An AI Bus intelligent agent interconnection bus implementation method
By using the AI Bus architecture and scheduling agent optimization algorithm, plug-and-play and dynamic task scheduling of multi-agent systems are achieved, solving the problems of communication complexity and uneven resource allocation in existing technologies, improving system scalability and execution efficiency, and reducing operation and maintenance costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-02
- Publication Date
- 2026-06-16
AI Technical Summary
Existing communication schemes for multi-agent systems suffer from complex point-to-point connections, uneven resource allocation, low execution efficiency, inability to achieve plug-and-play functionality, and inability to dynamically allocate tasks, resulting in poor system scalability and stability.
Adopting an AI Bus architecture, the system enables plug-and-play intelligent agents through components such as MCP Server, scheduling agent, MCP Host, AgentCards, and RPA process engine. It combines scheduling agent with optimization algorithms to perform intelligent scheduling of the entire task process. The system introduces a Sidecar service mesh agent to separate communication logic from business logic, achieving dynamic fault tolerance and standardized protocols to ensure system security and stability.
It significantly improves system scalability and maintainability, increases task execution efficiency by 1.5 times, reduces operation and maintenance costs by 90%, and ensures the system's communication reliability and security in complex network environments.
Smart Images
Figure CN122226541A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multi-agent system technology, and more specifically, to an AI Bus intelligent agent interconnection bus implementation method. Background Technology
[0002] With the rapid development of artificial intelligence and distributed systems technology, multi-agent systems, with their distributed collaboration capabilities, have become the core support for realizing intelligent operation in complex business scenarios. To achieve collaborative interaction among multiple agents, two main communication schemes have emerged in existing technologies. One is the agent interaction scheme based on direct communication protocols. This scheme uses standardized communication protocols such as FIPA and ACL to achieve point-to-point direct interaction between agents. Its core logic is that each agent pre-stores the communication addresses and interface specifications of other interacting agents, and transmits instructions, data, and task requests through dedicated direct connections. It is typically used in small to medium-sized fixed-topology multi-agent systems, such as distributed equipment collaboration systems in early industrial control scenarios. The other is the asynchronous communication scheme based on simple message queues. To alleviate the strong coupling problem caused by direct communication, this scheme introduces message middleware such as RabbitMQ and Kafka as relay nodes. Agents do not need to establish direct connections; instead, they achieve asynchronous message sending and receiving in a producer-consumer pattern through message queues. This is suitable for scenarios with large data volumes but simple task logic, such as distributed data distribution systems in internet services.
[0003] The two types of solutions mentioned above meet the basic communication needs of multiple agents in specific scenarios, but they also have the following shortcomings: In the direct communication solution, point-to-point connections need to be established between agents, and the number of connections in the system increases exponentially with the number of agents n, which is n(n-1) / 2. Any change in the address, interface adjustment, or offline of any agent will cause the communication of associated agents to be interrupted, and re-adaptation is required. Even in the simple message queue solution, agents still need to be bound to the topic and format of a specific queue. When adding or removing agents, the queue rules need to be reconfigured, and plug-and-play functionality cannot be achieved.
[0004] Existing communication carriers only handle data transmission and cannot understand task requirements and agent capabilities. Complex tasks require manual pre-setting of processes, and tasks cannot be dynamically allocated based on the agent's real-time load and response speed, leading to uneven resource allocation and low execution efficiency. Therefore, there is an urgent need to provide an AI Bus agent interconnection bus implementation method to solve the above problems. Summary of the Invention
[0005] To address the aforementioned technical issues, this invention provides an AI Bus intelligent agent interconnection bus implementation method. This invention breaks system coupling limitations through the AI Bus bus architecture, enabling plug-and-play intelligent agents and significantly improving system scalability; it combines scheduling intelligent agents with optimization algorithms to achieve intelligent scheduling of the entire task process, improving execution efficiency; full-link monitoring reduces operation and maintenance costs, and dynamic fault tolerance and standardized protocols ensure the safe and stable operation of the system.
[0006] To achieve the above objectives, the technical solution of the present invention is as follows:
[0007] An AI Bus intelligent agent interconnection bus implementation method includes the following steps:
[0008] S1. Establish the core communication hub of AI Bus, deploy MCP Server, scheduling agent, MCP Host, AgentCards, RPA process engine, and rule engine, configure A2A adaptation protocol and Sidecar service mesh proxy, and complete the communication link between the components; the AI Bus provides a unified communication and collaboration mechanism, supporting high-quality dataset fusion agents, AI tools, rule engine, and RPA process engine to interact through standardized protocols; the scheduling agent, as the core component of MCP Server, is responsible for service dialogue requests and scheduling between agents; the MCP Host is the running carrier of agents and tools, and Agents 1-n and AI Tools access AI Bus through the MCP Host;
[0009] S2. New intelligent agents or AI tools initiate a registration request through the MCP Host, which is then transmitted to the MCP Server via the AI Bus to complete standardized registration and access verification, and to include qualified intelligent agents or tools into the resource pool.
[0010] S3. The scheduling agent receives user service requests, identifies task requirements and expected output formats based on a large language model, and associates Agent Cards to initially screen the range of matching agents.
[0011] S4. The scheduling agent uses AI Bus to schedule the RPA process engine, rule engine and related agents to retrieve related data and rule information from multi-dimensional data sources and aggregate them to form a high-quality dataset.
[0012] S5. The scheduling agent inputs the task request and high-quality dataset into the large language model, generates a task list based on the agent's ability description and decomposes it into sub-tasks, and combines Agent Cards with real-time resource status to match and adapt the agent / tool sequence.
[0013] S6. The scheduling agent sends instructions to the target MCP Host through the AI Bus, specifying the agent / tool or RPA process template to execute subtasks. The Sidecar agent monitors the communication status and performs circuit breaker retry.
[0014] S7: The scheduling agent monitors resource load and task progress in real time. When resource changes, timeout risks, or failures occur, it triggers task reallocation and backup resource switching.
[0015] S8. Each execution unit feeds back the results through the AI Bus, and after the scheduling agent verifies and aggregates them, they are transmitted to the large language model to generate answers, or to the reporting agent to generate reports, or to the dialogue agent to output.
[0016] As a preferred embodiment of the present invention, when building the AI Bus core communication hub in S1, a bus architecture is adopted and a standardized interaction module with A2A adaptation protocol is integrated, and a unified communication protocol specification for CPU, memory, and network resources is preset; the Sidecar service mesh agent is deployed separately for each MCP Host, and traffic monitoring, data encryption, and circuit breaker retry functions are bound to achieve the separation of communication logic and business logic, and ensure the real-time and secure transmission of cross-module data and instructions.
[0017] As a preferred embodiment of the present invention, the core component deployment and scheduling agent initialization in S1 further includes the following specific configurations:
[0018] The deployed MCP Server is configured with full lifecycle management functions for intelligent agents, builds a dual-dimensional intelligent agent registration directory for business and functional domains, supports access control and security policy configuration, and schedules intelligent agents to establish communication links with large language models, AgentCards, RPA process engines, and rule engines.
[0019] Deployed Agent Cards can uniformly identify and encapsulate the capabilities of existing intelligent applications with different business capabilities. During system integration and business orchestration, the standardized capabilities carried by Agent Cards can be reused directly without repeated development, effectively improving integration efficiency and reducing development costs.
[0020] The deployed RPA process engine pre-registers data retrieval and task execution process templates. Each template corresponds to an independent process service interface and completes MCP Server registration. The rule engine configures the agent to call the rule base to support the large language model's learning and inference calling logic.
[0021] During the initialization of the scheduling agent, task priority weight rules and resource load thresholds are imported. And fault self-healing strategies, clearly defining dynamic task sets. Available resource set Define the core variables and complete the parameter configuration of the constraints and objective function of the autonomous scheduling optimization algorithm.
[0022] As a preferred embodiment of the present invention, the standardized registration and resource pool inclusion process of intelligent agents / tools in S2 specifically includes:
[0023] After a new intelligent agent / tool is deployed, it initiates a registration request to the AI Bus through its MCP Host. The request carries the core information of the intelligent agent / tool, including a unique code, service interface address, input / output parameter format, capability tags, and invocation rules.
[0024] AI Bus performs communication compatibility verification on registration requests based on preset protocol specifications, completes preliminary traffic routing configuration through the Sidecar proxy, and after the verification is passed, the registration request is passed to the MCP Server.
[0025] MCP Server updates available resource set The agent / tool is incorporated into the system resource pool, and the agent is scheduled to perceive resource pool updates in real time through AI Bus, so that it can obtain the capability attributes of the newly added resources without modifying the existing component configuration.
[0026] As a preferred embodiment of the present invention, the specific process of multi-source resource retrieval and data foundation construction in S4 is as follows:
[0027] The scheduling agent sends retrieval commands to the RPA process engine, rule engine, and related data analysis agents via AI Bus, specifying the retrieval dimensions and data requirements. The RPA process and rule engine work together to retrieve business data, rule information, and knowledge nodes that are strongly related to the task from multi-dimensional data sources such as the enterprise data source pool, local industry knowledge graph, and mechanism model library. All retrieval results are aggregated to the scheduling agent via AI Bus, and after deduplication, correlation, and formatting, they are aggregated to form the data foundation for task analysis.
[0028] As a preferred embodiment of the present invention, the specific process of task decomposition and constraint parameter and resource status processing in S5 is as follows:
[0029] Task decomposition: The scheduling agent synchronously inputs task requests and high-quality datasets into the large language model. Based on the agent capability descriptions in Agent Cards, it infers and generates a task list, decomposing it into a sequence of executable sub-tasks, and defining a dynamic task set. ;
[0030] The constraints are clearly defined: each subtask is specified. Deadline Priority weight and resource demand vector ;
[0031] Resource status collection: Real-time collection of various resources The state parameters are used to generate a resource availability vector. Calculate the normalized load rate of computing resources And satisfy Load safety constraints;
[0032] Task priority ranking: based on formula Calculate the urgency of the task, where The execution order is obtained by sorting the baseline processing time when resources are idle, arranged in ascending order. .
[0033] As a preferred embodiment of the present invention, the specific process of resource adaptability calculation and scheduling decision generation in S5 is as follows:
[0034] Resource suitability calculation: using the formula Calculate the fit between subtasks and resources; the higher the value, the better the fit.
[0035] Scheduling decision generation: Generating allocation decision variables based on autonomous scheduling optimization algorithms And it satisfies the following constraints:
[0036] Resource feasibility constraints: , ;
[0037] Task uniqueness constraint: , ;
[0038] Deadline constraints: , ;
[0039] Resource conflict adjustment: If resource conflicts exist, adjust them using the formula. Adjustments were made, including This represents the conflict penalty coefficient.
[0040] As a preferred embodiment of the present invention, the specific process of dynamic scheduling and fault self-healing in S7 is as follows:
[0041] Resource status and task time updates: by perception cycle Through formula Dynamic decay of occupied resources; through formula Real-time correction of estimated task processing time, among which This is the baseline processing time when resources are idle;
[0042] Dynamic scheduling is triggered when any of the following conditions are met:
[0043] New mission arrives: ;
[0044] Resource state mutation: ,in ;
[0045] Task timeout risk: ,in Second;
[0046] Fault self-healing adjustment: If resources In the event of a sudden malfunction, the assigned tasks Reassign to the standby resource with the highest fit and that meets the deadline constraint. ,Right now and .
[0047] As a preferred embodiment of the present invention, the specific process of result aggregation and output in S8 is as follows:
[0048] Results Feedback and Verification: After each execution unit completes its subtask, it feeds back the execution results, including intermediate data and analysis conclusions, to the scheduling agent through the AI Bus.
[0049] Answer generation and output: If the task requires direct answer output, the scheduling agent passes the result set to the large language model, which combines the task context to generate a complete answer with business value; if the task requires report output, the scheduling agent pushes the result set to the reporting agent to generate a standardized report, and finally outputs the corresponding results to the user through the dialogue agent.
[0050] As a preferred embodiment of the present invention, it further includes S9, the operation and maintenance monitoring stage, and achieves dynamic optimization through an autonomous scheduling optimization algorithm. The specific process is as follows:
[0051] Full-link trajectory recording: The scheduling agent records the complete trajectory of a task from request reception, task decomposition, agent invocation, result aggregation to output. The trajectory includes task decomposition logic, agent invocation order, response data, execution status, and allocation decision variables. The changes;
[0052] Communication metrics collection: The Sidecar agent collects communication metrics in real time, including traffic, latency, and error rate. After collection, the metrics are synchronized to the monitoring dashboard to visualize the communication status.
[0053] Integrated management and fault location: By combining the business data recorded by the scheduling agent and the communication indicators collected by the sidecar, an integrated observable system for business and communication is constructed; when task execution anomalies or communication failures occur, the fault nodes can be quickly located based on the trajectory records and indicator data. The fault nodes include agent capability mismatch, network fluctuations, and resource overload, supporting rapid fault diagnosis and system optimization.
[0054] Dynamic optimization is achieved through an autonomous scheduling optimization algorithm. The objective function of the algorithm is:
[0055] ,in The weighting coefficients are used to minimize the overall scheduling cost while satisfying hard constraints, ensuring efficient and stable system operation.
[0056] The beneficial technical effects of this invention are:
[0057] This invention constructs a unified communication hub through an AI Bus architecture, where all agents, tools, and engines interact indirectly through standardized protocols, reducing system connectivity complexity from traditional solutions. Down to This completely breaks down the coupling limitations of point-to-point communication or simple message queues. Simultaneously, the Sidecar service mesh agent separates communication logic from business logic. Functions such as traffic monitoring, encryption, and circuit breaking are independent of the core business of the agent. Changes to communication rules only require updating the Sidecar configuration, without modifying the agent code. Combined with the Agent Cards standardized registration mechanism, new agents / tools only need to complete core information registration to be automatically discovered and invoked by the system. The expansion cycle is shortened from days to minutes, significantly improving system maintainability and dynamic scalability.
[0058] By introducing a scheduling agent and an autonomous scheduling optimization algorithm, combined with the task reasoning capabilities of a large language model, the entire process of task parsing, decomposition, resource matching, and dynamic adjustment is made intelligent. Through a task priority ranking algorithm and a resource suitability calculation model, tasks and resources are accurately matched, avoiding resource overload or idleness. Simultaneously, the algorithm minimizes task timeout penalties, resource waste, and load imbalance through an objective function. Under hard constraints such as deadlines and resource load thresholds, it improves the efficiency of complex task decomposition and resource matching by 1.5 times, significantly shortening task response time.
[0059] The scheduling agent records the complete trajectory of a task from request reception, decomposition, invocation to output, including task logic, invocation order, execution status, and changes in decision variables. The Sidecar agent collects and visualizes communication metrics such as traffic, latency, and error rate in real time. When task anomalies or communication failures occur, the fault node can be quickly located through trajectory recording and metric data, reducing troubleshooting time from hours to minutes, improving system problem-solving efficiency by 90%, and significantly reducing operation and maintenance costs.
[0060] By dynamically updating resource status and task processing time through a sensing cycle, the system automatically triggers task reallocation when new tasks arrive, resource status changes abruptly, or there is a risk of task timeout. In the event of a sudden resource failure, the system can quickly switch tasks to the most suitable backup resource, preventing a chain reaction of failures. Through the A2A adaptation protocol, a unified communication standard for CPU, memory, and network resources is pre-defined. Combined with the data encryption function of the Sidecar agent, secure transmission of data and instructions across modules is achieved. Simultaneously, the Sidecar's circuit breaker and retry mechanism automatically handles communication anomalies, preventing data loss or transmission interruptions and ensuring the system's communication reliability in complex network environments, providing a secure and stable communication guarantee for multi-agent collaboration. Attached Figure Description
[0061] Figure 1 This is a diagram showing the overall components and connections of the AI Bus system of this invention.
[0062] Figure 2 This is a schematic diagram of the AI Bus intelligent agent interconnection bus system architecture of the present invention.
[0063] Figure 3 This is a flowchart of the scheduling agent of the present invention performing task reasoning and scheduling based on a large model. Detailed Implementation
[0064] In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0065] Combination Figure 1-3 The present invention provides the following embodiments:
[0066] An AI Bus intelligent agent interconnection bus implementation method includes the following steps:
[0067] S1. Establish the core communication hub of AI Bus, deploy MCP Server, scheduling agent, MCP Host, AgentCards, RPA process engine, and rule engine, configure A2A adaptation protocol and Sidecar service mesh proxy, and complete the communication link between the components; the AI Bus provides a unified communication and collaboration mechanism, supporting high-quality dataset fusion agents, AI tools, rule engine, and RPA process engine to interact through standardized protocols; the scheduling agent, as the core component of MCP Server, is responsible for service dialogue requests and scheduling between agents; the MCP Host is the running carrier of agents and tools, and Agents 1-n and AI Tools access AI Bus through the MCP Host;
[0068] S2. New intelligent agents or AI tools initiate a registration request through the MCP Host, which is then transmitted to the MCP Server via the AI Bus. The MCP Server extracts the core information of the intelligent agent, generates corresponding capability cards, and stores them in Agent Cards. This completes standardized registration and access verification, and includes qualified intelligent agents or tools in the resource pool.
[0069] S3. The scheduling agent receives user service requests, identifies task requirements and expected output formats based on a large language model, and associates Agent Cards to initially screen the range of matching agents.
[0070] S4. The scheduling agent uses AI Bus to schedule the RPA process engine, rule engine and related agents to retrieve related data and rule information from multi-dimensional data sources and aggregate them to form a data foundation.
[0071] S5. The scheduling agent inputs the task request and high-quality dataset into the large language model, generates a task list based on the agent's ability description and decomposes it into sub-tasks, and combines Agent Cards with real-time resource status to match and adapt the agent / tool sequence.
[0072] S6. The scheduling agent sends instructions to the target MCP Host through the AI Bus, specifying the agent / tool or RPA process template to execute subtasks. The Sidecar agent monitors the communication status and performs circuit breaker retry.
[0073] S7: The scheduling agent monitors resource load and task progress in real time. When resource changes, timeout risks, or failures occur, it triggers task reallocation and backup resource switching.
[0074] S8. Each execution unit feeds back the results through the AI Bus, and after the scheduling agent verifies and aggregates them, they are transmitted to the large language model to generate answers, or to the reporting agent to generate reports, or to the dialogue agent to output.
[0075] MCP Server and Agent Cards are two sets of resources, both of which communicate with intelligent applications through AI Bus. MCP Server is used to integrate resources and provide high-quality datasets for intelligent agents, while Agent Cards are used to enable "plug and play" capabilities for existing intelligent agents.
[0076] Agents in Agent Cards communicate via a gridded A2A mechanism and do not have a direct communication relationship with the MCP Server. Their capabilities are linked together by new agents that invoke and execute commands through large language model nodes. The large language model nodes, in turn, have their business processing logic completed by agent building tools, and they handle heterogeneous and multi-source resource calls according to standard construction methods.
[0077] The role of the RPA process engine is to orchestrate the priority strategy of service calls, while the role of the rules engine is to distinguish the intent and identify the corresponding agent for the call.
[0078] The role of the scheduling agent is divided into two cases according to the recognized intent: the first is to dynamically call the corresponding agent based on the result returned by the rule engine; the second is to realize the dynamic calling of service path routing calculation based on the result of the RPA process engine.
[0079] Furthermore, when building the AI Bus core communication hub in S1, a bus architecture is adopted and a standardized interaction module with A2A adaptation protocol is integrated, and a unified communication protocol specification for CPU, memory, and network resources is preset. The Sidecar service mesh agent is deployed separately for each MCP Host, and is bound to traffic monitoring, data encryption, and circuit breaker retry functions to achieve the separation of communication logic and business logic, and ensure the real-time and secure transmission of cross-module data and instructions.
[0080] The MCP Host serves as the operating platform for agents and tools, mounting Agent 1-n and AI Tools onto the corresponding MCP Host nodes and configuring the access link between the MCP Host and the AI Bus to ensure that all agents / tools can receive scheduling instructions through the MCP Host.
[0081] The bus architecture uses AI Bus as a unified hub, reducing the connection complexity of traditional point-to-point communication from... Down to This approach addresses the challenge of requiring system-wide adaptation for adding, removing, or changing configurations of intelligent agents at the architectural level. The standardized interaction module of the A2A adaptation protocol unifies the communication specifications of CPU, memory, and network resources, ensuring the interaction compatibility of different types of intelligent agents, tools, and engines. The Sidecar service mesh agent decouples communication assurance functions from business logic, monitors traffic in real-time to perceive transmission status, encrypts data to ensure interaction security, and uses a circuit breaker retry mechanism to automatically handle communication anomalies. These three elements work together to solve the pain point in existing technologies where communication security and stability depend on the design of the intelligent agent itself, achieving efficient and reliable transmission of data and instructions across modules.
[0082] Furthermore, the initialization of the core component deployment and scheduling agent in S1 also includes the following specific configurations:
[0083] The deployed MCP Server is configured with full lifecycle management functions for intelligent agents, builds a dual-dimensional intelligent agent registration directory for business and functional domains, supports access control and security policy configuration, and schedules intelligent agents to establish communication links with large language models, AgentCards, RPA process engines, and rule engines.
[0084] Deployed Agent Cards can uniformly identify and encapsulate the capabilities of existing intelligent applications with different business capabilities. During system integration and business orchestration, the standardized capabilities carried by Agent Cards can be reused directly without repeated development, effectively improving integration efficiency and reducing development costs.
[0085] The deployed RPA process engine pre-registers data retrieval and task execution process templates. Each template corresponds to an independent process service interface and completes MCP Server registration. The rule engine configures the agent to call the rule base to support the large language model's learning and inference calling logic.
[0086] During the initialization of the scheduling agent, task priority weight rules and resource load thresholds are imported. And fault self-healing strategies, clearly defining dynamic task sets. Available resource set Define the core variables and complete the parameter configuration of the constraints and objective function of the autonomous scheduling optimization algorithm.
[0087] Furthermore, the standardized registration and resource pool inclusion process for intelligent agents or tools in S2 specifically includes:
[0088] After a new intelligent agent or tool is deployed, it initiates a registration request to the AI Bus through its MCP Host. The request carries the core information of the intelligent agent or tool, including a unique code, service interface address, input / output parameter format, capability tags, and calling rules.
[0089] AI Bus performs communication compatibility verification on registration requests based on preset protocol specifications, completes preliminary traffic routing configuration through the Sidecar proxy, and after the verification is passed, the registration request is passed to the MCP Server.
[0090] MCP Server updates available resource set The agent / tool is incorporated into the system resource pool, and the agent is scheduled to perceive resource pool updates in real time through AI Bus, so that it can obtain the capability attributes of the newly added resources without modifying the existing component configuration.
[0091] The core information carried in the registration request provides a complete basis for the identification and capability matching of intelligent agents or tools, ensuring that their adaptability characteristics can be accurately obtained during subsequent scheduling; the protocol compatibility verification of AI Bus and the traffic routing configuration of Sidecar screen qualified access objects at the communication level, avoiding interaction failures caused by protocol incompatibility; the real-time update of the resource pool and the perception mechanism for scheduling intelligent agents enable plug-and-play of new resources, solving the problem that intelligent agent expansion requires modification of the entire system configuration in existing technologies. The core logic is to establish a unified management and automatic discovery mechanism for intelligent agent capabilities through standardized registration process and dynamic synchronization with resource pool.
[0092] Furthermore, the specific process of multi-source resource retrieval and data foundation construction in S4 is as follows:
[0093] The scheduling agent sends retrieval commands to the RPA process engine, rule engine, and related data analysis agents via AI Bus, specifying the retrieval dimensions and data requirements. The RPA process engine and rule engine work together to retrieve business data, rule information, and knowledge nodes that are strongly related to the task from multi-dimensional data sources such as the enterprise data source pool, local industry knowledge graph, and mechanism model library. All retrieval results are aggregated to the scheduling agent via AI Bus, and after deduplication, correlation, and formatting, they are aggregated to form the data foundation for task analysis.
[0094] The scheduling agent clearly defines the retrieval dimensions and requirements, avoiding blindness in the retrieval process; multi-dimensional data sources cover core information such as business data and knowledge nodes, ensuring the comprehensiveness of retrieval results; deduplication, association, and formatting processes eliminate data redundancy and format differences, enabling scattered retrieval results to form a structured and consistent data foundation, providing a complete and reliable data analysis basis for subsequent task decomposition and agent matching, and solving the problem of lack of unified data support for task processing in existing technologies.
[0095] Furthermore, the specific process of task decomposition, constraint parameter handling, and resource status processing in S5 is as follows:
[0096] Task decomposition: The scheduling agent synchronously inputs task requests and high-quality datasets into the large language model. Based on the agent capability descriptions in Agent Cards, it infers and generates a task list, decomposing it into a sequence of executable sub-tasks, and defining a dynamic task set. ;
[0097] The constraints are clearly defined: each subtask is specified. Deadline Priority weight , and resource demand vector ;
[0098] Resource status collection: Real-time collection of various resources The state parameters are used to generate a resource availability vector. Calculate the normalized load rate of computing resources And satisfy Load safety constraints;
[0099] Task priority ranking: based on formula Calculate the urgency of the task, where The execution order is obtained by sorting the baseline processing time when resources are idle, arranged in ascending order. .
[0100] The large language model, combined with a data foundation and agent capability description, decomposes tasks and leverages its semantic understanding and logical reasoning capabilities to transform complex tasks into a sequence of executable subtasks, addressing the pain point of existing technologies that require manual pre-setting of complex task processes. Clearly defined deadlines, priority weights, and resource requirement vectors provide quantitative constraints for task scheduling, ensuring precise matching between resource allocation and task requirements. Real-time resource status acquisition and load rate calculation dynamically reflect resource occupancy and load thresholds. The settings prevent resource overload from causing a decrease in execution efficiency; the task priority sorting formula combines priority weights and deadline pressure, so that urgent and important tasks are executed first, optimizing the task execution order and improving the overall task completion efficiency.
[0101] Furthermore, the specific process of resource suitability calculation and scheduling decision generation in S5 is as follows:
[0102] Resource suitability calculation: using the formula Calculate the fit between subtasks and resources; the higher the value, the better the fit.
[0103] Scheduling decision generation: Generating allocation decision variables based on autonomous scheduling optimization algorithms And it satisfies the following constraints:
[0104] Resource feasibility constraints: , ;
[0105] Task uniqueness constraint: , ;
[0106] Deadline constraints: , ;
[0107] Resource conflict adjustment: If resource conflicts exist, adjust them using the formula. Adjustments were made, including This represents the conflict penalty coefficient.
[0108] In the resource suitability formula, the numerator is obtained through... The average resource satisfaction rate is calculated, and the denominator incorporates the estimated task processing time, comprehensively reflecting the matching degree between resources and subtasks. This ensures that allocated resources not only meet task requirements but also possess efficient processing capabilities. Three constraints limit the scheduling boundaries from three core dimensions: resource supply, task execution uniqueness, and time nodes, avoiding unreasonable resource allocation or task execution timeouts. The resource conflict adjustment formula uses a conflict penalty coefficient. Resources that are already carrying too many tasks are penalized, and tasks are guided to be allocated to resources with lighter loads and higher adaptability, thus resolving the conflict problem of multiple tasks competing for the same resource and achieving balanced utilization of resources.
[0109] The specific steps of S6 are as follows: The scheduling agent sends task execution instructions to the target MCP Host through the AI Bus, specifying the subtask allocation scheme, execution priority and output requirements, and specifying the corresponding agent / tool or RPA process template to execute the subtask; The AI Bus transmits task parameters and data based on a standardized protocol, and the Sidecar agent monitors communication traffic, latency and error rate in real time. When a communication anomaly is detected, the circuit breaker retry mechanism is automatically triggered; After receiving the instructions, the execution unit on the MCP Host focuses on the core business logic to execute the subtask and feeds back the real-time execution status to the scheduling agent through the AI Bus.
[0110] Furthermore, the specific process of dynamic scheduling and fault self-healing in S7 is as follows:
[0111] Resource status and task time updates: by perception cycle For example, 1 second, through the formula Dynamic decay of occupied resources; through formula Real-time correction of estimated task processing time, among which This is the baseline processing time when resources are idle;
[0112] Dynamic scheduling is triggered when any of the following conditions are met:
[0113] New mission arrives: ;
[0114] Resource state mutation: ,in ;
[0115] Task timeout risk: ,in Second;
[0116] Fault self-healing adjustment: If resources Unexpected malfunction, The tasks assigned to them Reassign to the standby resource with the highest fit and that meets the deadline constraint. ,Right now and .
[0117] Perception cycle The system's settings enable real-time updates of resource status and task processing time. The dynamic resource decay formula accurately reflects resource consumption during task execution, while the task processing time correction formula, combined with resource load rate, makes processing time prediction more closely match the actual operating status. Dynamic scheduling trigger conditions cover three core scenarios: new task access, abnormal resource fluctuations, and task timeout risks, ensuring the system can respond promptly to dynamic changes and avoid resource waste or task timeouts caused by static scheduling. The fault self-healing adjustment logic uses the highest adaptability and deadline constraints as dual standards to quickly migrate tasks on faulty resources to backup resources, preventing chain-like fault propagation. This addresses the pain point in existing technologies where a single agent failure affects the execution of global tasks, improving the system's fault tolerance and availability.
[0118] Furthermore, the specific process of result aggregation and output in S8 is as follows:
[0119] Results Feedback and Verification: After each execution unit completes its subtask, it feeds back the execution results, including intermediate data and analysis conclusions, to the scheduling agent through the AI Bus.
[0120] Answer generation and output: If the task requires direct answer output, the scheduling agent passes the result set to the large language model, which combines the task context to generate a complete answer with business value; if the task requires report output, the scheduling agent pushes the result set to the reporting agent to generate a standardized report, and finally outputs the corresponding results to the user through the dialogue agent.
[0121] The execution results include intermediate data and analysis conclusions, providing a complete basis for subsequent aggregation and verification. The integrity verification of the scheduling agent ensures no results are missed, and the consistency verification eliminates contradictory or abnormal data, ensuring the reliability of the result set. The scenario-based output design based on task requirements adapts to different user scenarios. The large language model generates complete answers in combination with context, leveraging its language organization and logical integration capabilities to enhance the business value of the answers. The reporting agent generates standardized reports to meet the needs of structured display. Finally, the dialogue agent achieves efficient output, solving the problems of messy task results and single output format in existing technologies.
[0122] Furthermore, it also includes the S9 and operation and maintenance monitoring phases, and achieves dynamic optimization through an autonomous scheduling optimization algorithm. The specific process is as follows:
[0123] Full-link trajectory recording: The scheduling agent records the complete trajectory of a task from request reception, task decomposition, agent invocation, result aggregation to output. The trajectory includes task decomposition logic, agent invocation order, response data, execution status, and allocation decision variables. The changes;
[0124] Communication metrics collection: The Sidecar agent collects communication metrics in real time, including traffic, latency, and error rate. After collection, the metrics are synchronized to the monitoring dashboard to visualize the communication status.
[0125] Integrated management and fault location: By combining the business data recorded by the scheduling agent and the communication indicators collected by the sidecar, an integrated observable system for business and communication is constructed; when task execution anomalies or communication failures occur, the fault nodes can be quickly located based on the trajectory records and indicator data. The fault nodes include agent capability mismatch, network fluctuations, and resource overload, supporting rapid fault diagnosis and system optimization.
[0126] Dynamic optimization is achieved through an autonomous scheduling optimization algorithm. The objective function of the algorithm is:
[0127] ,in The weighting coefficients are used to minimize the overall scheduling cost while satisfying hard constraints, ensuring efficient and stable system operation.
[0128] The entire link trajectory records the complete business context of task execution, while communication metrics are collected in real time to reflect transmission status. The combined approach constructs an integrated business-communication observable system, breaking the limitations of existing technologies that separate business and communication monitoring. This allows fault location to trace both business logic problems and troubleshoot communication-level issues. The objective function defines the comprehensive scheduling cost from three dimensions: task timeout penalty, resource waste, and load balancing, with weighting coefficients... The algorithm can adjust the priority of resource utilization and load balancing according to system requirements. Under the premise of meeting hard constraints such as resource feasibility and deadlines, it minimizes the overall cost and realizes dynamic optimization of system resources, ensuring that the system continues to operate efficiently and stably in complex dynamic environments.
[0129] AI Bus intelligent agent interconnection bus inference process, such as Figure 3 As shown:
[0130] Initiating a task request: Users initiate search queries for specific business issues to the scheduling agent, such as "Accurately planning R&D paths based on past project data and market trends" or "Can the current production plan adjustment mechanism respond promptly to changes in market demand?"
[0131] Multi-source content retrieval: The scheduling agent passes task requests to the RPA process / rule engine, which retrieves business data, rule information, and knowledge nodes that are strongly related to the task from multiple dimensions of resources such as the enterprise data source pool and the local industry knowledge graph.
[0132] Related content aggregation: The RPA process / rule engine integrates the retrieved multi-source related content and feeds it back to the scheduling agent to form the data foundation for task analysis.
[0133] Large Language Model Inference Input: The scheduling agent synchronously inputs the task request and the aggregated relevant content into the large language model, providing it with a complete reasoning context.
[0134] Intelligent reasoning generates answers: Based on task requirements and multi-source data, the large language model performs deep semantic understanding, task logic decomposition, and solution derivation to generate complete answers with business value.
[0135] Answer Feedback and Application: The large language model feeds back the generated answers to the scheduling agent, which then outputs the final task solution to the user, supporting business decisions or process optimization.
[0136] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for implementing an AI Bus intelligent agent interconnection bus, characterized in that, Includes the following steps: S1. Establish the core communication hub of AI Bus, deploy MCP Server, scheduling agent, MCP Host, Agent Cards, RPA process engine, and rule engine, configure A2A adaptation protocol and Sidecar service mesh proxy, and complete the communication link between various components; the AI Bus provides a unified communication and collaboration mechanism, supporting high-quality dataset fusion agents, AI tools, rule engine, and RPA process engine to interact through standardized protocols; the scheduling agent, as the core component of MCP Server, is responsible for service dialogue requests and scheduling between agents; the MCP Host is the running carrier of agents and tools, and Agents 1-n and AI Tools access AI Bus through the MCP Host; S2. New intelligent agents or AI tools initiate a registration request through the MCP Host, which is then transmitted to the MCP Server via the AI Bus to complete standardized registration and access verification, and to include qualified intelligent agents or tools into the resource pool. S3. The scheduling agent receives user service requests, identifies task requirements and expected output formats based on a large language model, and associates Agent Cards to initially screen the range of matching agents. S4. The scheduling agent uses AI Bus to schedule the RPA process engine, rule engine and related agents to retrieve related data and rule information from multi-dimensional data sources and aggregate them to form a high-quality dataset. S5. The scheduling agent inputs the task request and high-quality dataset into the large language model, generates a task list based on the agent's ability description and decomposes it into sub-tasks, and combines Agent Cards with real-time resource status to match and adapt the agent / tool sequence. S6. The scheduling agent sends instructions to the target MCP Host through the AI Bus, specifying the agent / tool or RPA process template to execute subtasks. The Sidecar agent monitors the communication status and performs circuit breaker retry. S7: The scheduling agent monitors resource load and task progress in real time. When resource changes, timeout risks, or failures occur, it triggers task reallocation and backup resource switching. S8. Each execution unit feeds back the results through the AI Bus, and after the scheduling agent verifies and aggregates them, they are transmitted to the large language model to generate answers, or to the reporting agent to generate reports, or to the dialogue agent to output.
2. The AI Bus intelligent agent interconnection bus implementation method according to claim 1, characterized in that, When building the AI Bus core communication hub in S1, a bus architecture is adopted and a standardized interaction module with A2A adaptation protocol is integrated. The unified communication protocol specifications for CPU, memory, and network resources are preset. The Sidecar service mesh agent is deployed separately for each MCP Host and is bound to traffic monitoring, data encryption, and circuit breaker retry functions to achieve the separation of communication logic and business logic and ensure the real-time and secure transmission of cross-module data and instructions.
3. The AI Bus intelligent agent interconnection bus implementation method according to claim 2, characterized in that, The initialization of the core component deployment and scheduling agent in S1 also includes the following specific configurations: The deployed MCP Server is configured with full lifecycle management functions for intelligent agents, builds a dual-dimensional intelligent agent registration directory for business and functional domains, supports access control and security policy configuration, and schedules intelligent agents to establish communication links with large language models, Agent Cards, RPA process engines, and rule engines. Deployed Agent Cards can uniformly identify and encapsulate the capabilities of existing intelligent applications with different business capabilities. During system integration and business orchestration, the standardized capabilities carried by Agent Cards can be reused directly without repeated development, effectively improving integration efficiency and reducing development costs. The deployed RPA process engine pre-registers data retrieval and task execution process templates. Each template corresponds to an independent process service interface and completes MCP Server registration. The rule engine configures the agent to call the rule base to support the large language model's learning and inference calling logic. During the initialization of the scheduling agent, task priority weight rules and resource load thresholds are imported. And fault self-healing strategies, clearly defining dynamic task sets. Available resource set Define the core variables and complete the parameter configuration of the constraints and objective function of the autonomous scheduling optimization algorithm.
4. The AI Bus intelligent agent interconnection bus implementation method according to claim 1, characterized in that, The standardized registration and resource pool inclusion process for intelligent agents or tools in S2 specifically includes: After a new intelligent agent or tool is deployed, it initiates a registration request to the AI Bus through its MCP Host. The request carries the core information of the intelligent agent or tool, including a unique code, service interface address, input / output parameter format, capability tags, and calling rules. AI Bus performs communication compatibility verification on registration requests based on preset protocol specifications, completes preliminary traffic routing configuration through the Sidecar proxy, and after the verification is passed, the registration request is passed to the MCP Server. MCP Server updates available resource set The agent or tool is included in the system resource pool, and the agent is scheduled to perceive resource pool updates in real time through AI Bus, so that it can obtain the capability attributes of the newly added resources without modifying the existing component configuration.
5. The AI Bus intelligent agent interconnection bus implementation method according to claim 1, characterized in that, The specific process of multi-source resource retrieval and data foundation construction in S4 is as follows: The scheduling agent sends retrieval commands to the RPA process engine, rule engine, and related data analysis agents via AI Bus, specifying the retrieval dimensions and data requirements. The RPA process engine and rule engine work together to retrieve business data, rule information, and knowledge nodes that are strongly related to the task from multi-dimensional data sources such as the enterprise data source pool, local industry knowledge graph, and mechanism model library. All retrieval results are aggregated to the scheduling agent via AI Bus, and after deduplication, correlation, and formatting, they are aggregated to form the data foundation for task analysis.
6. The AI Bus intelligent agent interconnection bus implementation method according to claim 1, characterized in that, The specific process of task decomposition, constraint parameters, and resource status processing in S5 is as follows: Task decomposition: The scheduling agent synchronously inputs task requests and high-quality datasets into the large language model. Based on the agent capability descriptions in AgentCards, it infers and generates a task list, decomposing it into a sequence of executable sub-tasks, and defining a dynamic task set. ; The constraints are clearly defined: each subtask is specified. Deadline Priority weight and resource demand vector ; Resource status collection: Real-time collection of various resources The state parameters are used to generate a resource availability vector. Calculate the normalized load rate of computing resources And satisfy Load safety constraints; Task priority ranking: based on formula Calculate the urgency of the task, where The execution order is obtained by sorting the baseline processing time when resources are idle, arranged in ascending order. .
7. The AI Bus intelligent agent interconnection bus implementation method according to claim 6, characterized in that, The specific process of resource suitability calculation and scheduling decision generation in S5 is as follows: Resource suitability calculation: using the formula Calculate the fit between subtasks and resources; the higher the value, the better the fit. Scheduling decision generation: Generating allocation decision variables based on autonomous scheduling optimization algorithms And it satisfies the following constraints: Resource feasibility constraints: , ; Task uniqueness constraint: , ; Deadline constraints: , ; Resource conflict adjustment: If resource conflicts exist, adjust them using the formula. Adjustments were made, including This represents the conflict penalty coefficient.
8. The AI Bus intelligent agent interconnection bus implementation method according to claim 1, characterized in that, The specific process of dynamic scheduling and fault self-healing in S7 is as follows: Resource status and task time updates: by perception cycle Through formula Dynamic decay of occupied resources; through formula Real-time correction of estimated task processing time, among which This is the baseline processing time when resources are idle; Dynamic scheduling is triggered when any of the following conditions are met: New mission arrives: ; Resource state mutation: ,in ; Task timeout risk: ,in Second; Fault self-healing adjustment: If resources Sudden failure ( ), and assign tasks Reassign to the standby resource with the highest fit and that meets the deadline constraint. ,Right now and .
9. The AI Bus intelligent agent interconnection bus implementation method according to claim 1, characterized in that, The specific process of result aggregation and output in S8 is as follows: Results Feedback and Verification: After each execution unit completes its subtask, it feeds back the execution results, including intermediate data and analysis conclusions, to the scheduling agent through the AI Bus. Answer generation and output: If the task requires direct answer output, the scheduling agent passes the result set to the large language model, which combines the task context to generate a complete answer with business value; if the task requires report output, the scheduling agent pushes the result set to the reporting agent to generate a standardized report, and finally outputs the corresponding results to the user through the dialogue agent.
10. The AI Bus intelligent agent interconnection bus implementation method according to claim 1, characterized in that, It also includes the S9 and operation and maintenance monitoring phases, and achieves dynamic optimization through an autonomous scheduling optimization algorithm. The specific process is as follows: Full-link trajectory recording: The scheduling agent records the complete trajectory of a task from request reception, task decomposition, agent invocation, result aggregation to output. The trajectory includes task decomposition logic, agent invocation order, response data, execution status, and allocation decision variables. The changes; Communication metrics collection: The Sidecar agent collects communication metrics in real time, including traffic, latency, and error rate. After collection, the metrics are synchronized to the monitoring dashboard to visualize the communication status. Integrated management and fault location: By combining the business data recorded by the scheduling agent with the communication indicators collected by the sidecar, an integrated observable system for business and communication is constructed. When task execution abnormalities or communication failures occur, the fault nodes can be quickly located based on the trajectory records and indicator data. The fault nodes include agent capability mismatch, network fluctuations, and resource overload, supporting rapid fault diagnosis and system optimization. Dynamic optimization is achieved through an autonomous scheduling optimization algorithm. The objective function of the algorithm is: ,in The weighting coefficients are used to minimize the overall scheduling cost while satisfying hard constraints, ensuring efficient and stable system operation.