resolver. The task decomposition engine uses an LLM-based task planning algorithm to decompose complex tasks into a sub-task network represented by a directed acyclic graph. The enterprise application layer provides standardized application interfaces for specific enterprise business scenarios.

2. The multi-agent operating system for enterprise computing according to claim 1, characterized in that, The LLM inference engine supports mainstream large language models through a unified model adapter, and supports automatic model selection and load balancing based on task type. The mainstream large language models include the GPT series, GLM series, and Llama series. The intelligent agent runtime layer adopts the Actor model design, with each intelligent agent running as an independent Actor in a secure sandbox, and is configured with standardized lifecycle management interfaces such as init(), run(), learn(), and shutdown() for the intelligent agents; The agent registration center maintains a global capability directory and uses capability profile vectors to represent the capability features of agents, supporting agent capability matching and dynamic discovery based on semantic similarity.

3. The multi-agent operating system for enterprise computing according to claim 1, characterized in that, The knowledge graph module is built on a graph database and supports efficient querying of hundreds of millions of nodes; the vector database is implemented using either Milvus or FAISS and supports millisecond-level similarity retrieval of billions of vectors. The security sandbox provides each agent with an independent, isolated operating environment, supporting resource quota limits and process-level isolation.

4. A multi-agent communication method, applied to the enterprise computing-oriented multi-agent operating system as described in any one of claims 1-3, characterized in that, Includes the following steps: (1) Establish a multi-agent message bus as the core hub for communication between agents, and adopt a hybrid communication mode that combines publish / subscribe and point-to-point communication; (2) Five standardized communication protocols are implemented on the multi-agent message bus, namely, task protocol, negotiation protocol, notification protocol, state synchronization protocol, and security authentication protocol. Among them, the task protocol defines the standard message format for task publication, acceptance, execution, and reporting; the negotiation protocol adopts a contract network-based negotiation mechanism to support the complete negotiation process of bidding, tendering, winning the bid, and signing the contract; the notification protocol realizes event notification, state change broadcasting, and alarm push; the state synchronization protocol adopts an incremental synchronization mechanism to periodically synchronize the agent's running status; and the security authentication protocol adopts JWT-based token authentication and TLS 1.3-based communication encryption.

5. The multi-agent communication method according to claim 4, characterized in that, The multi-agent message bus adopts a layered design, including a transport layer and a message layer. The transport layer supports two transport protocols, TCP and WebSocket, and automatically selects the optimal protocol according to the communication scenario. WebSocket is used for long connection scenarios, and TCP is used for batch transmission scenarios.

6. The multi-agent communication method according to claim 5, characterized in that, The message layer defines a unified message format. Each message contains a message header and a message body. The message header contains the message ID, sender ID, receiver ID, timestamp, and priority. The message body contains the task type, message content, and additional parameters.

7. A knowledge graph construction and multi-path reasoning method, applied to the enterprise computing-oriented multi-agent operating system as described in any one of claims 1-3, characterized in that, Includes the following steps: S1. Multi-source data acquisition: Raw data is acquired in parallel from enterprise documents, business databases, external knowledge bases, and interaction logs; S2, NLP Preprocessing and Entity Extraction: The raw data is segmented, named entity recognition is performed, relation extraction is performed, and attribute extraction is performed. The NER model is fine-tuned based on the BERT architecture. S3. Knowledge Graph Construction: The extracted entities, relations and attributes are stored in the graph database in the form of RDF triples; S4. Vectorized storage: The text content is converted into a vector representation through an embedding model and stored in a vector database; S5. Multi-path reasoning: It integrates three paths—rule-based forward chain reasoning, graph neural network-based graph reasoning, and LLM-based comprehensive reasoning—to be executed in parallel. The reasoning results are integrated through a weighted fusion mechanism, and the fusion weights are dynamically adjusted based on the reasoning confidence and historical accuracy. S6. Intelligent Decision Output: Generates recommendation suggestions, risk warnings, and intelligent question-and-answer responses based on the fused reasoning results; S7. Feedback Learning and Knowledge Update: Based on decision-making effect feedback data, automatically update the knowledge graph and optimize the inference model parameters.

8. The knowledge graph construction and multi-path reasoning method according to claim 7, characterized in that, The entity relationships in the knowledge graph include four categories: hierarchical relationships, association relationships, temporal relationships, and causal relationships. It is also equipped with a reasoning update module, which periodically cleans up outdated and redundant knowledge in the knowledge graph.

9. An enterprise scenario adaptation method, applied to the multi-agent operating system for enterprise computing as described in any one of claims 1-3, characterized in that, Includes the following steps: T1. Scene Feature Extraction and Analysis: Perform multi-dimensional analysis of business features, data features, and interaction features of the enterprise scenarios accessed; T2. Agent Template Matching: Select the combination of agent templates that best matches the enterprise scenario from the pre-set agent template library. The template library covers five types of agent templates: management, analysis, execution, knowledge, and perception. T3. Dynamic assembly of capability components: Based on the needs of enterprise scenarios, the five modules of perception, cognition, decision-making, execution and learning of the intelligent agent are dynamically selected and their parameters are configured. T4. Collaboration Strategy Generation: Based on enterprise scenario characteristics and template combinations, automatically generate the optimal task allocation strategy, communication topology, and conflict resolution rules; T5. Deployment Configuration and Initialization: Automatically complete container orchestration, resource configuration, knowledge base initialization, and API interface exposure; T6. Operational Data Acquisition and Iterative Optimization: Continuously collect system operation data and optimize collaborative strategies and agent parameters through A / B testing and reinforcement learning algorithms.

10. The enterprise scenario adaptation method according to claim 9, characterized in that, The intelligent agent template library contains pre-set intelligent agent templates and best practice configurations for seven major industry scenarios: healthcare, industrial manufacturing, financial services, education and research, retail e-commerce, smart cities, and supply chain management.

Images