Enterprise agent low-code arrangement and version management platform and method

Abstract

The invention provides an enterprise agent low-code arrangement and version management platform and method, and the method comprises the steps: implanting an agent into each node, obtaining registration information through an agent management system, generating or associating agent scheduling process nodes, and forming an agent scheduling process node set; configuring a transmission instruction ID, downloading the instruction ID block by the agent management system node, and arranging an agent workflow for the instruction ID; the instruction ID is triggered, task order information is uploaded to an alliance chain task order block, the agent management system downloads block data, agent scheduling process nodes and order types are obtained, and available agent sets are screened; generating a task allocation block by adopting an improved adaptive genetic algorithm; and downloading the task distribution block, acquiring the interface data by the corresponding agent to execute the task, and uploading a result to the task result block. According to the invention, the problems of efficient agent scheduling management and effective application of cloud and edge end resources in a complex business scene of an enterprise are solved.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an enterprise intelligent agent low-code orchestration and version management platform. Background Technology

[0002] Currently, in the wave of enterprise digital transformation, intelligent agent technology, as a crucial pillar driving business automation and intelligent upgrades, has demonstrated irreplaceable value. Although intelligent agent technology has made some progress in enterprise applications, many solutions often struggle to adapt flexibly to dynamic and ever-changing business environments, particularly in task allocation and resource coordination. Existing methods typically cannot effectively address the uneven distribution of resources across different nodes, nor can they rationally split and schedule tasks based on their complexity. This limitation often leads to resource waste or task delays during peak periods or in complex scenarios, impacting the smoothness of overall business operations. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides an enterprise intelligent agent low-code orchestration and version management platform and method, which can achieve comprehensive improvements in production efficiency, resource utilization, product quality, and system flexibility.

[0004] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, a low-code orchestration and version management method for enterprise intelligent agents is proposed. The specific steps of this method are as follows: Intelligent agents are implanted into cloud nodes or edge nodes, and the registration information of the intelligent agents is uploaded to the consortium blockchain. The intelligent agent management system obtains the registration information to generate or associate intelligent agent scheduling process nodes, forming a set of intelligent agent scheduling process nodes. The application system node configures the transmission instruction ID and its order type, and uploads it to the consortium blockchain instruction ID block. The intelligent agent management system node downloads the instruction ID block and orchestrates the intelligent agent workflow based on the instruction ID of the intelligent agent scheduling process node set. The application system node triggers the instruction ID to upload task order information to the consortium blockchain task order block. The intelligent agent management system downloads the block data, parses the instruction ID, obtains the intelligent agent scheduling process node and order type, and filters the available intelligent agent set. The agent scheduling and management module adopts an improved adaptive genetic algorithm to optimize the scheduling and generate task allocation blocks based on task orders and the set of available agents. Each node downloads the task allocation block, the corresponding agent obtains the interface data to execute the task, and uploads the results to the task result block.

[0005] Among them, cloud nodes or edge nodes are implanted with intelligent agents, and the registration information of the intelligent agents is uploaded to the consortium blockchain. The intelligent agent management system obtains the registration information to generate or associate intelligent agent scheduling process nodes, forming a set of intelligent agent scheduling process nodes, and also includes: Intelligent agents are implanted into cloud nodes or edge nodes. The intelligent agent management system generates corresponding intelligent agent scheduling process nodes for the same intelligent agent implanted into each node, forming a set of intelligent agent scheduling process nodes. The intelligent agent management system dynamically evaluates the availability and scalability of intelligent agents on each node based on real-time information on remaining resources and working status of intelligent agents at each node, automatically registers or deregisters intelligent agents corresponding to each node, and completes the automatic reduction and expansion of intelligent agents on each node.

[0006] The application system nodes configure and transmit instruction IDs and their order types, and upload them to the consortium blockchain instruction ID block. The intelligent agent management system nodes download the instruction ID block and orchestrate intelligent agent workflows based on the instruction IDs using the intelligent agent scheduling process node set. The system also includes: Application system nodes are configured with transmission command IDs. Based on the scenario of the business command, the order type corresponding to the current command ID is defined, including regular orders and divertable orders. In addition, divertable orders also include those based on resource threshold diversion and those based on feature strength threshold diversion. The intelligent agent management system nodes orchestrate intelligent agent workflows for task order types corresponding to instruction IDs and upload process configuration blocks. The data blocks for consensus among all nodes are minimized, which improves data transmission efficiency and saves network resources.

[0007] Specifically, the application system node triggers the instruction ID to upload task order information to the consortium blockchain task order block. The intelligent agent management system downloads the block data, parses the instruction ID, obtains the intelligent agent scheduling process node and order type, and filters the set of available intelligent agents. The system also includes: Cloud nodes and edge nodes servers collect the latest remaining computing power, remaining memory and other remaining resource information in real time. If there are fluctuations, the latest remaining resource information is uploaded to the remaining resource information block and consensus is reached among all nodes of the consortium chain. Cloud nodes and edge node servers obtain real-time working status information of all smart agents on their respective nodes, generate smart agent state blocks, and reach consensus with each node of the consortium blockchain; The intelligent agent management system nodes obtain the latest task order blocks, the latest remaining resource information blocks, and the latest intelligent agent status blocks in real time. They parse the instruction IDs in the task order blocks to obtain the intelligent agent scheduling process nodes and order types, and filter the available intelligent agent set.

[0008] Furthermore, the intelligent agent management system nodes acquire the latest task order blocks, the latest remaining resource information blocks, and the latest intelligent agent status blocks in real time, parse the instruction IDs in the task order blocks, obtain the intelligent agent scheduling process nodes and order types, and filter the set of available intelligent agents. This also includes: When the order type is a regular order, the set of available agents is filtered. When the order type is a divertable order, before filtering the set of available agents, the data diversion processing module of the agent management system obtains the data to be processed for the task order based on the data interface corresponding to the pre-configured pre-input module of the instruction ID and the condition information corresponding to the current task order. Furthermore, based on the divertable orders, the diversion factors are determined to be resource demand thresholds or feature intensity thresholds. Adaptive windowing or feature intensity threshold analysis is applied to perform data sharding on the task order data to be processed, obtain the conditions or primary key information corresponding to the data shards, generate split task orders, associate them with the original task orders, and upload them to the task order block.

[0009] Specifically, based on the divertable orders, the diversion factor is determined to be either a resource demand threshold or a feature strength threshold. Adaptive windowing or feature strength threshold analysis is applied to perform data sharding on the task order data to be processed. The conditions or primary key information corresponding to the data shards are obtained, and the split task orders are generated, associated with the original task orders, and uploaded to the task order block. This also includes: The diversion factor for divertable orders is the resource demand threshold. Based on the time dimension or other feature dimensions, the data to be processed in the task order is sharded by an adaptive window. When the data resource demand in this window interval is equal to the threshold, the task orders in this interval and the corresponding conditions in this interval are obtained, and associated with the current task order corresponding to the instruction ID, and uploaded to the task order block.

[0010] Specifically, based on the divertable orders, the diversion factor is determined to be either a resource demand threshold or a feature strength threshold. Adaptive windowing or feature strength threshold analysis is applied to perform data sharding on the task order data to be processed. The conditions or primary key information corresponding to the data shards are obtained, and the split task orders are generated, associated with the original task orders, and uploaded to the task order block. This also includes: The diversion factor for divertable orders is the feature intensity threshold. A data feature curve is generated, and the set of intervals exceeding the threshold is extracted by comparing the baseline curve as the D_11 time. The remaining set of conditions, T1, is used as the D_12 time condition to associate with the original task order and upload it to the task order block.

[0011] Specifically, based on the divertable orders, the diversion factor is determined to be either a resource demand threshold or a feature strength threshold. Adaptive windowing or feature strength threshold analysis is applied to perform data sharding on the task order data to be processed. The conditions or primary key information corresponding to the data shards are obtained, and the split task orders are generated, associated with the original task orders, and uploaded to the task order block. This also includes: The diversion factor for divertable orders is the feature intensity threshold. By clustering and statistically analyzing the key features and derived features of task order processing data, bounding boxes are generated when the deviation from the historical center exceeds the threshold. The high-intensity feature intervals are extracted as D_21 to D_2N and associated with the original task order to be assigned to cloud nodes. The remaining low-intensity feature intervals are associated with the original task order to be assigned to edge nodes.

[0012] Furthermore, the agent scheduling and management module employs an improved adaptive genetic algorithm to optimize scheduling and generate task allocation blocks based on task orders and the set of available agents. It also includes: The task orders are encoded as chromosome length m×n to generate the initial population. The fitness function is R minus the completion time Fj, where R is k multiplied by the maximum Fj. The inter-group roulette wheel selection method is used to select two points for reverse mutation. The nonlinear adaptive operators pc and pm combine population fitness changes with elite retention to iterate until termination, output the optimal scheduling scheme and upload the task allocation block.

[0013] On the other hand, embodiments of the present invention also disclose an enterprise agent low-code orchestration and version management platform, applicable to any of the above-mentioned enterprise agent low-code orchestration and version management methods, including: The consortium blockchain management module is used to access and manage cloud nodes, various edge nodes, intelligent agent management system nodes, data acquisition nodes of various IoT devices, and various application system nodes. The agent registration module is used to automatically install, register, and upload agent registration blocks on various cloud or edge devices. The agent deregistration module is used to automatically uninstall various agents and deregister and upload agent deregistration blocks at various cloud or edge terminals; The data splitting and processing module is used for adaptive data splitting of task order data to be processed; The agent management module is used for the generation and management of agent scheduling process nodes; The intelligent agent scheduling module is used for the intelligent allocation of task orders; The instruction configuration module is used to configure and generate instructions that will be triggered by the application system in the future; The intelligent agent orchestration module manages the intelligent agent orchestration of instructions generated by the application system based on business requirements; The agent version management module is used to manage and view the orchestrated agent processes corresponding to instructions.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention addresses the challenge of achieving efficient intelligent agent scheduling and management, and effectively utilizing cloud and edge resources in complex enterprise business scenarios. By constructing a consortium blockchain underlying management module for each node, rapid information transmission and consensus among nodes are achieved, and smart contracts enable the orderly allocation of resources across nodes. Based on the intelligent agent management system, adaptive data partitioning of tasks can be implemented, distributing large tasks across multiple nodes for execution, thus improving the utilization rate of resources on each node. An improved adaptive genetic algorithm is used to optimize intelligent agent scheduling on edge nodes, quickly finding near-optimal scheduling schemes under complex constraints and multiple objectives. The invention monitors the resource usage status of each node and the frequency of intelligent agent scheduling, calculates the availability and scalability indicators of intelligent agents, performs dynamic analysis, and automatically registers or deregisters relevant intelligent agents at various stages, effectively utilizing node and intelligent agent resources and saving enterprise server resources. Attached Figure Description

[0015] Figure 1 : A flowchart of a low-code orchestration and version management platform and method for enterprise intelligent agents.

[0016] Figure 2 A genetic algorithm flowchart for an enterprise intelligent agent low-code orchestration and version management platform and method. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This embodiment of an enterprise intelligent agent low-code orchestration and version management platform and method may specifically include: First, it's important to clarify the different types of orders: Regular orders: The task execution of an order can only be performed by one agent, and one agent can only execute one order task at a time. Splittable orders: These typically involve large amounts of data that need to be distributed among various agents for task execution, or tasks with high feature strength that need to be assigned to the cloud. Multiple service nodes can simultaneously assign the same type of agent to handle the data processing, with each agent executing only one order task at a time. Splittable orders can be pre-processed by slicing data based on resource thresholds or feature strength thresholds. Task orders: As mentioned below, an order refers to an instruction order initiated by the system that requires an agent to execute business logic. Each agent's scheduling process node can be invoked by multiple task orders corresponding to multiple instruction IDs simultaneously.

[0019] Furthermore, a consortium blockchain is a blockchain network jointly managed and maintained by a limited number of pre-selected organizations with related businesses. Permissions within the network (such as recording, reading, and writing) are not open to everyone but are allocated according to rules agreed upon by the consortium members. Essentially, a consortium blockchain is a "cooperative" distributed ledger technology. Smart contracts in a consortium blockchain enable the high performance, fine-grained control, and compliance required by applications. It sacrifices some openness and complete decentralization, but the application of smart contracts achieves higher efficiency, better privacy protection, stronger controllability, and compliance.

[0020] A consortium blockchain comprises cloud nodes, edge nodes, an intelligent agent management system, data acquisition nodes for various IoT devices, and nodes from various application systems. This consortium blockchain architecture enables rapid information transmission and consensus among nodes, and utilizes blockchain smart contracts to ensure the orderly allocation of resources, avoiding scheduling conflicts between nodes and intelligent agents due to data transmission delays. The orchestration technology in this invention prioritizes the use of intelligent agents at edge nodes; for instructions or tasks that cannot be completed by edge intelligent agents, cloud intelligent agents are activated, effectively utilizing both cloud and edge resources.

[0021] Intelligent agents are embedded in cloud nodes or edge nodes, and their registration information is uploaded to the consortium blockchain. The intelligent agent management system obtains the registration information to generate or associate intelligent agent scheduling process nodes, forming a set of intelligent agent scheduling process nodes, specifically including: New intelligent agents are implanted into cloud nodes or edge nodes according to enterprise business needs. When a new intelligent agent is implanted, the node that added the intelligent agent automatically uploads the intelligent agent registration information, such as the node IP where the intelligent agent is deployed, the intelligent agent name, the type of intelligent agent, and the corresponding functional description, to the consortium blockchain block, and this information is then shared with all consortium blockchain nodes. After obtaining the intelligent agent registration information block, the intelligent agent management system generates an intelligent agent scheduling process node based on that intelligent agent name and associates it with the current intelligent agent registration information. If an intelligent agent scheduling process node with that intelligent agent name already exists, no new intelligent agent scheduling process node is generated; the current intelligent agent registration information is directly associated with the corresponding intelligent agent scheduling process node, thus forming a set of intelligent agent scheduling process nodes.

[0022] The agent management system dynamically evaluates the availability and scalability of agents on each node based on real-time acquisition of remaining resource information and agent working status information. It automatically registers or deregisters agents corresponding to each node, and automatically reduces and expands the number of agents on each node. Specifically, this includes: The agent management system statistically analyzes the usage frequency of each type of agent, the remaining resources of each node of that agent, and the usage frequency of other agents on each node of that agent to obtain the availability index of the agent at a given node, as follows: The test duration is measured in hours (S). P represents the number of times the current agent is invoked within a time unit of S hours; Set H is the set of minimum remaining resources (H1, H2, H3...Hn) of each edge node where the current agent is located in hour S; Set L is a set of L1, L2, L3...Ln representing the average number of times each edge node where the current agent is located is used by other agents in S hours; The current availability set of the agent at each node is K = (P / L) * H, where K is a set of K1, K2, K3...Kn; When any value in K1, K2, K3...Kn is less than a pre-set reduction threshold, the number Z of values ​​in set K that are less than the pre-set reduction threshold is counted. If Z equals K, the agent of one node is retained, and the agents of other nodes are automatically deregistered. Each node uploads the current agent deregistration information to the agent deregistration block, and consensus is reached among all nodes in the consortium blockchain. If Z is greater than 0 and less than K, the agents of nodes in set K that are less than the reduction threshold are automatically deregistered. Each node uploads the current agent deregistration information to the agent deregistration block, and consensus is reached among all nodes in the consortium blockchain. If Z equals 0, then the nodes of the current agent do not need to deregister the current agent.

[0023] Additionally, it includes an agent management system that analyzes the scalability of each agent, as detailed below: Y1 represents the minimum number of nodes required for scalability; Set N is the set of minimum remaining resources (N1, N2, N3...Nn) for each edge node that is not registered with the current agent in hour S; When there is a value in K1, K2, K3...Kn that is greater than the preset expansion threshold, the number Y of values ​​in set K that are greater than the preset expansion threshold is counted. If Y is less than Y1, the current agent's scalability is 0, and there is no need to register the current agent on each node. If Y is greater than or equal to Y1, the agent can be registered on Y1 nodes that have not yet registered. The platform automatically selects the largest node from set N to automatically register the current agent. Each node uploads the current agent registration information to the agent deregistration block and reaches consensus with each node in the consortium blockchain.

[0024] As described above, by monitoring the resource usage status of each node and the scheduling frequency of intelligent agents, calculating the availability and scalability indicators of intelligent agents, performing dynamic analysis, and automatically registering or deregistering relevant intelligent agents at each stage, the resources of each node and intelligent agent can be effectively utilized, saving enterprise server resources.

[0025] Furthermore, the application system nodes configure the transmission instruction ID and its order type, and upload them to the consortium blockchain instruction ID block. The intelligent agent management system nodes download the instruction ID block and orchestrate the intelligent agent workflow based on the intelligent agent scheduling process node set for the instruction ID, specifically including: Based on the needs of enterprise business processes, transmission instructions are configured in the business operation process of application system nodes, and a unique instruction ID is generated for them. Based on the data generated by the business, the order type corresponding to the current instruction ID is marked as a regular order or a divertable order. If it is a divertable order, the diversion factors are also configured. The diversion factors include diversion based on resource demand threshold and diversion based on feature strength threshold. The instruction ID information is then uploaded to the consortium blockchain instruction ID block.

[0026] The intelligent agent management system node downloads the instruction ID block. Based on the enterprise's business instruction ID processing needs, the corresponding intelligent agent workflow is orchestrated on the intelligent agent management system. The instruction ID is selected, the name of the intelligent agent scheduling process node invoked by the current instruction ID order to be processed is configured, the name of the data input interface for the current business to be processed is selected as the intelligent agent's pre-input, and the name of the data output interface after data processing is selected as the intelligent agent's post-output. The intelligent agent management system node uploads this configuration information and configuration time information to the process configuration block and reaches consensus with all nodes in the consortium blockchain. Before or after processing the data corresponding to a specific task order, the data block for consensus among all nodes is minimized to improve data transmission efficiency and save network resources.

[0027] As mentioned above, based on business requirements, the corresponding intelligent agent call configuration is orchestrated for each instruction ID in the intelligent agent management system node. For example, when calling the process block corresponding to the instruction ID, the system automatically calls the instruction ID process configuration block information with the latest block generation time.

[0028] Furthermore, the application system node triggers the instruction ID to upload task order information to the consortium blockchain task order block. The intelligent agent management system downloads the block data, parses the instruction ID, obtains the intelligent agent scheduling process node and order type, and filters the set of available intelligent agents. Specifically, this includes: When an application system node issues a trigger command at a certain business stage, such as trigger command ID1, the application system node uploads command ID1 and determines the task order information corresponding to the current command ID1. If the current order is a regular order, it uploads the current command ID1 and the key information corresponding to the current task order, such as the primary key of the data to be processed by the intelligent agent, or the corresponding time range, etc. This key information serves as the condition for obtaining interface data for the data input interface of the aforementioned business to be processed. If the current order is a divisible order, in addition to uploading the key information corresponding to the task order, it also needs to upload the capacity of the data corresponding to this task order, and then calculate the computing power or resource requirements required by the intelligent agent to process this task order. The application system node uploads the above information to the task order block and reaches consensus with all nodes of the consortium blockchain.

[0029] As mentioned above, based on the block data uploaded with the instruction ID, the intelligent agent management system node parses the key features of the data to be processed by the intelligent agent rather than the data itself, thereby reducing the frequency of large-scale data transmission, improving efficiency, and reducing the waste of ineffective resources.

[0030] Additionally, it should be noted that cloud nodes and edge node servers continuously monitor the latest remaining computing power, remaining memory, and other remaining resource information. If there are fluctuations, the latest remaining resource information will be uploaded to the remaining resource information block and shared with all nodes of the consortium blockchain; if there are no fluctuations, no action will be taken.

[0031] It should also be noted that when the intelligent agents of cloud nodes and edge nodes start or stop executing tasks, they upload the latest status information of the corresponding intelligent agents to the intelligent agent status block, including the IP of the cloud node or edge node to which they belong, the name of the intelligent agent, and the working status information of the intelligent agent when starting or stopping execution, and then reach consensus with each node of the consortium blockchain.

[0032] Furthermore, the intelligent agent management system nodes acquire the latest task order blocks, the latest remaining resource information blocks, and the latest intelligent agent status blocks in real time. They parse the instruction IDs in the task order blocks to obtain the intelligent agent scheduling process nodes and order types, and filter the set of available intelligent agents. Specifically, this includes: The intelligent agent management system nodes download the latest task order blocks, the latest remaining resource information blocks, and the latest intelligent agent status blocks. They perform real-time parsing of the downloaded task order block data to obtain the corresponding instruction ID1 and the information corresponding to the current task order. Based on the instruction ID1, they retrieve the pre-configured intelligent agent scheduling process node corresponding to instruction ID1 and the order type marked by instruction ID1. According to the cloud node or edge node to which the intelligent agent belongs, associated with the intelligent agent's scheduling process node, they obtain the latest remaining resource information and the working status of the corresponding intelligent agent on the cloud node and edge node. Intelligent agents in the working state of "started execution" are marked as 1, intelligent agents in the working state of "finished execution" are marked as 0, and intelligent agents whose remaining resources are less than a pre-set threshold are marked as 1. Let this be set M1, which includes the corresponding cloud node IP or edge node IP and the intelligent agent name. If set M1 is empty, instruction ID1 waits for allocation; if set M1 is not empty, it selects a set of intelligent agents that meet the requirements of the current task order, and uses the intelligent agent management system to adaptively split the task data, distributing a large task to multiple nodes for execution, improving the utilization rate of resources on each node.

[0033] If the order type corresponding to the current instruction ID is a regular order, and if the agent set M1 contains edge node agents, then all edge node agents in set M1 will be selected as candidate set M. If the agent set M1 does not contain edge node agents, then the cloud node agents in set M1 will be selected as candidate set M.

[0034] If the order type corresponding to the current instruction ID is a divertable order, the data diversion processing module of the intelligent agent management system obtains the condition information of the task order corresponding to the instruction ID and the data interface corresponding to the pre-configured access front-end input, obtains the data to be processed for the task order based on this condition, and at the same time, according to the diversion factor information corresponding to the current instruction ID, firstly, adaptive window slicing is used to divert the data of the task order corresponding to the current instruction ID, specifically including: If the current instruction ID's distribution factor is a resource requirement threshold, information from the time dimension or other feature dimensions is obtained. For example, data segmentation can be performed using a self-adaptive time window. For instance, the start time of the time condition in the task order block information of the current instruction ID can be used as the start time of the time interval. This time interval is then progressively shifted forward, and the data capacity and processing resource requirements of this time interval are statistically analyzed in real time. If the resource requirement equals a pre-set threshold, the shifting and expansion of the time interval stops, and this time interval is used as the time condition for the first task order D1 split from the task order corresponding to the first instruction ID. The first task order D_01 is then associated with the current task order corresponding to the instruction ID. The end time node in the instruction ID's time condition is used as the start time node to continue shifting forward. Similarly, the split task orders D_02, D_03...D_0N are obtained, thus completing the splitting and distribution of the task order for the current instruction ID. The split task orders and their corresponding time condition information are then uploaded to the task order block and shared with all nodes of the consortium blockchain.

[0035] If the triage factor for the current instruction ID is a feature strength threshold, the techniques for triaging task orders based on the current instruction ID include, but are not limited to: The data distribution and processing module of the intelligent agent management system generates curves of data features or derived features corresponding to the task order with the current instruction ID based on the time dimension, and generates a baseline curve based on historical data. For example, based on time-dimensional data segmentation, the starting time point is when the value of the data feature or derived feature corresponding to the task order with the current instruction ID exceeds a pre-set threshold, and the ending time point is when the value of the data feature or derived feature corresponding to the task order with the current instruction ID returns to the baseline. This time interval is extracted and added to the set T0 of time intervals where the feature intensity exceeds the threshold. Then, all time intervals where the value of the data feature or derived feature corresponding to the task order with the current instruction ID exceeds the threshold are obtained and added to T0. T0 serves as the time condition for task order D_11, which is split from the current instruction ID, and task order D_11 is associated with the current task order corresponding to the instruction ID. The remaining time intervals corresponding to the instruction ID that are not added to T0 are set as T1. T0 serves as the time condition for task order D_12, which is split from the current instruction ID, and task order D_11 is associated with the current task order corresponding to the instruction ID. As described above, the task order information corresponding to the current instruction ID is processed by data slicing to generate task order D_11 after splitting the task order corresponding to the instruction ID and its corresponding time condition T0, as well as task order D_12 and its corresponding time condition T1. A new task order block associated with the current task order corresponding to the instruction ID is generated and consensus is reached among all nodes of the consortium blockchain.

[0036] Additionally, the data splitting and processing module of the intelligent agent management system acquires regions or data features with abnormal feature strength. Data splitting and processing involves clustering the key features of the task order corresponding to the current instruction ID with the key features of historical data or model training data. It then determines whether the features in the current task order deviate from the center of the set of key features in historical data or model training data by more than a threshold. For example, in image recognition task orders, if this feature value is the center, a bounding box is generated. The pixels within this bounding box are taken as order D_21, which is then associated with the current task order corresponding to the instruction ID. This further splits into task orders D_22, D_23...D_2N, which are also associated with the current task order corresponding to the instruction ID. These task orders with feature strengths greater than the threshold (D_21, D_22, D_23...D_2N) are then assigned to intelligent agents on cloud nodes for queuing processing. If the task order corresponding to the current instruction ID does not have any regions or features with abnormal feature strength, it is designated as task order D_20, associated with the current task order corresponding to the instruction ID, and assigned to intelligent agents on edge nodes for processing. As mentioned above, the intelligent agent management system uploads the split order information and feature data, along with the allocation information to edge nodes or cloud nodes, to the task order block and reaches consensus with each node of the consortium blockchain.

[0037] Furthermore, the cloud-based node agents invoke large-scale pre-trained deep learning models to perform detailed analysis of the local high-dimensional data uploaded to the cloud nodes. The cloud-based models possess stronger feature representation and pattern recognition capabilities, enabling them to identify subtle and complex defects that edge models struggle to distinguish. The cloud-based models not only provide the final classification but also output a measure of prediction uncertainty. For samples with high uncertainty, they can be automatically pushed to a human expert interface for review, and the review results are used as new labeled data. Based on the difficult samples and expert annotations accumulated in the cloud, a more powerful next-generation detection model is periodically trained and generated through a federated learning framework.

[0038] It should be noted that the original task orders that have been split will no longer participate in the agent allocation and execution of the current process scheduling node. When all the split task orders associated with the original task order complete their task execution based on this process node, then the original task order will have completed its task execution based on this process node.

[0039] Furthermore, based on the aforementioned instruction task orders and the split instruction task orders, as well as the current status of the agent, the agent scheduling and management module uses an improved adaptive genetic algorithm to optimize agent scheduling for edge node agents, quickly finding a near-optimal scheduling scheme under complex constraints and multiple objectives.

[0040] The following mathematical symbols are explained as follows: M: node agent; T: execution time; Sch: scheduling plan; Optsch: shortest execution time scheme; ST: start execution time; ET: end execution time; Jobid: task order number; Procid: agent scheduling process node number; Fitness: fitness function; k: proportional coefficient; R: normalized intermediate value.

[0041] During business execution, let M = {1, 2, 3, ..., m} be the node agents, N = {1, 2, 3, ..., n} be the orders, and Sch = N = {1, 2, 3, ..., m × n} be the set of all process nodes. The order of constraints between processes is as follows: for each order j, the order can only be processed when the previous step is completed and the node of the next process i is idle. Let j be the start time. Let j be the execution time. Let j be the completion time. The order execution time varies depending on the agent selection scheme, and the optimal agent selection and execution sequence scheme directly constitute the optimal scheduling scheme.

[0042] The business execution agent scheduling problem generally needs to satisfy the following conditions: the execution order of all orders on the agents is given in advance, and orders should be executed on the agents in the predetermined order; at any given time, the execution of a process node of an order can only be performed on one agent, and one agent can only execute the process node business of one order; all orders have the same execution priority, and an order can only be completed after all its execution tasks have been completed; except in special cases, the execution time of orders is determined and remains unchanged. Order j is executed on agent m, denoted as... =1, otherwise record as =0. The specific mathematical model is: (1) (2) (3) (4) (5) Equation (1) represents the Optsch scheme with the shortest completion time; Equation (2) indicates that each order j must be executed continuously once it starts and the task cannot be terminated prematurely; Equation (3) indicates that each agent can execute at most one order at the same time; the constraints Equations (4) and (5) ensure the accuracy and rationality of the processing.

[0043] In the business intelligence agent scheduling problem, the order execution order can be regarded as the chromosome arrangement in genetic behavior. The algorithm mainly changes the order execution order by iterating the chromosome population, thereby completing the optimization task.

[0044] The algorithm design includes the following steps: encoding and decoding, generating the initial population, chromosome selection, crossover and mutation, and algorithm termination, as detailed below: (1) Chromosome encoding and decoding. Since the dimensions of the genetic algorithm solution and the business agent scheduling solution are different, the scheduling scheme can be mapped to chromosomes through encoding. The number of agent nodes is m and the number of orders is n. Therefore, the length of the encoding is m×n. The encoding is randomly arranged by the randperm function as an initial feasible solution. Pop the initial solutions to form a whole and constitute the initial population.

[0045] (2) Fitness Function Design. In the business intelligence agent execution scheduling problem, the ultimate goal is to minimize the completion time. The fitness function is selected by recording the code of each chromosome and the completion time of chromosome i. Select the maximum completion time (max) This is scaled up to a certain ratio (designed to be 1.2–1.5 times), with the ratio being k. Then, an intermediate reference value R is obtained, which is subtracted from the completion time of each chromosome. The larger the difference, the higher the fitness value of that chromosome, i.e., the shorter the completion time. Therefore, the fitness function is designed as follows: (6) (7) (3) Generate the initial population. A roulette wheel strategy is adopted between groups to make individuals with higher fitness values ​​more likely to be selected as the parent of new individuals, thereby guiding the optimization direction of the algorithm.

[0046] (4) Chromosome selection. Chromosome selection is performed using a normalized probability roulette wheel. Random numbers are generated, and the interval in which they fall determines which chromosome to replicate. From the existing pop populations, pop chromosomes are selected again using a roulette wheel to form a new population.

[0047] (5) Genetic operations. The crossover operation uses a two-point crossover method to randomly generate two points. , This results in two paternal chromosomes. , exist , The exchange of gene segments between the two points forms two new chromosomes. and and to and The algorithm performs validity processing by comparing the two exchanged segments, retaining the code that cannot be canceled out and replacing it sequentially. The mutation method is reverse mutation, which randomly generates two points on the chromosome, and reverses the chromosome segments between the two points to achieve the mutation effect. The algorithm terminates when the maximum number of iterations is reached.

[0048] (6) This invention employs a novel, nonlinearly varying adaptive crossover and mutation operator improvement strategy. The operator changes according to the fitness value of the population, maintaining diversity in the early stages of the population while preventing the destruction of superior individuals in the later stages. The crossover operator (pc) and mutation operator (pm) are designed as follows: (8) Where Pc is the crossover probability of the current individual; Pc1 is the maximum crossover probability; Pc2 is the minimum crossover probability; f is the fitness value of the current individual; favg is the average fitness value of all individuals in the current population; fmax is the maximum fitness value in the current population; and β is the adaptive adjustment factor.

[0049] (9) Where Pm is the mutation probability of the current individual; Pm1 is the maximum mutation probability; Pm2 is the minimum mutation probability; f, favg, fmax, and β have the same meaning as in formula (8).

[0050] Explanation of the function of the parameters: (1) Pc1, Pc2, Pm1, Pm2: control the upper and lower bounds of crossover and mutation probabilities to prevent the probabilities from being too high or too low.

[0051] (2) F, favg, fmax: Reflect the degree of superiority or inferiority of an individual in the population, used to dynamically adjust probabilities: If the individual fitness f is close to fmax (good individual), then β approaches 0, Pc and Pm decrease, thus protecting the good solution; If an individual's fitness f is low (a poor individual), then β approaches 1, Pc and Pm increase, enhancing search ability.

[0052] (3) Used to determine the convergence state of a population: If the value is ≤0.5, it indicates that the population has not yet converged, and adaptive adjustment should be used; If the value is greater than 0.5, it indicates that the population has converged. In this case, the maximum probability Pc1 or Pm1 should be used to prevent the population from getting trapped in local optima.

[0053] In equations (8) and (9), regardless of the current mutation probability, the improved crossover mutation rate is not equal to 0, and in the early stages of evolution, it is unlikely to fall into a local optimum. As the population iterates, the crossover mutation probability gradually increases, and accelerating the search speed can find the optimal solution as soon as possible. When the current individual is close to the optimal individual, the crossover mutation probability gradually decreases, avoiding the current individual from falling into a local optimum. This operator can not only retain the excellent coding scheme in the iteration of solving the business agent scheduling problem, but also escape local optima, and has a good effect in solving the business agent scheduling problem. At the same time, this invention applies the elite preservation strategy to the algorithm, that is, retains the best chromosome in this generation to prevent the optimal individual from being lost or destroyed due to crossover mutation, thereby accelerating the evolution process. The application process of the improved adaptive genetic algorithm is as follows: Figure 2 As shown, the optimal scheduling scheme is finally obtained, and then the allocation scheme of real-time task instructions is determined.

[0054] The intelligent agent management system nodes upload the final task order allocation data to the task allocation block and reach consensus with all consortium blockchain nodes. Each cloud node or edge node downloads the task allocation block data, parses the corresponding task allocation data, and if the current task allocation data contains information about a corresponding intelligent agent task assigned to the current node, the corresponding task order enters the current intelligent agent's execution queue, waiting for the previous task order to complete. When the current task order enters execution, the intelligent agent obtains the task order block information corresponding to the current task order, and based on the task order block information, obtains the interface condition information for calling the corresponding pending data, including but not limited to the time conditions and feature data primary key information mentioned above. Then, based on this information, it calls the corresponding interface data of the pre-configured data input interface, executes the task, and outputs the execution result data to the pre-configured data output interface. The execution result, along with the task instruction ID, is uploaded to the task instruction result block and reached consensus with all consortium blockchain nodes. The corresponding nodes then download the task instruction result block to proceed to the next business task.

[0055] This invention uses a genetic algorithm to allocate the intelligent agents that can execute the currently parallel task instruction orders in real time, avoiding queuing and waiting at an edge node, distributing resource requests across multiple nodes, effectively utilizing the resources of each server's intelligent agents, and improving business execution efficiency.

[0056] Furthermore, each node downloads the task allocation block, the corresponding agent obtains interface data to execute the task, and uploads the results to the task result block, specifically including: Each node downloads and parses the task allocation block. If the task is allocated to the current node, the task order enters the execution queue of the corresponding agent of the current node.

[0057] The current node's execution agent obtains the condition information and primary key information from the task order block, obtains the data to be processed in the task order based on the pre-configured data input interface corresponding to the instruction ID of the task order, executes the task, outputs the result data of the task execution to the pre-configured data output interface of the instruction ID, and periodically clears the processed data to avoid resource waste.

[0058] The current node's executing agent uploads the processing result status to the corresponding task execution result block and reaches consensus among all nodes in the consortium blockchain. If the order is not related to the original task order, a result block for the current task order is generated; if it is related, it determines that all split task orders have been completed, and then a result block for the original task order is generated.

[0059] Specifically, for task orders processed by edge node smart agents or cloud node smart agents, if the current order is not associated with an original task order with an instruction ID, the current order completes data processing based on this smart agent process. If the current order has an original task order associated with an instruction ID, it is determined whether all task orders associated with the original task order have been executed based on this smart agent process. If all have been executed, the original task order with the corresponding instruction ID generates a task instruction result block and reaches consensus with each node of the consortium blockchain. If there are incomplete task orders, the original task order is not processed.

[0060] It should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should be considered within the scope of protection of this invention.

Claims

1. A method for low-code orchestration and version management of enterprise intelligent agents, characterized in that, The method includes: Intelligent agents are implanted into cloud nodes or edge nodes, and the registration information of the intelligent agents is uploaded to the consortium blockchain. The intelligent agent management system obtains the registration information to generate or associate intelligent agent scheduling process nodes, forming a set of intelligent agent scheduling process nodes. The application system node configures the transmission instruction ID and its order type, and uploads it to the consortium blockchain instruction ID block. The intelligent agent management system node downloads the instruction ID block and orchestrates the intelligent agent workflow based on the instruction ID of the intelligent agent scheduling process node set. The application system node triggers the instruction ID to upload task order information to the consortium blockchain task order block. The intelligent agent management system downloads the block data, parses the instruction ID, obtains the intelligent agent scheduling process node and order type, and filters the available intelligent agent set. The agent scheduling and management module adopts an improved adaptive genetic algorithm to optimize the scheduling and generate task allocation blocks based on task orders and the set of available agents. Each node downloads the task allocation block, the corresponding agent obtains the interface data to execute the task, and uploads the results to the task result block.

2. The method as described in claim 1, characterized in that, The cloud nodes or edge nodes are embedded with intelligent agents, and the registration information of the intelligent agents is uploaded to the consortium blockchain. The intelligent agent management system obtains the registration information to generate or associate intelligent agent scheduling process nodes, forming a set of intelligent agent scheduling process nodes, including: Intelligent agents are implanted into cloud nodes or edge nodes. The intelligent agent management system generates corresponding intelligent agent scheduling process nodes for the same intelligent agent implanted into each node, forming a set of intelligent agent scheduling process nodes. The intelligent agent management system dynamically evaluates the availability and scalability of intelligent agents on each node based on real-time information on remaining resources and working status of intelligent agents at each node, automatically registers or deregisters intelligent agents corresponding to each node, and completes the automatic reduction and expansion of intelligent agents on each node.

3. The method as described in claim 1, characterized in that, The application system node configures the transmission instruction ID and its order type, and uploads it to the consortium blockchain instruction ID block. The intelligent agent management system node downloads the instruction ID block and orchestrates the intelligent agent workflow based on the instruction ID according to the intelligent agent scheduling process node set, including: Application system nodes are configured with transmission command IDs. Based on the scenario of the business command, the order type corresponding to the current command ID is defined, including regular orders and divertable orders. In addition, divertable orders also include those based on resource threshold diversion and those based on feature strength threshold diversion. The intelligent agent management system nodes orchestrate intelligent agent workflows for task order types corresponding to instruction IDs and upload process configuration blocks. The data blocks for consensus among all nodes are minimized, which improves data transmission efficiency and saves network resources.

4. The method as described in claim 1, characterized in that, The application system node triggers the instruction ID to upload task order information to the consortium blockchain task order block. The intelligent agent management system downloads the block data, parses the instruction ID, obtains the intelligent agent scheduling process node and order type, and filters the set of available intelligent agents, including: Cloud nodes and edge nodes servers collect the latest remaining computing power, remaining memory and other remaining resource information in real time. If there are fluctuations, the latest remaining resource information is uploaded to the remaining resource information block and consensus is reached among all nodes of the consortium chain. Cloud nodes and edge node servers obtain real-time working status information of all smart agents on their respective nodes, generate smart agent state blocks, and reach consensus with each node of the consortium blockchain; The intelligent agent management system nodes obtain the latest task order blocks, the latest remaining resource information blocks, and the latest intelligent agent status blocks in real time. They parse the instruction IDs in the task order blocks to obtain the intelligent agent scheduling process nodes and order types, and filter the available intelligent agent set.

5. The method as described in claim 4, characterized in that, The intelligent agent management system nodes acquire the latest task order blocks, the latest remaining resource information blocks, and the latest intelligent agent status blocks in real time. They parse the instruction IDs in the task order blocks to obtain the intelligent agent scheduling process nodes and order types, and filter the available intelligent agent set, including: When the order type is a regular order, the set of available agents is filtered. When the order type is a divertable order, before filtering the set of available agents, the data diversion processing module of the agent management system obtains the data to be processed for the task order based on the data interface corresponding to the pre-configured pre-input module of the instruction ID and the condition information corresponding to the current task order. Furthermore, based on the divertable orders, the diversion factors are determined to be resource demand thresholds or feature intensity thresholds. Adaptive windowing or feature intensity threshold analysis is applied to perform data sharding on the task order data to be processed, obtain the conditions or primary key information corresponding to the data shards, generate split task orders, associate them with the original task orders, and upload them to the task order block.

6. The method as described in claim 5, characterized in that, The process involves determining the diversion factor based on divertable orders, using either a resource demand threshold or a feature strength threshold. Adaptive windowing or feature strength threshold analysis is applied to perform data sharding on the task order data to be processed. The conditions or primary key information corresponding to each data shard are obtained, and the split task orders are generated, associated with the original task orders, and uploaded to the task order block. This includes: The diversion factor for divertable orders is the resource demand threshold. Based on the time dimension or other feature dimensions, the data to be processed in the task order is sharded by an adaptive window. When the data resource demand in this window interval is equal to the threshold, the task orders in this interval and the corresponding conditions in this interval are obtained, and associated with the current task order corresponding to the instruction ID, and uploaded to the task order block.

7. The method as described in claim 5, characterized in that, The process of determining the diversion factor based on divertable orders as a resource demand threshold or feature strength threshold, applying adaptive windowing or feature strength threshold analysis, performing data sharding on the task order data to be processed, obtaining the conditions or primary key information corresponding to the data shards, generating split task orders, associating them with the original task orders, and uploading them to the task order block, also includes: The diversion factor for divertable orders is the feature intensity threshold. A data feature curve is generated, and the set of intervals exceeding the threshold is extracted by comparing the baseline curve as the D_11 time. The remaining set of conditions, T1, is used as the D_12 time condition to associate with the original task order and upload it to the task order block.

8. The method as described in claim 5, characterized in that, The process of determining the diversion factor based on divertable orders as a resource demand threshold or feature strength threshold, applying adaptive windowing or feature strength threshold analysis, performing data sharding on the task order data to be processed, obtaining the conditions or primary key information corresponding to the data shards, generating split task orders, associating them with the original task orders, and uploading them to the task order block, also includes: The diversion factor for divertable orders is the feature intensity threshold. By clustering and statistically analyzing the key features and derived features of task order processing data, bounding boxes are generated when the deviation from the historical center exceeds the threshold. The high-intensity feature intervals are extracted as D_21 to D_2N and associated with the original task order to be assigned to cloud nodes. The remaining low-intensity feature intervals are associated with the original task order to be assigned to edge nodes.

9. The method as described in claim 1, characterized in that, The agent scheduling and management module employs an improved adaptive genetic algorithm to optimize scheduling and generate task allocation blocks based on task orders and the set of available agents, including: The task orders are encoded as chromosome length m×n to generate the initial population. The fitness function is R minus the completion time Fj, where R is k multiplied by the maximum Fj. The inter-group roulette wheel selection method is used to select two points for reverse mutation. The nonlinear adaptive operators pc and pm combine population fitness changes with elite retention to iterate until termination, output the optimal scheduling scheme and upload the task allocation block.

10. An enterprise intelligent agent low-code orchestration and version management platform, employing the method described in any one of claims 1 to 9, characterized in that, include: The consortium blockchain management module is used to access and manage cloud nodes, various edge nodes, intelligent agent management system nodes, data acquisition nodes of various IoT devices, and various application system nodes. The agent registration module is used to automatically install, register, and upload agent registration blocks on various cloud or edge devices. The agent deregistration module is used to automatically uninstall various agents and deregister and upload agent deregistration blocks at various cloud or edge terminals; The data splitting and processing module is used for adaptive data splitting of task order data to be processed; The agent management module is used for the generation and management of agent scheduling process nodes; The intelligent agent scheduling module is used for the intelligent allocation of task orders; The instruction configuration module is used to configure and generate instructions that will be triggered by the application system in the future; The intelligent agent orchestration module manages the intelligent agent orchestration of instructions generated by the application system based on business requirements; The agent version management module is used to manage and view the orchestrated agent processes corresponding to instructions.

Images