A method for coordinating and scheduling, executing in parallel and self-evolving of intelligent agents of a platform of a million-level industry

Abstract

The application belongs to the technical field of artificial intelligence multi-agent cooperation, and discloses a method for ten-thousand-level industry agent cooperation scheduling, parallel execution and self-evolution of a platform, wherein a ten-thousand-level agent cluster covering thousands of industries is constructed, a single master agent receives and issues a unified task instruction, the instruction is automatically decomposed into sub-tasks that can be executed in parallel, ten-thousand industry agents are scheduled to execute tasks, communicate with each other, learn from each other, and based on the task execution results, the agent model is automatically optimized and upgraded. The application solves the technical pain points of the existing multi-agent system, such as insufficient industry coverage, inability to large-scale parallel, difficulty in cross-agent knowledge sharing and inability to self-evolution, realizes the platform-level intelligent cooperation capability of single instruction triggering, ten-thousand-agent concurrency, cross-industry interconnection and lifelong iteration, greatly improves the processing efficiency and intelligence level of complex business scenarios, and is suitable for large-scale operation and management scenarios of digitalization and intelligentization of thousands of industries.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence multi-agent collaborative technology, specifically involving a method for collaborative scheduling, parallel execution, and self-evolution of tens of thousands of industry-level intelligent agents on a platform. Background Technology

[0002] Artificial intelligence technology is rapidly penetrating all industries, and single intelligent agents or small-scale intelligent agent collaboration are no longer sufficient to handle complex, diverse, and large-scale business processing needs. Existing multi-agent systems mainly suffer from the following shortcomings: First, intelligent agents have limited coverage, usually only developed for a single industry or specific functional scenarios, lacking the ability to cover the entire industry and unable to solve cross-industry collaboration problems.

[0003] Second, the task execution mode is outdated. Existing systems mostly use serial execution or small-scale parallel execution, which cannot support tens of thousands of intelligent agents to execute online at the same time, resulting in low task processing efficiency and failing to meet the needs of high-frequency and high-volume business.

[0004] Third, the lack of an effective interconnection mechanism between intelligent agents means that each agent is in an information silo state, unable to learn from each other and share experiences, which prevents the agents from rapidly improving their capabilities.

[0005] Fourth, the lack of a self-evolution mechanism means that the intelligent agent models rely heavily on offline training and become fixed, and cannot iterate and optimize themselves through feedback data during online task execution. As a result, the system's capabilities stagnate after long-term operation.

[0006] Therefore, in order to meet the full-ecosystem operation needs of the Dianji platform, there is an urgent need for a technical solution that can achieve large-scale concurrency of tens of thousands of intelligent agents, cross-industry interconnection and mutual learning, and support lifelong self-iteration. Summary of the Invention

[0007] To overcome the aforementioned technical problems, this invention provides a method for collaborative scheduling, parallel execution, and self-evolution of tens of thousands of industry-specific intelligent agents on a platform. The aim is to build a super intelligent agent ecosystem covering thousands of industries, enabling tens of thousands of intelligent agents to execute tasks efficiently at the same time through a single instruction entry point, promoting mutual learning and co-evolution among intelligent agents, thereby significantly improving the platform's ability and efficiency in handling complex tasks.

[0008] The present invention adopts the following technical solution: A method for collaborative scheduling, parallel execution, and self-evolution of tens of thousands of industry intelligent agents on a platform includes the following steps: S1. Cluster Initialization: Build a cluster of tens of thousands of intelligent agents covering thousands of industries on the Dianjita platform. Each intelligent agent has a unique industry business capability tag, industry knowledge base, communication interface and initial iteration model. S2, Command Access: The unified task command input from the outside is received by the unique central control intelligent agent within the platform, and the task command is subjected to intent recognition, target extraction and constraint condition parsing to generate a global task planning graph; S3. Task decomposition: The overall control agent automatically decomposes the complex global task into multiple independently executable parallel subtasks according to the global task planning graph, and constructs the dependency relationship and execution priority between the subtasks. S4. Parallel Scheduling and Execution: The central control agent selects a suitable target agent group from the tens of thousands of agents clusters based on the industry attributes corresponding to each sub-task and the current load status of the agent, issues corresponding sub-task instructions, and schedules the target agent group to start and execute the sub-tasks in parallel. S5. Cross-Agent Interconnection Learning: During task execution, the target intelligent agent group shares execution status, intermediate results, industry experience and optimal strategies in real time through the communication protocol built into the platform, realizing cross-industry and cross-scenario mutual learning and knowledge integration. S6. Result Aggregation: The overall control agent automatically aggregates the subtask execution results of all target agents, performs data verification, deduplication, fusion and sorting, and generates and outputs the global final task results. S7. Self-iterative optimization: Based on the performance evaluation data of this task, a multi-dimensional reward function is constructed. The model parameters of the overall control agent and the participating industry agents are updated through reinforcement learning algorithms to realize the autonomous iteration and capability evolution of the agents.

[0009] Preferably, the number of the target intelligent agents group to be adapted in step S4 ranges from 1,000 to 50,000, supports dynamic intelligent expansion according to task complexity, and all participating intelligent agents adopt a synchronous execution mechanism without serial waiting.

[0010] Preferably, the cross-agent interconnection learning in step S5 specifically includes: converting heterogeneous data from different industries into unified feature vectors through a semantic alignment protocol, sharing best practice cases in the execution process, optimizing its own industry rule base and execution strategy, and realizing the transfer and complementarity of capabilities.

[0011] Preferably, the multi-dimensional reward function in step S7 includes task completion rate, execution efficiency, result accuracy, cross-agent collaborative score, and resource utilization rate. A global reward signal is generated through weighted calculation to drive the agent to update the model.

[0012] This invention also discloses a collaborative scheduling, parallel execution, and self-evolution system for tens of thousands of industry intelligent agents on a platform, used to implement the above-mentioned method, including: The 10,000-level industry intelligent agent cluster module is used to deploy and manage tens of thousands of distributed intelligent agents covering thousands of industries; The central control intelligent agent scheduling module, as the system's sole entry point, is responsible for instruction parsing, task decomposition, and global scheduling. The parallel task execution module is used to schedule the selected group of agents to execute subtasks in parallel. The cross-agent interconnection learning module is used to enable real-time communication, experience sharing, and knowledge transfer between agents; The self-iterative optimization module is used to update the agent model through reinforcement learning based on execution feedback data. The results aggregation and output module is used to summarize, verify, and output the global task results.

[0013] Preferably, the tens of thousands of industry intelligent agent cluster modules are deployed using a distributed containerized architecture, which supports the dynamic creation, destruction, and elastic expansion of intelligent agents, ensuring low-latency communication and stable operation of large-scale intelligent agents.

[0014] Preferably, the overall control intelligent agent scheduling module integrates an intent recognition engine and a dynamic programming algorithm, which can adjust the task decomposition strategy and resource allocation scheme according to real-time environmental data.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Achieves simultaneous execution of tens of thousands of intelligent agents, doubling efficiency: This invention is the first to support tens of thousands of intelligent agents to execute tasks in parallel, shortening complex tasks that originally required months or days to be processed to hours or even minutes, thus improving processing efficiency.

[0016] 2. Break down industry barriers and build an ecological closed loop: Through the interconnected learning of intelligent agents from various industries, the problem of the capability ceiling of a single intelligent agent has been solved, and the cross-empowerment of business capabilities and ecological integration have been realized.

[0017] 3. Possesses self-evolution capability and zero-cost iteration: Unlike the traditional "train once, use forever" model, this invention continuously optimizes the model through online task data, realizing the autonomous evolution of the agent and reducing the cost of continuous operation and training.

[0018] 4. High system stability and strong scalability: It adopts a distributed containerized architecture, supports the dynamic addition and removal of intelligent agents, and a single point of failure does not affect the whole system, which can easily support the unlimited expansion of business scale. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to specific embodiments.

[0021] Example 1: Operation of an Omnichannel Spring Promotion Campaign 1. Cluster Configuration The Dianji platform has completed the deployment of 30,000 industry-specific intelligent agent nodes, covering intelligent agents in all consumer sectors, including retail, catering, beauty, apparel, home appliances, FMCG, maternal and infant products, home furnishings, local services, and private domain operations. Each intelligent agent has an industry-specific knowledge base, execution strategy library, communication interface, and iterative model, supporting dynamic load awareness and elastic scheduling.

[0022] 2. Issuance of instructions Platform operators issue task instructions to the central control agent through a unified portal: "Plan and execute a nationwide spring renewal omnichannel promotional campaign, covering online malls, offline stores, private domain communities, short video platforms, and live streaming scenarios, and output complete campaign plans, category-specific copywriting, product recommendation strategies, pricing systems, inventory linkage solutions, and customer segmentation and outreach strategies." The central control agent completes intent recognition, target extraction, and constraint parsing, and generates a global task planning graph.

[0023] 3. Task breakdown The overall control agent, based on a dynamic programming algorithm, automatically decomposes the global task into 12 major categories and 236 parallelizable subtasks, including: -Overall rules and gameplay design of the event - Generation of full-category pricing strategies and discount rules - Personalized marketing copy tailored to specific user groups - Intelligent SKU recommendation and bundled package design -Online and offline inventory synchronization and allocation strategy Store Display and Salesperson Scripts - Private Community Reach Script -Short video / live stream content and scripts - Activity data tracking and performance monitoring rules - Risk control and price protection strategies -Customer service response database automatically generated - Construction of an activity review indicator system The system automatically establishes subtask dependencies and marks non-dependent subtasks as to be executed in parallel.

[0024] 4. Parallel Scheduling and Execution The central control agent selects and matches 18,200 target agents from 30,000 agents based on industry tags, real-time load, and historical execution efficiency, including: - 10,000 smart agents for retail stores - 5000 intelligent agents in the beauty product category - 3000 smart agents for clothing matching - 200 private domain operation intelligent agents The central control agent adopts a non-serialized waiting synchronization mechanism, which sends out all subtasks at once, triggering large-scale parallel execution.

[0025] 5. Cross-agent interconnected learning During execution, the intelligent agents interact in real time through the platform's unified communication protocol: -The apparel smart agent will synchronize high-conversion copywriting structure and matching logic with the beauty and home smart agents; - Retail intelligence agents share optimal discount rules and in-store conversion scripts for stores; - Private domain intelligent agents output high-open-rate title templates and interactive scripts; - All agents transform heterogeneous experiences into unified feature vectors through a semantic alignment protocol, thereby completing knowledge fusion and policy optimization.

[0026] 6. Results Aggregation The central control agent aggregates the execution results of 18,200 agents in real time, completing data verification, deduplication, fusion, sorting, and logical alignment. Within two hours, it outputs a complete "National Spring Omnichannel Promotion Master Plan," including: -326 sets of detailed activity gameplay -3000+ product category copywriting examples -1200 product combination strategies - Nationwide Store Tiered Implementation Manual - Omnichannel outreach script library - Real-time data monitoring dashboard rules 7. Iterative optimization After the campaign went live, the system continuously collected metrics such as conversion rate, average order value, click-through rate, store visit rate, and inventory turnover rate to construct a multi-dimensional reward function. Through reinforcement learning, the system updated the model parameters of 18,200 agents and the overall control agent.

[0027] After iteration, the efficiency of similar tasks increased by 35%, the accuracy of the solutions increased by 42%, and the capabilities evolved on their own.

[0028] Example 2: Enterprise-level Compliance and Risk Control Review 1. Scenario and Cluster Preparation A large corporate group commissioned a platform to conduct an annual comprehensive compliance and risk control review, covering areas such as contracts, finance, invoices, human resources, supply chain, data security, intellectual property, and anti-commercial bribery. The platform utilizes a dedicated cluster of compliance intelligent agents to ensure data isolation and secure execution.

[0029] 2. Task Instructions and Analysis The central control AI receives the instruction: "Conduct a comprehensive compliance review of all contract documents, financial vouchers, invoice ledgers, employee code of conduct implementation records, supply chain access and performance records, and data security logs for the entire group in the previous year, identify risk points, abnormal behaviors, non-compliant clauses, and potential fraud, and output a complete compliance review report and rectification suggestions." The central control agent completes intent parsing, scope definition, and review standard matching, generating a global review task planning diagram.

[0030] 3. Task breakdown The overall review task was broken down into 8 major categories and 168 parallel review sub-tasks: - Contract terms compliance scan - Verification of invoice authenticity and relevance - Detection of anomalies in financial documents -Employee behavior risk screening Supplier Admission and Performance Review Data security and access control auditing - Intellectual Property Infringement Risk Assessment - Anti-commercial bribery and conflict of interest review 4. Parallel scheduling and execution The central control agent matches 5,000 dedicated, compliant agents based on their professional capabilities. - 2000 legal contract intelligent agents - 1500 smart agents for financial invoices - 1000 HR risk control intelligent agents - 500 supply chain compliance intelligent agents All agents start up synchronously and execute in parallel without waiting, automatically reviewing over 100,000 documents, vouchers, and logs.

[0031] 5. Cross-agent interconnected learning During execution, the intelligent agent shares risk characteristics and identification rules in real time. -The legal intelligence agent synchronizes the characteristics of high-risk contract terms with the financial intelligence agent; -The financial intelligence agent shares rules regarding abnormal invoices and abnormal transaction records with the legal and supply chain intelligence agents; - The human resources intelligent agent outputs high-risk behavioral characteristics to collaboratively improve the overall risk identification rate.

[0032] - By integrating cross-domain knowledge, we can achieve the discovery of complex correlation risks that a single intelligent agent cannot accomplish.

[0033] 6. Results Aggregation and Risk Output The central control intelligent agent aggregates all review results, completes risk classification, evidence chain assembly, and duplicate risk merging, and outputs a complete compliance review report within one business day, including: - Overall Compliance Score High / Medium / Low Risk List -Abnormal contracts, invoices, and transaction details -Risk Sources and Correlation Analysis -Suggestions for rectification and timeline -Directions for Institutional Optimization During this review, the system successfully identified a hidden commercial fraud and irregular fund transfer through cross-verification by legal and financial intelligence agents, preventing economic losses of several million yuan.

[0034] 7. Self-evolutionary iteration The system calculates reward signals based on review accuracy, risk recall, efficiency, and collaboration, and uses these signals to update the compliance agent model through reinforcement learning, optimizing the risk identification rule base. Subsequent similar review tasks saw a 28% increase in risk detection rate and a 60% reduction in review time.

[0035] Although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to the above embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for collaborative scheduling, parallel execution, and self-evolution of tens of thousands of industry intelligent agents on a platform, characterized in that: Includes the following steps: S1. Cluster Initialization: Build a cluster of tens of thousands of intelligent agents covering thousands of industries on the Dianjita platform. Each intelligent agent has a unique industry business capability tag, industry knowledge base, communication interface and initial iteration model. S2, Command Access: The unified task command input from the outside is received by the unique central control intelligent agent within the platform, and the task command is subjected to intent recognition, target extraction and constraint condition parsing to generate a global task planning graph; S3. Task decomposition: The overall control agent automatically decomposes the complex global task into multiple independently executable parallel subtasks according to the global task planning graph, and constructs the dependency relationship and execution priority between the subtasks. S4. Parallel Scheduling and Execution: The central control agent selects a suitable target agent group from the tens of thousands of agents clusters based on the industry attributes corresponding to each sub-task and the current load status of the agent, issues corresponding sub-task instructions, and schedules the target agent group to start and execute the sub-tasks in parallel. S5. Cross-Agent Interconnection Learning: During task execution, the target intelligent agent group shares execution status, intermediate results, industry experience and optimal strategies in real time through the communication protocol built into the platform, realizing cross-industry and cross-scenario mutual learning and knowledge integration. S6. Result Aggregation: The overall control agent automatically aggregates the subtask execution results of all target agents, performs data verification, deduplication, fusion and sorting, and generates and outputs the global final task results. S7. Self-iterative optimization: Based on the performance evaluation data of this task, a multi-dimensional reward function is constructed. The model parameters of the overall control agent and the participating industry agents are updated through reinforcement learning algorithms to realize the autonomous iteration and capability evolution of the agents.

2. The method according to claim 1, characterized in that, The number of target intelligent agents in step S4 ranges from 1,000 to 50,000, and supports dynamic intelligent expansion based on task complexity. All participating intelligent agents adopt a synchronous execution mechanism without serial waiting.

3. The method according to claim 1, characterized in that, The cross-agent interconnection learning described in step S5 specifically includes: converting heterogeneous data from different industries into unified feature vectors through semantic alignment protocols, sharing best practice cases in the execution process, optimizing one's own industry rule base and execution strategy, and realizing the transfer and complementarity of capabilities.

4. The method according to claim 1, characterized in that, The multi-dimensional reward function mentioned in step S7 includes task completion rate, execution efficiency, result accuracy, cross-agent collaborative score, and resource utilization. A global reward signal is generated through weighted calculation to drive the agent to update the model.

5. A collaborative scheduling, parallel execution, and self-evolution system for tens of thousands of industry intelligent agents on a platform, used to implement the method described in any one of claims 1-4, characterized in that, include: The 10,000-level industry intelligent agent cluster module is used to deploy and manage tens of thousands of distributed intelligent agents covering thousands of industries; The central control intelligent agent scheduling module, as the system's sole entry point, is responsible for instruction parsing, task decomposition, and global scheduling. The parallel task execution module is used to schedule the selected group of agents to execute subtasks in parallel. The cross-agent interconnection learning module is used to enable real-time communication, experience sharing, and knowledge transfer between agents; The self-iterative optimization module is used to update the agent model through reinforcement learning based on execution feedback data. The results aggregation and output module is used to summarize, verify, and output the global task results.

6. The system according to claim 5, characterized in that, The tens of thousands of industry intelligent agent cluster modules are deployed using a distributed containerized architecture, which supports the dynamic creation, destruction, and elastic scaling of intelligent agents, ensuring low-latency communication and stable operation of large-scale intelligent agents.

7. The system according to claim 5, characterized in that, The overall control intelligent agent scheduling module integrates an intent recognition engine and a dynamic programming algorithm, which can adjust the task decomposition strategy and resource allocation scheme according to real-time environmental data.

Images