Intelligent agent generation method and device, medium and product

Abstract

The invention discloses an agent generation method and device, a medium and a product. The method comprises the steps of obtaining task description information, and performing semantic analysis to obtain a semantic analysis result; determining a plurality of sub-tasks and task dependency relationships and task constraint conditions among the sub-tasks based on the semantic analysis result; performing information matching according to the description information of the subtasks and node description information in a capability graph to obtain a first capability combination; generating a first agent set and an agent interaction relationship according to the task dependency relationship, the task constraint condition and the first capability combination; and scheduling and operating the agents in the first agent set according to the task dependency relationship and the agent interaction relationship to complete the generation process of the target multi-agent set. According to the technical scheme, the problem that a static agent structure cannot meet service requirements is solved, the agent combination matched with the task can be determined through semantic analysis and the capability atlas, and the agent determination flexibility and the task adaptability are improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and in particular to a method, device, medium and product for generating intelligent agents. Background Technology

[0002] Large-model-based agent systems, as key execution units in artificial intelligence systems, possess capabilities such as environmental perception, autonomous decision-making, and goal-oriented execution, and can be widely applied in multiple fields.

[0003] While existing multi-agent systems (MAS) and multi-agent collaboration technologies achieve a certain degree of flexibility and scalability when facing complex tasks and changing environments through task decomposition and role collaboration, most agents rely on pre-set static agent structures and cannot adaptively generate or adjust agent combinations according to different tasks. This leads to problems such as response lag, resource waste, or collaboration failure when facing new tasks, sudden changes, or cross-scenario collaboration. Summary of the Invention

[0004] This invention provides a method, device, medium, and product for generating intelligent agents, which can determine combinations of intelligent agents that match tasks through semantic parsing and capability graphs, thereby improving the flexibility of agent determination and the adaptability to tasks.

[0005] In a first aspect, embodiments of the present invention provide a method for generating an intelligent agent, the method comprising:

[0006] In response to the triggering of agent generation, the task description information used to guide agent generation is obtained, and the task description information is semantically parsed to obtain the semantic parsing result;

[0007] Based on the semantic parsing results, the task description information is decomposed into multiple subtasks, and the task dependencies between subtasks and the task constraints of each subtask are determined.

[0008] The first capability combination is obtained by matching the task description information of each subtask with the node description information in the first capability graph.

[0009] A first set of intelligent agents is generated based on task dependencies, task constraints, and a combination of first capabilities, and the interaction relationships between the intelligent agents in the first set of intelligent agents are determined.

[0010] Based on task dependencies and agent interaction relationships, the agents in the first agent set are scheduled and run, and in response to the successful operation of the agents in the first agent set, the process of generating the target multi-agent set is completed.

[0011] In a second aspect, embodiments of the present invention provide an intelligent agent generation device, the device comprising:

[0012] The semantic parsing module is used to respond to the trigger of agent generation, obtain task description information to guide agent generation, and perform semantic parsing on the task description information to obtain the semantic parsing result;

[0013] The subtask determination module is used to decompose the task description information into multiple subtasks based on the semantic parsing results, and determine the task dependencies between subtasks and the task constraints of each subtask.

[0014] The capability combination determination module is used to match the task description information of each subtask with the node description information in the first capability graph to obtain the first capability combination.

[0015] The agent set determination module is used to generate a first agent set based on task dependencies, task constraints, and a first capability combination, and to determine the interaction relationships between agents in the first agent set.

[0016] The agent generation module is used to schedule and run agents in the first agent set according to task dependencies and agent interaction relationships, and to complete the generation process of the target multi-agent set in response to the successful operation of agents in the first agent set.

[0017] Thirdly, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the intelligent agent generation method as described in any of the embodiments of the present invention.

[0018] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent agent generation method as described in any of the embodiments of the present invention.

[0019] Fifthly, embodiments of the present invention also provide a computer program product, including a computer program that, when executed by a processor, implements the agent generation method as described in any of the embodiments of the present invention.

[0020] In this embodiment of the invention, in response to the triggering of agent generation, task description information for guiding agent generation is obtained, and semantic parsing is performed on the task description information to obtain semantic parsing results. Based on the semantic parsing results, the task description information is decomposed into multiple sub-tasks, and the task dependencies between sub-tasks and the task constraints of each sub-task are determined. Information matching is performed between the task description information of each sub-task and the node description information in the first capability graph to obtain a first capability combination. A first agent set is generated based on the task dependencies, task constraints, and the first capability combination, and the agent interaction relationships in the first agent set are determined. Agents in the first agent set are scheduled and run according to the task dependencies and agent interaction relationships, and the generation process of the target multi-agent set is completed in response to the successful operation of the agents in the first agent set. The technical solution of this invention solves the problem that static agent structures cannot meet business requirements. It can determine agent combinations that match tasks through semantic parsing and capability graphs, improving the flexibility of agent determination and adaptability to tasks. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of an agent generation method provided in an embodiment of the present invention;

[0023] Figure 2 A flowchart illustrating another method for generating an intelligent agent provided in an embodiment of the present invention;

[0024] Figure 3 A flowchart illustrating another method for generating an intelligent agent provided in an embodiment of the present invention;

[0025] Figure 4 This is a schematic diagram of the intelligent agent generation process provided in an embodiment of the present invention;

[0026] Figure 5 This is a schematic diagram of the structure of an intelligent agent generation device provided in an embodiment of the present invention;

[0027] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0028] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0029] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance. The acquisition, storage, use, and processing of data in the technical solutions of this application all comply with the relevant provisions of national laws and regulations.

[0030] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the relevant content of the solution.

[0031] Figure 1 This is a flowchart illustrating an agent generation method provided in an embodiment of the present invention. This embodiment is applicable to scenarios involving agent generation. The method can be executed by an agent generation device, which can be implemented in software and / or hardware and integrated into a computer device.

[0032] like Figure 1 As shown, the agent generation method includes the following steps:

[0033] S110. In response to the triggering of agent generation, obtain task description information used to guide agent generation, and perform semantic parsing on the task description information to obtain the semantic parsing result.

[0034] This embodiment can be applied to natural language processing tasks in various fields, such as government affairs, education, manufacturing, transportation, medical care, finance, and scientific research.

[0035] In response to the generation of the intelligent agent, the system obtains natural language task description information input by the user. This task description information may include the core business objectives of the task, execution scenario limitations, execution constraints, and output format requirements, etc., to guide the intelligent agent in generating task results corresponding to the task description information, such as generating a market analysis report for a specific field.

[0036] Semantic parsing of task descriptions using natural language processing (NLP) transforms ambiguous natural language task descriptions into structured information that the system can recognize. This semantic parsing can be performed using one or more models, such as the Large Language Model (LLP), BERT, and TextCNN, or a combination of several models.

[0037] The semantic parsing results can include the core intent of the task, the execution object, the business scenario, and the target result.

[0038] This embodiment obtains the semantic parsing results of the task description information through natural language processing, so as to facilitate subsequent task decomposition based on the semantic parsing results.

[0039] S120. Based on the semantic parsing results, the task description information is decomposed into multiple sub-tasks, and the task dependencies between sub-tasks and the task constraints of each sub-task are determined.

[0040] Based on the semantic parsing results, the original task corresponding to the task description information is decomposed into multiple independently executable subtasks, and the task dependencies between subtasks and the task constraints of each subtask are determined.

[0041] Determining the task dependencies between subtasks and the task constraints of each subtask can be achieved by using pre-trained language models such as BERT and GPT to perform deep semantic parsing of the task description information, extracting action sequence words, constraint keywords, and business logic words from the text.

[0042] Based on the extracted temporal and logical vocabulary, the execution relationships between subtasks are analyzed using graph neural network reasoning models or rule-based reasoning engines, thereby identifying directed or weighted directed graphs that represent task dependencies.

[0043] The parsed constraint keywords are matched with a pre-defined constraint type library. The constraint category is determined by a text classification model such as TextCNN. Combined with specific values ​​and requirements, the corresponding task constraints are defined for each subtask from dimensions such as execution time limit, data format, permission scope, and result acceptance criteria.

[0044] Task dependencies can include sequential dependencies, parallel dependencies, and conditional dependencies, such as "first complete the complaint content classification, and then perform after-sales solution matching".

[0045] The constraints for each subtask may include execution time limits, data format requirements, permission scope, and result acceptance criteria.

[0046] This embodiment reduces the complexity of task execution by breaking it down, and ensures that the execution logic of subtasks conforms to business rules by defining dependencies and constraints.

[0047] S130. Match the task description information of each subtask with the node description information in the first capability graph to obtain the first capability combination.

[0048] The first capability graph can be a pre-built, standardized capability index library, appearing as a dynamically updated graph in real time. For example, by continuously detecting newly added available tools or components inside and outside the system, their functional descriptions are converted into vector representations (i.e., node description information) using a word embedding model. The tools or components are then used as nodes in the first capability graph, and the vectors are stored in the graph as semantic feature attributes of the nodes, thereby achieving continuous updating of the first capability graph.

[0049] Similar to the original task, subtasks are also represented using natural language, i.e., task description information. After determining the task description information of the subtasks, a word embedding model is also used to convert the task description information of the subtasks into vector representations. This allows for information matching between the task description information of each subtask and the node description information in the first capability graph, for example, by calculating the cosine similarity between vectors.

[0050] By matching information, suitable capability units are matched for each subtask to obtain the first capability combination required to complete the original task.

[0051] This embodiment utilizes the indexing characteristics of the capability map to achieve precise mapping between the sub-task required capabilities and existing available capability units, thereby realizing resource matching.

[0052] S140. Generate a first set of intelligent agents based on task dependencies, task constraints, and the combination of first capabilities, and determine the interaction relationships between the intelligent agents in the first set of intelligent agents.

[0053] In this embodiment, the hierarchical mapping relationship from capability to agent is as follows: a capability can correspond to a functional component that implements that capability, and at least one component or tool of a capability is combined together to generate the corresponding agent.

[0054] Therefore, by combining task dependencies, task constraints, and the first capability combination, the system groups and maps the capability units of each subtask to the corresponding intelligent agent carrier, or creates intelligent agents with corresponding capabilities, and integrates them to form the first intelligent agent set; that is, the first intelligent agent set includes multiple intelligent agents with different functions, and multiple intelligent agents with different functions realize the tasks corresponding to the task description information.

[0055] While determining the first set of agents, the interaction relationships between agents are determined based on the dependencies of sub-tasks, such as data transmission relationships, result feedback relationships, and collaborative triggering relationships.

[0056] This embodiment transforms capability combinations into an executable intelligent agent system, clarifying the collaborative rules of multiple intelligent agents.

[0057] S150. Schedule and run the agents in the first agent set according to the task dependency relationship and agent interaction relationship, and complete the generation process of the target multi-agent set in response to the successful operation of the agents in the first agent set.

[0058] Based on task dependencies and agent interactions, agents in the first agent set are launched according to preset scheduling rules, and the running status and execution results of each agent are monitored. Preset scheduling rules include priority scheduling and resource load scheduling.

[0059] Once all agents have successfully completed their tasks, the process of generating the target multi-agent set is complete, yielding the task results of the original task corresponding to the task description information that guided the agent generation.

[0060] The technical solution of this invention, in response to the triggering of agent generation, acquires task description information to guide agent generation, performs semantic parsing on the task description information to obtain semantic parsing results; based on the semantic parsing results, the task description information is decomposed into multiple sub-tasks, and the task dependencies between sub-tasks and the task constraints of each sub-task are determined; information matching is performed between the task description information of each sub-task and the node description information in a first capability graph to obtain a first capability combination; a first agent set is generated based on the task dependencies, task constraints, and the first capability combination, and the agent interaction relationships in the first agent set are determined; agents in the first agent set are scheduled and run according to the task dependencies and agent interaction relationships, and the generation process of the target multi-agent set is completed in response to the successful operation of the agents in the first agent set. The technical solution of this invention solves the problem that static agent structures cannot meet business requirements, and can determine agent combinations that match tasks through semantic parsing and capability graphs, improving the flexibility of agent determination and adaptability to tasks.

[0061] Figure 2 This is a flowchart illustrating an agent generation method provided in an embodiment of the present invention. This embodiment belongs to the same inventive concept as the agent generation methods in the above embodiments, and further describes the process of determining a first capability combination. This method can be executed by an agent generation device, which can be implemented in software and / or hardware and integrated into an electronic device with application development capabilities.

[0062] like Figure 2 As shown, the agent generation method in this embodiment includes the following steps:

[0063] S210. In response to the triggering of agent generation, obtain task description information used to guide agent generation, and perform semantic parsing on the task description information to obtain the semantic parsing result.

[0064] S220. Based on the semantic parsing results, the task description information is decomposed into multiple sub-tasks, and the task dependencies between sub-tasks and the task constraints of each sub-task are determined.

[0065] For example, the task parsing module based on the large model performs dependency parsing and key phrase extraction to determine task dependencies, decomposing the user task R into a set of subtasks V = {T1, T2, …, T}. n}, determine the set of task dependencies The elements in E are ordered pairs (T) consisting of two subtasks. i ,T j ), where T i <T j Subtask T i The execution order is earlier than that of subtask T. j That is, T i It is T j The preceding dependent subtasks are determined. This leads to the determination of the dependency graph G=(V,E) representing the task dependencies, and the task constraints for each subtask.

[0066] Specifically, by leveraging the dependency parsing capabilities of the large model, the sentence structure of the user task description is analyzed, such as verb temporal order and logical connectors, to identify the sequential relationship of task actions. At the same time, key phrases such as business actions and constraints are extracted to decompose the user task R into a set of subtasks V={T1,T2,…,Tn} with independent execution meaning.

[0067] By combining the action sequence relationship obtained from syntactic analysis, the association of which subtasks are executed first is stored in set E in the form of ordered pairs. Finally, a task dependency graph is constructed through subtask nodes V and dependency edges E.

[0068] S230. Calculate the semantic similarity between each node in the first capability graph and each subtask, as well as the input and output format matching degree, based on the task description information of each subtask and the node description information in the first capability graph.

[0069] The system organizes its internal and external callable capability resources in the form of a graph. Nodes in the first capability graph represent functional modules, edges represent dependencies and combinations, and each module is accompanied by capability metadata, which may include functionality, input / output formats, and execution constraints. Modules support semantic retrieval and matching operations. Based on the sub-task requirements output by the task parsing module, semantic similarity calculation and contextual constraint filtering are used to calculate the semantic similarity between each node and each sub-task in the first capability graph, as well as the input / output format matching degree.

[0070] Specifically, the process of calculating the matching degree between the first capability graph node and the sub-task through semantic similarity calculation and context constraint filtering can be as follows: First, the task description information of the sub-task is converted into a vector through embedding, and then the cosine similarity is calculated between the vector and the semantic feature vector of the node in the first capability graph to obtain the semantic similarity between the two. The semantic feature vector includes the vector representation of the tool or component function description information corresponding to the node.

[0071] Semantic similarity can be calculated using the following formula:

[0072] ;

[0073] in A vector describing the information of a subtask. For the semantic feature vectors of the capability graph nodes, cosine similarity is used to calculate the correlation between task intent and capability.

[0074] Then, by combining contextual constraints for filtering, we can verify the degree of matching between the input data format and output requirements of the subtask and the input and output parameter formats of the tool calls corresponding to the graph nodes.

[0075] In this embodiment, to reduce the scale of vector computation, preliminary filtering can be performed based on the inverted index of task keywords and capability tags to obtain a candidate capability set C', and then the matching degree of C' can be calculated. For example, to improve the matching efficiency between subtasks and nodes in the first capability graph, a semantic feature vector index is constructed based on the FAISS or Milvus vector database. The semantic feature vectors of each node in the graph are stored in the vector library and an index is constructed. By utilizing the approximate nearest neighbor retrieval capabilities of FAISS and Milvus, a retrieval complexity of O(logN) is achieved, quickly locating the candidate node with the highest semantic similarity to the subtask, and significantly reducing the matching time in large-scale node scenarios.

[0076] Furthermore, the results of high-frequency task-capability matching can be cached to improve the speed of secondary calls.

[0077] S240. The ability matching score of each node with each subtask is obtained by weighting the semantic similarity, the matching degree of the input and output formats and the historical success rate of the node's capabilities.

[0078] After calculating semantic similarity and input / output format matching, the comprehensive ability matching score is calculated by weighted summation, taking into account the historical success rate of the first ability graph node.

[0079] For example, the overall ability matching score can be calculated using the following formula:

[0080] ;

[0081] in, Indicates semantic similarity. Indicates the degree of matching between input and output formats. The node's historical success rate represents its capabilities, β and γ are adjustable weights, c represents the node, and t represents the subtask.

[0082] S250. Determine the nodes that match each subtask based on the ability matching score to obtain the first ability combination.

[0083] The node with the highest overall ability matching score is determined as the node for subtask matching, thus obtaining the first ability combination {C1,…,Cm}.

[0084] S260. Generate a first set of intelligent agents based on task dependencies, task constraints, and the first capability combination, and determine the interaction relationships between the intelligent agents in the first set of intelligent agents.

[0085] S270. Schedule and run the agents in the first agent set according to the task dependency relationship and agent interaction relationship, and in response to the successful operation of the agents in the first agent set, complete the process of generating the target multi-agent set.

[0086] During task execution, the system collects and analyzes the agent's operational status and dynamically updates the subtask set and dependency graph G. Specifically, when a subtask fails, the execution result generates new branch conditions, or external environmental changes necessitate adjustments to the task path, the system will partially or completely reconstruct the task graph corresponding to the task dependencies based on the latest feedback.

[0087] Specifically, in one optional implementation, during the execution of agents in the first set of agents, in response to the failure of the first agent in the first set of agents, the task dependencies and / or task constraints associated with the subtasks corresponding to the first agent are adjusted.

[0088] When running the first set of intelligent agents, if the first intelligent agent fails to execute its associated subtask, the system first marks the subtask node as abnormal, triggering the task dependency chain interruption detection mechanism; then, based on the first capability graph, it re-searches for capability nodes that match the subtask and generates an alternative execution plan; finally, it dynamically adjusts the task dependencies associated with the subtask according to the alternative plan, such as inserting new subtask nodes, modifying the direction of the original dependency edges, or relaxing or adjusting the task constraints of the subtask, such as extending the execution time limit or simplifying data format requirements, thereby reconstructing the task execution path and ensuring that the blocked overall task can continue to advance along the new path.

[0089] In one alternative implementation, during the operation of agents in the first set of agents, in response to a new task branching condition generated by a second agent in the first set of agents during operation, a new subtask is added based on the new task branching condition, and the task dependencies are updated based on the new subtask.

[0090] During the operation phase of the first set of intelligent agents, if a new task branch condition is generated during the execution of the second intelligent agent, the system first decomposes the new subtask with independent execution significance based on the business logic of the branch condition; then, combined with the structure of the original task dependency relationship, the new subtask is integrated into the task dependency graph as a node, and the pointing and association logic of the edges are updated, such as setting the new subtask as a predecessor / desire of a certain node, or as a new path node of the condition branch, and finally completing the dynamic reconstruction of the task dependency relationship to ensure that the execution logic of the new branch task is consistent with the original task chain.

[0091] In one alternative implementation, during the operation of agents in the first set of agents, in response to the subtask corresponding to the third agent in the first set of agents being identified as a redundant or invalid task node, the third agent is deleted and the task dependencies are updated.

[0092] By using preset node validity rules, such as whether the subtask output provides actual data support for subsequent tasks and whether the execution result is a necessary condition, the subtask corresponding to the third agent is identified as a redundant or invalid task node. Then, the third agent is directly deleted, and its association with other agents is severed. Finally, based on the structure of the original task dependency graph, the subtask connection logic before and after the deleted node is reorganized, the dependency edge corresponding to the node is deleted, and a new directed edge is re-established to complete the dynamic update of task dependency relationships, ensuring the simplicity and execution efficiency of the task chain.

[0093] Through the above adjustment mechanism, the task parsing module and the scheduling execution module can realize the iterative evolution of the subtask set {T1,…,Tn} and the dependency relationship G, so that the system can maintain adaptability and robustness to the dynamic environment during the execution process.

[0094] The technical solution of this embodiment, in response to the triggering of agent generation, acquires task description information to guide agent generation, and performs semantic parsing on the task description information to obtain semantic parsing results; based on the semantic parsing results, the task description information is decomposed into multiple sub-tasks, and the task dependencies between sub-tasks and the task constraints of each sub-task are determined; the semantic similarity and input-output format matching degree of each node in the first capability graph are calculated based on the task description information of each sub-task and the node description information in the first capability graph; a weighted calculation is performed based on the semantic similarity, input-output format matching degree, and the historical success rate of the node's capabilities to obtain the capability matching score of each node and each sub-task; the nodes matching each sub-task are determined based on the capability matching score to obtain the first capability combination. A first agent set is generated based on the task dependencies, task constraints, and the first capability combination, and the agent interaction relationships in the first agent set are determined; the agents in the first agent set are scheduled and run based on the task dependencies and agent interaction relationships, and the generation process of the target multi-agent set is completed in response to the successful operation of the agents in the first agent set. The technical solution of this invention solves the problem that static agent structures cannot meet business needs. It can determine the combination of capabilities that match the task by matching subtask semantics with capability graphs, and further determine the combination of agents, thereby decoupling the task, capabilities and agents, and further improving the flexibility of agent determination and adaptability to the task.

[0095] Figure 3 This is a flowchart illustrating an agent generation method provided in an embodiment of the present invention. This embodiment belongs to the same inventive concept as the agent generation methods in the above embodiments, and further describes the process of determining a first set of agents. This method can be executed by an agent generation device, which can be implemented in software and / or hardware and integrated into an electronic device with application development capabilities.

[0096] like Figure 3 As shown, the agent generation method in this embodiment includes the following steps:

[0097] S310. In response to the triggering of agent generation, obtain task description information used to guide agent generation, and perform semantic parsing on the task description information to obtain the semantic parsing result.

[0098] S320. Based on the semantic parsing results, the task description information is decomposed into multiple subtasks, and the task dependencies between subtasks and the task constraints of each subtask are determined.

[0099] S330. Match the task description information of each subtask with the node description information in the first capability graph to obtain the first capability combination.

[0100] S340. Determine the number of agents and their input / output responsibilities based on task dependencies and task constraints.

[0101] Features such as subtask parallelism and task constraints can be derived based on task dependencies, and the number of agents can be determined based on these derived features. According to the data flow logic in the task dependencies, the input responsibilities of each agent can be clarified, such as receiving the output or raw data of the preceding subtask, as well as the output responsibilities, such as meeting the input format of the following subtask and conforming to its own constraints, to ensure that the collaboration of agents is consistent with the task execution logic.

[0102] S350. Under the constraints of a preset maximum number of attempts and a preset optimization objective, select the corresponding capability from the first capability combination to execute the sub-task based on the number of agents and the input / output responsibilities of the agents, so as to generate the first set of agents.

[0103] Initiate an iterative trial with a maximum of no more than a preset maximum number of attempts, such as M. The iterative trial aims to generate the optimal agent combination A={A1,…,Ak} based on the first capability combination. The objective function is the preset optimization goal.

[0104] In the optimization objective function, SR(A) represents the success rate of agent combination A in completing the task, with a value range of [0,1]. COST(A) represents the resource consumption cost required to execute the combination. REUSE(A) represents the reuse rate of historical agent structures or capability modules in the combination, which is a proportional value. represents the weighting coefficient, used to adjust the importance of each optimization objective; A represents the set of agents, which is the decision variable in the optimization process.

[0105] Based on the first capability combination and task requirements, an initial set of intelligent agents is constructed. If the initial generation does not meet the preset criteria, the optimization iteration continues based on the previous result, with a maximum of M attempts. During the iteration process, if the parsing result meets the preset criteria, the loop terminates; otherwise, the first capability combination is updated according to the current iteration parsing result, and structured parsing is re-executed.

[0106] The analysis results are the actual performance data of the agent combination in the iteration, including task success rate, resource consumption cost, capability reuse rate, and verification results of whether its execution logic matches task dependencies and constraints.

[0107] Preset criteria are the pre-defined rules for determining whether a combination meets the criteria. They may include optimization target thresholds, task adaptation requirements, feasibility conditions, etc.

[0108] If the first capability combination involves tools or functional components that do not yet exist in the system, the tool configuration module will be invoked to build the new tool; if the initial build fails, a maximum of M build correction attempts will be made.

[0109] S360. Schedule and run the agents in the first agent set according to the task dependency relationship and agent interaction relationship, and in response to the successful operation of the agents in the first agent set, complete the generation process of the target multi-agent set.

[0110] The system performs execution flow orchestration and real-time scheduling on the generated multi-agent system. Based on the task graph structure with task dependencies, it selects appropriate execution strategies, such as sequential, parallel, and branch execution, and coordinates information flow and state synchronization among agents. The system features an extensible execution strategy plugin mechanism, supporting dynamic adjustment of the execution plan based on task attributes and execution feedback, ensuring efficient collaborative operation of the multi-agent system.

[0111] Task scheduling uses DAG (Directed Acyclic Graph) topological sorting, supporting parallel, serial, and branch execution.

[0112] The scheduling process can be as follows: input the task dependency graph G, calculate the in-degree of each task (i.e., the number of dependencies), select tasks with an in-degree of 0 and add them to the batch set in sequence, and tasks within the same batch can be executed in parallel.

[0113] In one optional implementation, during the application process of obtaining the target multi-agent set, the running status information of each agent in the target multi-agent set is obtained; in response to any information anomaly in the running status information, the agent generation process based on the target multi-agent set is initiated; wherein, the running status information includes at least one of response time, error rate, and resource usage.

[0114] The system collects and analyzes the operational status of each agent, generating a feedback signal stream. Under triggered threshold conditions, the system's reconfiguration logic can be activated, such as reassigning tasks, replacing capability modules, and reconfiguring agents, enabling the system to possess self-awareness and self-optimization capabilities, thereby improving task success rate and robustness.

[0115] During the generation of target agents, failed tasks can be reassigned to backup agents with the same capabilities, and the candidate agent with the best historical performance can be selected first. The capability modules that perform poorly in the failed agents can be replaced by retrieving the optimal capability replacement through the capability graph scheduling module. The structure, task division, parameter configuration, and inference strategy of the agents are optimized as a whole. This may include operations such as merging / splitting, policy replacement, and execution logic reorganization.

[0116] The choice of refactoring strategy can be expressed by the following formula:

[0117] ;

[0118] Where S = {Reassign, Replace, Rebuild}, Gain SR ω(s) represents the estimated improvement in success rate brought about by the strategy, Cost(s) represents the additional resource cost required to implement the strategy, Risk(s) represents the uncertainty risk that the strategy may introduce, and ω1, ω2, ω3 represent the weighting coefficients.

[0119] The score for each candidate strategy is calculated using this formula, and the target strategy is determined based on the score.

[0120] The technical solution of this embodiment, in response to the triggering of agent generation, obtains task description information to guide agent generation, and performs semantic parsing on the task description information to obtain semantic parsing results; based on the semantic parsing results, the task description information is decomposed into multiple sub-tasks, and the task dependencies between sub-tasks and the task constraints of each sub-task are determined; information matching is performed between the task description information of each sub-task and the node description information in the first capability graph to obtain a first capability combination; the number of agents and the input / output responsibilities of agents are determined according to the task dependencies and task constraints; under the constraints of a preset maximum number of attempts and a preset optimization target, the corresponding capabilities are selected from the first capability combination to execute sub-tasks according to the number of agents and the input / output responsibilities of agents, so as to generate a first agent set. Agents in the first agent set are scheduled and run according to the task dependencies and agent interaction relationships, and the generation process of the target multi-agent set is completed in response to the successful operation of agents in the first agent set. The technical solution of this invention solves the problem that static intelligent agent structures cannot meet business needs. It can define intelligent agents by task dependence and constraints, and select suitable capabilities by combining capability graphs, thereby solving the problem that intelligent agents are difficult to adjust with tasks and improving their flexibility and adaptability.

[0121] In a specific instance of agent generation, the process of agent generation in the agent generation system corresponding to the agent generation method is illustrated as follows: Figure 4 As shown, in the first step, the user inputs a natural language task, and the task parsing module performs semantic parsing on the input statement to extract information such as task intent, subtask structure, and constraints.

[0122] The second step is for the capability graph scheduling module to match the required capability modules in the graph based on the parsed task intent, and generate a candidate set of capability combinations.

[0123] Third, the agent generation module dynamically derives the number of agents, their responsibilities, and interaction relationships based on the capability combination results and task structure requirements, forming a multi-agent collaborative system structure. The system instantiates the agents and registers them with the task scheduler, formally entering the task execution process.

[0124] The fourth step involves the execution scheduling module parsing the task and dynamically planning the subtask execution process to schedule the agent in parallel or serial mode.

[0125] Fifth, during task execution, the feedback sensing module continuously collects the operational status of each agent, such as response time, error rate, and resource usage. Once an anomaly is detected, the reconstruction trigger is activated, and the system automatically performs task reallocation and agent weight generation.

[0126] The sixth step involves lifecycle management and agent reuse. The system persistently stores the behavioral data and capability call logs of each agent after execution, including role definitions and execution success rates. When similar tasks arrive later, existing agent structures or their capability modules are quickly matched and reused, saving generation time. For agents that have not been called for a long time, the lifecycle manager releases their resources and archives or downgrades them, optimizing system resource scheduling efficiency.

[0127] The lifecycle management module manages the lifecycle states of all agents in the system, including stages such as activation, running, idle, frozen, and deregistration. This module supports historical data persistence, recording information such as agent capability call logs, task execution results, and collaboration history. When a new task is triggered, archived agent structures or capability combinations can be reused, significantly shortening generation time. Simultaneously, for agents that have not been used for a long time, the module supports automatic resource reclamation and archiving strategies, optimizing system resource utilization efficiency.

[0128] Reuse matching can be determined using the following formula:

[0129] ;

[0130] Indicates task similarity, with values ​​ranging from [0,1]; Sim cap (A i H j The value ) represents the similarity of capability sets, that is, the degree of matching between the capabilities required to achieve the current task and the capabilities possessed by the archived agents, with a value range of [0,1]. This represents the weighting coefficient. When the score exceeds the threshold, the historical agent structure is directly reused to reduce generation costs.

[0131] Historical reuse processes may include:

[0132] (1) Task Trigger: Receive new task description and capability requirements;

[0133] (2) Candidate matching: Retrieve agents with similar structures / capabilities from the historical agent database;

[0134] (3) Reuse evaluation: Calculate the ReuseScore and select those with scores higher than the threshold θ. reuse Candidates;

[0135] (4) Reuse and adaptation: directly reuse or adjust historical intelligent agent modules as needed;

[0136] (5) Execution record persistence: Write the execution performance back to the historical database and update the performance data;

[0137] Periodically check the agent's most recent usage time (last_used) in the history database. If current_time - last_used > T... idle And ReuseScore < θ reuse If it is, then it is marked as recyclable.

[0138] In summary, this embodiment starts with task semantic parsing, connects the entire process from natural language requirements to multi-agent structure construction and collaborative execution, and achieves high adaptability, high intelligence and high efficiency operation of the agent collaborative system through capability graph scheduling, adaptive feedback adjustment and life cycle reuse mechanism.

[0139] This embodiment employs an adaptive agent generation mechanism based on task semantic understanding. It extracts task intent, subtask structure, and constraints from user-input natural language task descriptions and, combined with a capability graph, automatically derives the number of agents, their responsibilities, and interaction relationships. This multi-agent collaborative system does not rely on manually preset templates; it can adaptively generate and instantiate task execution units at runtime, improving the system's adaptability and scalability to diverse task scenarios. A modular capability scheduling method based on capability graphs and intent matching is introduced. Various atomic capability modules are modeled using functional semantics as an index, and after task parsing, capability combinations that meet specific conditions can be retrieved from the graph based on task intent and contextual requirements. This mechanism supports the decoupling and binding of capabilities, roles, and agents, and implements dynamic scheduling strategies through a configurable weight mechanism, providing precise and flexible capability support for agent generation. It also supports an agent adaptive reconstruction mechanism driven by execution process feedback, automatically triggering the multi-agent system agent reconstruction process by real-time sensing of the operational status of each agent in the system. This process includes operations such as redistributing task steps, changing role capability modules, and merging or splitting agents, enabling the system to have the ability to self-regulate and evolve when faced with sudden anomalies or environmental changes, thus ensuring task continuity and success rate.

[0140] Figure 5 This is a schematic diagram of the intelligent agent generation device provided in an embodiment of the present invention. This embodiment is applicable to scenarios involving intelligent agent generation. The device can be implemented in software and / or hardware and integrated into a computer device.

[0141] like Figure 5 As shown, the agent generation device includes: a semantic parsing module 410, a subtask determination module 420, an agent combination determination module 430, an agent set determination module 440, and an agent generation module 450.

[0142] The semantic parsing module 410 is used to obtain task description information to guide the generation of intelligent agents in response to the triggering of agent generation, and to perform semantic parsing on the task description information to obtain the semantic parsing result; the subtask determination module 420 is used to decompose the task description information into multiple subtasks based on the semantic parsing result, and to determine the task dependency relationship between subtasks and the task constraint condition of each subtask; the capability combination determination module 430 is used to match the task description information of each subtask with the node description information in the first capability graph to obtain the first capability combination; the intelligent agent set determination module 440 is used to generate the first intelligent agent set according to the task dependency relationship, task constraint condition and the first capability combination, and to determine the interaction relationship of the intelligent agents in the first intelligent agent set; the intelligent agent generation module 450 is used to schedule and run the intelligent agents in the first intelligent agent set according to the task dependency relationship and the interaction relationship of the intelligent agents, and to complete the generation process of the target multi-intelligent agent set in response to the successful operation of the intelligent agents in the first intelligent agent set.

[0143] The technical solution of this invention, in response to the triggering of agent generation, acquires task description information to guide agent generation, performs semantic parsing on the task description information to obtain semantic parsing results; based on the semantic parsing results, the task description information is decomposed into multiple sub-tasks, and the task dependencies between sub-tasks and the task constraints of each sub-task are determined; information matching is performed between the task description information of each sub-task and the node description information in a first capability graph to obtain a first capability combination; a first agent set is generated based on the task dependencies, task constraints, and the first capability combination, and the agent interaction relationships in the first agent set are determined; agents in the first agent set are scheduled and run according to the task dependencies and agent interaction relationships, and the generation process of the target multi-agent set is completed in response to the successful operation of the agents in the first agent set. The technical solution of this invention solves the problem that static agent structures cannot meet business requirements, and can determine agent combinations that match tasks through semantic parsing and capability graphs, improving the flexibility of agent determination and adaptability to tasks.

[0144] In an alternative implementation, during the operation of agents in the first agent set, the agent generation module 450 is further configured to:

[0145] In response to the failure of the first agent in the first agent set, adjust the task dependencies and / or task constraints associated with the subtasks corresponding to the first agent.

[0146] In an alternative implementation, during the operation of agents in the first agent set, the agent generation module 450 is further configured to:

[0147] In response to a new task branching condition generated by a second agent in the first agent set during operation, a new subtask is added based on the new task branching condition, and the task dependency relationship is updated based on the new subtask.

[0148] In an alternative implementation, during the operation of agents in the first agent set, the agent generation module 450 is further configured to:

[0149] In response to the subtask corresponding to the third agent in the first agent set being identified as a redundant or invalid task node, the third agent is deleted and the task dependency is updated.

[0150] In one alternative embodiment, the apparatus further includes:

[0151] An exception handling module is used to acquire the running status information of each agent in the target multi-agent set during the application process of acquiring the target multi-agent set; in response to any information exception in the running status information, to initiate the agent generation process based on the target multi-agent set; wherein, the running status information includes at least one of response time, error rate and resource usage.

[0152] In one alternative implementation, the energy combination determination module 430 is specifically used for:

[0153] The semantic similarity and input / output format matching degree of each node in the first capability graph are calculated based on the task description information of each subtask and the node description information in the first capability graph. The capability matching score of each node and each subtask is obtained by weighting the semantic similarity, input / output format matching degree and the historical success rate of the node's capability. The nodes that match each subtask are determined based on the capability matching score to obtain the first capability combination.

[0154] In one alternative implementation, the agent set determination module 440 is specifically used for:

[0155] The number of agents and their input / output responsibilities are determined based on task dependencies and constraints. Under the constraints of a preset maximum number of attempts and a preset optimization objective, the corresponding capabilities are selected from the first capability combination to execute sub-tasks, thereby generating the first set of agents.

[0156] The intelligent agent generation device provided in the embodiments of the present invention can execute the intelligent agent generation method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of executing the method.

[0157] Figure 6 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0158] like Figure 6 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0159] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0160] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as intelligent agent generation methods.

[0161] In some embodiments, the agent generation method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the agent generation method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the agent generation method by any other suitable means (e.g., by means of firmware).

[0162] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0163] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0164] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0165] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0166] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0167] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0168] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0169] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the agent generation method as provided in any embodiment of this application.

[0170] In implementing the computer program product, computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0171] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0172] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for generating an intelligent agent, characterized in that, include: In response to the triggering of agent generation, task description information used to guide agent generation is obtained, and semantic parsing is performed on the task description information to obtain semantic parsing results; Based on the semantic parsing results, the task description information is decomposed into multiple sub-tasks, and the task dependencies between the sub-tasks and the task constraints of each sub-task are determined. The first capability combination is obtained by matching the task description information of each subtask with the node description information in the first capability graph. A first set of intelligent agents is generated based on the task dependencies, the task constraints, and the first capability combination, and the interaction relationships between the intelligent agents in the first set of intelligent agents are determined. Based on the task dependencies and agent interaction relationships, the agents in the first agent set are scheduled and run, and in response to the successful operation of the agents in the first agent set, the generation process of the target multi-agent set is completed.

2. The method according to claim 1, characterized in that, The process of running the agents in the first set of agents also includes: In response to the failure of the first agent in the first agent set, the task dependencies and / or task constraints associated with the subtasks corresponding to the first agent are adjusted.

3. The method according to claim 1, characterized in that, The process of running the agents in the first set of agents also includes: In response to a new task branching condition being generated by a second agent in the first agent set during operation, a new subtask is added based on the new task branching condition, and the task dependency relationship is updated based on the new subtask.

4. The method according to claim 1, characterized in that, The process of running the agents in the first set of agents also includes: In response to the subtask corresponding to the third agent in the first set of agents being identified as a redundant or invalid task node, the third agent is deleted and the task dependency is updated.

5. The method according to claim 1, characterized in that, Also includes: During the application process of acquiring the target multi-agent set, the running status information of each agent in the target multi-agent set is acquired; In response to any information anomaly in the running status information, an agent generation process based on the target multi-agent set is initiated; The operational status information includes at least one of the following: response time, error rate, and resource usage.

6. The method according to claim 1, characterized in that, The step of matching the task description information of each subtask with the node description information in the first capability graph to obtain the first capability combination includes: Calculate the semantic similarity and input / output format matching degree between each node in the first capability graph and each subtask based on the task description information of each subtask and the node description information in the first capability graph; The capability matching score of each node and each subtask is obtained by weighting the semantic similarity, the input and output format matching degree and the historical success rate of the node's capability; The first capability combination is obtained by determining the node that matches each of the sub-tasks based on the capability matching score.

7. The method according to claim 1, characterized in that, The step of generating a first set of intelligent agents based on the task dependencies, the task constraints, and the first capability combination, and determining the interaction relationships among the intelligent agents in the first set of intelligent agents, includes: The number of agents and their input / output responsibilities are determined based on the task dependencies and task constraints. Under the constraints of a preset maximum number of attempts and a preset optimization target, the corresponding capability is selected from the first capability combination to execute the sub-task based on the number of agents and the input / output responsibilities of the agents, so as to generate a first set of agents.

8. An electronic device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the agent generation method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the agent generation method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the agent generation method as described in any one of claims 1-7.

Images