Task processing method and electronic equipment

Abstract

The invention discloses a task processing method and electronic equipment, and relates to the field of artificial intelligence, a first agent responds to a task processing request input by a user, and generates and executes an execution process comprising a plurality of task steps; the second agent responds to the situation that the task step currently executed by the first agent is a to-be-verified step, personalized verification is conducted on an output result generated when the first agent executes the to-be-verified step on the basis of personalized information of the user, and a corresponding verification guide instruction is generated and fed back to the first agent; and the first intelligent agent executes a subsequent processing step for the to-be-verified step based on the verification guide instruction so as to control the execution process to be continuously executed until a task processing result conforming to the personalized information is obtained.

Description

Technical Field

[0001] This application relates primarily to the field of artificial intelligence technology, and more specifically to a task processing method and an electronic device. Background Technology

[0002] With the application and rapid development of artificial intelligence, intelligent agents based on AI models such as large language models have been widely used in many business fields to autonomously perform tasks such as information generation and tool invocation to meet user needs.

[0003] However, due to the inherent "illusion" problem of artificial intelligence models, such as the model's reasoning-generated results / decisions not matching the user's actual needs, the tasks performed by the agent are inaccurate, requiring the user to input multiple times to continuously adjust the agent's task execution process and results, thus reducing the user's experience with the agent. Summary of the Invention

[0004] To address the above problems, this application provides the following technical solution:

[0005] A first aspect of this application provides a task processing method, the method comprising:

[0006] The first intelligent agent responds to the user's input task processing request, generates and executes an execution flow that includes multiple task steps;

[0007] The second agent responds to the fact that the task step currently being executed by the first agent is a step to be verified. Based on the user's personalized information, the second agent performs personalized verification on the output result generated by the first agent in executing the step to be verified, and generates a corresponding verification guidance instruction to be fed back to the first agent.

[0008] Based on the verification guidance instruction, the first intelligent agent executes subsequent processing steps for the step to be verified, so as to control the execution process to continue until a task processing result that conforms to the personalized information is obtained.

[0009] Optionally, the first intelligent agent, based on the verification guidance instruction, executes subsequent processing steps for the step to be verified, including any of the following:

[0010] The first intelligent agent, in response to the verification guidance instruction indicating that the personalized verification has passed, continues to execute the next task step;

[0011] In response to the verification guidance instruction indicating that the personalization verification failed, the first intelligent agent inputs the verification guidance instruction as a prompt message into the first model, guiding the first model to re-execute the current task steps.

[0012] Optionally, based on the user's personalized information, the output result of the first intelligent agent performing the verification step is performed in a personalized manner to generate corresponding verification guidance instructions, including:

[0013] Based on the attributes of the step to be verified, the target verification logic is determined among multiple candidate verification logics.

[0014] Based on the user's personalized information, the target verification logic is used to perform personalized verification on the output results generated by the first intelligent agent in performing the verification step, and generate corresponding verification guidance instructions.

[0015] Optionally, based on the user's personalized information, the target verification logic performs personalized verification on the output result generated by the first intelligent agent performing the verification step, generating corresponding verification guidance instructions, including:

[0016] Determine the object to be verified included in the output result of the first intelligent agent executing the step to be verified;

[0017] The target verification logic is used to evaluate the degree of conformity of the object to be verified relative to the user's personalized information.

[0018] Based on the degree of compliance, a verification guidance instruction is generated for the currently executed task step.

[0019] Optionally, the execution flow is generated based on the first agent's understanding of the task processing request's intent. Before generating the execution flow, it further includes:

[0020] The second intelligent agent performs personalized verification on the intent understanding results generated by the first intelligent agent based on the user's personalized information, and generates corresponding verification guidance instructions to be fed back to the first intelligent agent.

[0021] The first intelligent agent inputs the verification guidance instruction as a prompt into the first model, guiding the first model to re-understand the intent of the task processing request, and generates the execution flow based on the intent understanding result that conforms to the personalized information.

[0022] Optionally, the execution process is collaboratively generated by the first intelligent agent and the second intelligent agent based on the user's personalized information.

[0023] Optionally, the personalized verification process is implemented by the second intelligent agent through the second model, including:

[0024] The second intelligent agent sends the output result to the cloud, processes the output result through a data processing model, and feeds back the feature data of the generated output result to the second intelligent agent;

[0025] The second intelligent agent inputs the feature data into the personalization model and performs personalized verification on the feature data based on the stored personalized information of the user.

[0026] The second model is the personalized model deployed on the electronic device used by the user, or it includes the personalized model and the data processing model deployed in the cloud.

[0027] A second aspect of this application provides a task processing method, the method being executed by a first intelligent agent in an electronic device, comprising:

[0028] In response to a user's task processing request, an execution flow containing multiple task steps is generated;

[0029] Perform the task steps described above to generate the corresponding output results;

[0030] If the task step is identified as a step to be verified, a verification guidance instruction for the output result of the step to be verified is received; the verification guidance instruction is generated based on the user's personalized information verification.

[0031] Based on the verification guidance instruction, subsequent processing steps for the step to be verified are executed to control the execution process to continue until a task processing result that conforms to the personalized information is obtained.

[0032] A third aspect of this application provides a task processing method, the method being executed by a second intelligent agent in an electronic device, comprising:

[0033] In response to the execution of the step to be verified, the output result generated by the execution of the step to be verified is obtained; the step to be verified is a task step included in the execution flow generated based on the user-input task processing request.

[0034] Based on the user's personalized information, the output results are personalized for verification, and corresponding verification guidance instructions are generated.

[0035] Output the verification guidance command;

[0036] The verification guidance instruction is used to control subsequent processing steps for the step to be verified.

[0037] A fourth aspect of this application provides an electronic device, the electronic device comprising:

[0038] At least one memory, and a computer program stored in the memory;

[0039] At least one processor capable of running intelligent agents;

[0040] The intelligent agent can execute the computer program through the processor to implement the steps of the task processing method provided in the first, second, or third aspect. Attached Figure Description

[0041] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0042] Figure 1 This is a flowchart illustrating the task processing method proposed in Embodiment 1 of this application;

[0043] Figure 2 This is a flowchart illustrating the task processing method proposed in Embodiment 2 of this application;

[0044] Figure 3 This is a flowchart illustrating the task processing method proposed in Embodiment 3 of this application;

[0045] Figure 4 This is a flowchart illustrating the task processing method proposed in Embodiment 4 of this application;

[0046] Figure 5 This is a flowchart illustrating the task processing method proposed in Embodiment 5 of this application;

[0047] Figure 6 This is a flowchart illustrating the task processing method proposed in Embodiment Six of this application;

[0048] Figure 7 This is a flowchart illustrating the task processing method proposed in Embodiment 7 of this application;

[0049] Figure 8 This is a flowchart illustrating the task processing method proposed in this application under a business scenario example;

[0050] Figure 9 This is a schematic diagram of the task processing device proposed in Embodiment 1 of this application;

[0051] Figure 10 This is a schematic diagram of the task processing device proposed in Embodiment 2 of this application;

[0052] Figure 11 This is a schematic diagram of the task processing device proposed in Embodiment 3 of this application;

[0053] Figure 12 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0054] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is only for explaining specific embodiments and is not intended to limit the application. The embodiments of this application are described below with reference to the accompanying drawings. It will be understood by those skilled in the art that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0055] The terms "first," "second," etc., used in the context of this application and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0056] It is understood that before using the technical solutions disclosed in the embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained. For example, in response to receiving a user's active request, a pop-up window may be used, and a textual prompt message may be presented in the pop-up window to clearly inform the user that the requested operation will require the acquisition and use of the user's personal information. The user can choose whether to provide personal information to the electronic device, application, server, or storage medium or other software or hardware that performs the operation of the technical solution of this application based on the prompt message. This application does not limit the prompt message and the method of user authorization implementation. The data involved in this technical solution (including but not limited to the data itself, the acquisition or use of data) shall comply with the requirements of relevant laws and regulations and related provisions.

[0057] Regarding the technical problems described in the background section, this paper proposes using external tools or databases to verify the accuracy of the objective facts of the content generated by the model's inference. However, this post-fact-checking method is only applicable to objectively verifiable facts and cannot perceive the user's personalized information. For example, in a task request where the user inputs a request to edit content related to their favorite sports star, it cannot check whether the edited object A is a football player, which can easily lead to the edited result not conforming to the user's preferences. Another approach is to strengthen the prompt word engineering by adding static prompt words (such as "Please consider user preferences") before the user input, attempting to guide the model to generate personalized results. However, this method highly relies on the model's ability to understand and execute vague and general instructions, resulting in unstable and uncontrollable effects. It cannot perform real-time and reliable verification of whether the generated content meets the user's personalized needs during the inference process, especially for complex multi-step tasks, and cannot guarantee that the final result meets the user's personalized needs.

[0058] Furthermore, a general reliability enhancement mechanism is proposed, such as dynamic confidence assessment of model decisions. This mechanism can identify the uncertainty of the model itself (such as low-confidence decision points). However, its verification logic is general and inward-looking, and it is not designed for the key dimension of user personalized preferences. As a result, it cannot accurately identify scenarios where the model has high confidence but conflicts with user personalized preferences, such as recommending a steak restaurant to a vegetarian with high confidence. This reduces the accuracy and personalization of the model's inference results and affects the user experience.

[0059] To further improve the above-mentioned problems, this application proposes a novel task processing method. It introduces a dedicated second agent for personalized verification, which collaborates with a first agent executing each task step. Since the second agent performs real-time personalized verification on the output of the step to be verified currently executed by the first agent, and the verification logic is built based on the user's personalized information, it can accurately identify and correct conflicts between the output and the user's personalized information. Only then will the first agent be guided to execute the next task step, until all task steps are completed. This ensures that the final task processing result obtained by the first agent meets the user's personalized needs, eliminating the need for the user to repeatedly input prompts to adjust the model's task processing logic and eliminating the need for post-verification of the task processing results. This improves task processing efficiency and reliability, ensuring that the task processing result satisfies both the corresponding business objectives and the user's personalized information. The task processing method proposed in this application will be described in detail below with reference to the accompanying drawings.

[0060] Reference Figure 1This is a schematic diagram of the signaling flow for the task processing method proposed in Embodiment 1 of this application. The method proposed in this embodiment is applied to an electronic device, which can be a smartphone, laptop, smart wearable device, smart home device, vehicle terminal, or business terminal, etc., without limitation. Figure 1 As shown, the task processing method proposed in this embodiment may include, but is not limited to:

[0061] Step S11: The first intelligent agent responds to the task processing request input by the user, generates and executes an execution flow containing multiple task steps;

[0062] The task processing request can be an unstructured natural language request input by the user in text, voice, or other forms. It can include the user's natural language expression of the task to be processed / task processing requirements. Depending on the actual needs, it can also include relevant information required for the execution of the task, such as at least one task processing object among the text data, image data, and video data to be processed. For example, the user uploads the task processing object in the input box output by the first intelligent agent and inputs a natural language description in text form to describe the task that the first intelligent agent will perform on the task processing object, such as after uploading a video file, inputting "editing information related to sports stars", etc.

[0063] Of course, the user's input natural language can also include instructions to indicate the task processing object. For example, the user can directly input "Edit player-related information from the video file stored at xx (path)" or "Edit yesterday's live xxx game" in text, voice, or other forms in the input box output by the first intelligent agent. It is evident that the task processing object can originate from a specified storage device, external database, external device, or real-time acquisition device, determined by the source description of the task processing object included in the task processing request. This application addresses the content of the task processing request and its input format. Optionally, for corresponding types of signals in natural language input via voice or other forms, subsequent processing can be performed, or the signals can be converted to text information before subsequent processing; there are no restrictions on this.

[0064] In this way, the first intelligent agent can parse the task processing request to understand the core objectives and constraints of the user's task / intent (such as time, resources, roles, etc.), as well as the input / output requirements (such as the task's starting conditions and expected results / goals). Based on this, it can perform task planning, break down complex tasks into subtasks (such as indivisible atomic tasks), identify the logical relationships between different subtasks (such as sequential / temporal dependencies, parallel execution relationships, branching relationships, etc.), supplement each subtask with relevant task information (such as operation actions, required resources, or judgment conditions, etc.), and generate a complete task flow for implementing complex tasks as ordered task steps, such as a structured, executable sequence of task steps.

[0065] Therefore, in order to generate an execution flow that satisfies the task processing request, the capabilities of the first intelligent agent can be broken down into core components such as intent understanding, task decomposition, step planning, flow generation, and format output. To implement each capability, its technical architecture can be divided into the following layers: an input layer for receiving task processing requests expressed by the user in natural language; an understanding layer for completing intent understanding (such as intent recognition, entity extraction, and constraint parsing based on model / rule-based natural language processing techniques); a planning layer for task decomposition (such as recursive decomposition or splitting based on domain knowledge graphs) and step rules (such as the dependency analyzer identifying the temporal / logical dependencies of subtasks); and an output layer for feeding back the execution flow, which converts the planned multiple task steps into a structured format (such as sequence diagrams, flowcharts, or lists), to the task executor. Depending on actual needs, the capabilities and modules decomposed above can be merged or further decomposed. This application does not restrict the implementation methods of each layer and module and its capabilities in the technical architecture of the first intelligent agent.

[0066] For example, if the user inputs a task processing request as "help me edit a highlight reel of a football match", the task processing request is parsed according to the method described above. The user's desired video theme, content scope, and objectives are understood, i.e., task planning is performed, generating three task steps: "analyze reference video editing styles", "analyze highlight clips of editing materials", and "integrate highlight clips for video compositing". This forms a structured and executable execution flow, such as a task step sequence.

[0067] It should be noted that the implementation method of the first intelligent agent responding to task processing requests and generating execution flows includes, but is not limited to, the content described above. The first intelligent agent parses the task processing request and determines that the user has input the various task steps required to complete the task. That is, the task processing request contains natural language descriptions of multiple task steps. The first intelligent agent can directly convert these natural language descriptions of multiple task steps into a structured, executable execution flow. In this case, the first intelligent agent can recognize the content contained in the task processing request and quickly generate the execution flow after data format conversion. Of course, the first intelligent agent can also generate an optimized execution flow containing more / more detailed task steps based on the task step descriptions / execution flow descriptions in the task processing request, by refining the task steps or adding task steps, to improve the reliability and accuracy of task processing.

[0068] For example, if a user inputs a task request such as "Help me write the company's second quarter work summary report, including business data, problem analysis, and improvement plans, and generate a writing process," the first agent, based on the user's input requirements for the work summary report, retrieves relevant business data (such as sales revenue, customer volume, etc.) and administrative work data (such as attendance data) from the company's internal database. It then determines the structure of the work summary report, such as title → introduction → business data → administrative work → problem analysis → improvement plan → conclusion. Based on this, it generates an execution process with this structure for generating the work summary report. This execution process can include task steps that correspond one-to-one with each component of the structure, such as refining the title, generating the introduction, collecting business data, and compiling administrative work statistics, among other task steps for generating the work summary report.

[0069] After the first agent determines the execution flow for fulfilling the task processing request, it can execute the current task step according to the execution relationship indicated by the flow. This can be done by allocating execution resources through a task scheduler or workflow engine, or by calling internal functions or external tools. As needed, the agent can record task step execution status information, execution results (i.e., the generated output), and execution time, generating corresponding execution logs. Furthermore, the first agent can be configured with an exception handling module based on business requirements. This module can execute any of the following exception handling strategies—retry, alternative strategies, manual intervention, or process termination—in case the current task step fails, preventing prolonged lingering at the current task step and thus reducing user experience by preventing excessively long wait times for request responses. After executing a task step, the first agent obtains the corresponding output, which can be used as input for the next task step.

[0070] Therefore, the first intelligent agent in this application is not only the generator of the task execution process, but also the executor of the task execution process. This application does not restrict the implementation method of the first intelligent agent generating and executing each task step. It can be implemented by directly utilizing the functions of the first intelligent agent, or by calling a model or other external tools.

[0071] The first intelligent agent can be a general task execution process generation and execution tool (intelligent program), or it can be optimized into an intelligent execution carrier that focuses on specific business scenarios, possesses industry knowledge accumulation, and fits the business closed loop. In this case, the first intelligent agent can be a business intelligent agent for a specific business scenario, which can drive the implementation of step S11 based on business attributes. This ensures that the execution process generation and task step execution are implemented around the standard operating procedures and business plans of the specific business, rather than generalized task processing. Therefore, the execution process generated by the business intelligent agent for user request tasks is more in line with the actual business scenario of the enterprise, with stronger applicability and business value, thereby improving the reliability and accuracy of the execution results (output results) of the task steps. This application does not limit the construction method of business intelligent agents applicable to various business scenarios. The design parameters of the general intelligent agent can be adjusted based on the business data and operating procedures of the corresponding business scenario to obtain a business intelligent agent specifically for task processing in that business scenario, but it is not limited to this method of obtaining business intelligent agents.

[0072] Based on this, after receiving a task processing request from a user, the electronic device can first analyze the business category to which the requested task belongs, and then send the task processing request to the corresponding business agent. In this case, the business agent acts as the first agent to implement the task processing method proposed in this application. Optionally, if the user inputs the task processing request in the input interface of the business system or in the input box output by the business agent, the received task processing request can be directly input into the business agent (i.e., the first agent) of the business system to implement the method proposed in this application.

[0073] Step S12: In response to the fact that the task step currently being executed by the first intelligent agent is a step to be verified, the second intelligent agent performs personalized verification on the output result generated by the first intelligent agent in executing the step to be verified based on the user's personalized information, and generates a corresponding verification guidance instruction.

[0074] In this embodiment, a personalized verification agent, distinct from the first agent, is used as an independent second agent to perform real-time, interventional personalized verification of the steps to be verified executed by the first agent. In other words, a personalized verification mechanism is embedded in the reasoning of the first agent, with successful personalized verification serving as the trigger condition for the first agent to execute the next task step. The step to be verified can be at least one task step included in the execution flow, typically an indispensable key task step for achieving the corresponding task. This step can be determined by evaluating its importance; this application does not limit the method used to determine it.

[0075] In one possible implementation, during the generation of an execution flow containing multiple task steps, this application may assess the importance of each task step based on at least one of the following: context information of the session scenario in which the task processing request is located, the user's personalized information (implemented by a second intelligent agent in this case), and business rules of the business scenario to which the task processing request belongs. Specifically, it may assess the importance / impact of each task step on fulfilling the user's requested task. Based on this assessment, at least one of all task steps is selected as a step to be verified. After the first intelligent agent executes each step to be verified, the second intelligent agent performs personalized verification of the output result generated by that step. If the output result meets the user's personalized needs, the first intelligent agent then executes the next task step. The importance assessment of each task step can be implemented by the first or second intelligent agent using its own decision-making / deep thinking capabilities, or by calling a model, or by the first and second intelligent agents working together; this application does not impose any limitations on this.

[0076] In another possible implementation, this application may also allow the first intelligent agent to determine whether the current task step in the execution process is a step to be verified, either during the execution of the current task step or after obtaining the output result of the execution of the current task step, and before executing the next task step. If it is a step to be verified, its output result is then transmitted to the second intelligent agent for personalized verification, and the execution of the next task step using the output result is interrupted. Of course, the second intelligent agent may also monitor the processing of the first intelligent agent, and when it determines that the first intelligent agent has obtained the output result of the step to be verified, it actively reads the output result and controls the first intelligent agent to interrupt the execution of the next task step, thereby avoiding the use of output results that do not meet the user's personalized needs as input for the next task step, invalid execution of the next task step, and reduction of the success rate and accuracy of task step execution.

[0077] Furthermore, the process of detecting whether the task step currently being executed by the first agent belongs to the verification step can also be implemented by the second agent, avoiding increased resource consumption of the first agent. In this case, the second agent can monitor the execution progress of the first agent's execution flow. While the first agent is executing each task step, during execution, or after completion but before executing the next task step, the second agent uses its own capabilities or a calling model to determine whether the currently executing task step belongs to the verification step. If it does, the second agent actively reads the output result generated by the first agent executing that task step and controls the first agent to interrupt the execution flow; or it sends a personalized verification request for that task step to the first agent, which responds to the personalized verification request, confirming that the second agent is allowed to perform personalized verification of the output result of the current task step, and transmits the output result to the second agent. Thus, if the task step currently being executed by the first agent does not belong to the verification step, the second agent will not interfere with the first agent's execution flow, and the first agent can normally execute the next task step.

[0078] Therefore, this application allows the first intelligent agent to determine that the currently executing task step is a verification step, send the received output to the second intelligent agent, trigger the second intelligent agent to execute the personalized verification step, and control the execution flow to interrupt execution. Alternatively, the second intelligent agent can determine that the task step currently being executed or to be executed by the first intelligent agent is a verification step, actively read the output generated by the first intelligent agent executing the task step, and trigger the first intelligent agent to control the execution flow to interrupt execution, such as by sending an interrupt command for the execution flow to the first intelligent agent.

[0079] It should be noted that this application does not restrict the implementation method of how to detect whether a task step belongs to a verification step. In addition to the automatic detection method implemented by the intelligent agent described above, it can also be determined by manual intervention as needed. For example, during the review of the generated execution flow by the reviewer, the reviewer selects at least one task step as a step to be verified, configures a verification identifier for the task step, and when the first intelligent agent or the second intelligent agent recognizes that the currently executed task step is configured with a verification identifier, the first intelligent agent actively or passively transmits the currently obtained output result to the second intelligent agent, and the second intelligent agent performs a personalized verification step on the output result. This application does not restrict the implementation method of the personalized verification step.

[0080] The personalized verification process for the output of the step to be verified can be implemented by a second intelligent agent based on stored user personalized information. This personalized information refers to a set of information that is directly related to personalized verification rules, dimensions, and standards, and is user-specific or business scenario-specific. This ensures that the verification process accurately matches the user's business habits, compliance requirements, target preferences, format preferences, and resource permissions, ensuring that the verified output meets the user's expectations / personalized needs and is adapted to the business scenario. In this application, the user's personalized information can be determined based on at least one of the following: the user's historical behavior data, preference settings, feedback data on task processing results, and business rule constraints such as industry / enterprise-specific rules, compliance requirements, and operating procedures. In addition to user-level subjective preferences, it can also include objective rules at the business level. If any of the information contained therein changes, it can be dynamically updated based on the changed information.

[0081] Historical behavior data can originate from business systems within the corresponding business scenarios, the storage of electronic devices used by users, or other application platforms or device nodes managed by user identifiers. This application does not restrict its source or content. User preference settings can be completed in the personalized configuration interface of the business system, application, or operating system, such as preferences for business objects, business processing styles, etc. This application does not restrict the method or content of preference settings. Feedback data on task processing results can be the processing results of tasks that are the same as or similar to the task requested by the user, or feedback data on the final results obtained by the first intelligent agent directly completing each task step when the user requests a task for the first time. This includes input / expression / emotion indicating whether the user is satisfied with the task processing results, correction information for all or part of the task processing results that are not satisfactory, etc. Feedback data can be obtained through relevant function buttons, feedback interfaces, or image acquisition methods. This application does not restrict the representation format of feedback data. Therefore, personalized user information can be obtained through at least one of the following methods: proactive collection (such as the aforementioned preference settings or user feedback data), passive learning (such as extraction based on historical behavioral data), and synchronization with business systems (such as business rule constraint information obtained from business systems), to ensure that the information accurately matches user needs.

[0082] In practical applications, the user's personalized information can be directly stored by the second intelligent agent, or it can be converted into structured data and stored, or the structured data can be further broken down into executable verification conditions / rules, such as machine-executable instructions in the format of "field + operator + threshold" and then stored. Alternatively, a user profile (i.e., a personalized self-portrait) / user model can be built based on the user's personalized information, and corresponding verification logic can be configured in the second intelligent agent to enable the second intelligent agent to have personalized verification capabilities. In this way, after obtaining the output result to be verified, it can verify whether the output result conforms to the verification conditions / rules according to its own personalized verification capabilities and the stored or identified verification conditions / rules, that is, verify whether the execution of the current task step by the first intelligent agent conforms to the user's personalized information.

[0083] After obtaining the personalized verification result according to, but not limited to, the method described above, the second intelligent agent can generate corresponding verification guidance instructions. Specifically, if the personalized verification passes, and the matching degree between the output result and the user's personalized information is greater than a matching threshold (or the deviation is less than a deviation threshold), the generated verification guidance instruction can be a verification pass instruction / information. Conversely, if the personalized verification fails, and the matching degree is less than a matching threshold (or the deviation is greater than a deviation threshold), the generated verification guidance instruction can be a correction guidance instruction / information. This application does not limit the content of the verification guidance instructions.

[0084] Step S13: The second agent sends the verification guidance instruction back to the first agent;

[0085] In step S14, the first intelligent agent executes subsequent processing steps for the step to be verified based on the verification guidance instruction, so as to control the execution process to continue until a task processing result that conforms to the personalized information is obtained.

[0086] Following the above analysis, the second agent in this application intercepts the first agent's execution of the next task step based on the output result before the first agent executes / completes the step to be verified, or before executing the next task step of the step to be verified. Through the aforementioned personalized verification feedback verification guidance instruction, especially when the output result does not conform to the personalized information, the agent can promptly correct the step to be verified, ensuring that its output result conforms to the personalized information before the first agent executes the next task step. Therefore, the subsequent processing steps for the step to be verified in step S14 can be either re-executing the step to be verified or triggering the execution of the next task step. That is, if the verification guidance instruction indicates that the personalized verification has passed, the first agent continues to execute the next task step; if the verification guidance instruction indicates that the personalized verification has failed, the first agent re-executes the current task step (the step to be verified) based on the verification guidance instruction. It is evident that this application, during the dynamic execution of the task, utilizes a real-time monitoring and control mechanism based on the personalized information remembered by the second agent for verification and proactive correction, ensuring the degree of personalization of the step to be verified.

[0087] In this scenario, when the first agent invokes the first model to execute task steps, if the first agent receives a verification guidance instruction indicating that the personalized verification failed, it can input this instruction (such as correcting guidance information) as a prompt into the first model, guiding it to re-execute the current task step. For example, if the current task step is "analyze the highlight segments of the edited material," and the user is actually interested in a video clip of player A scoring a goal, but the first agent edits a video clip of player B assisting, this clearly does not match the user's preference. The second agent sends a verification guidance instruction to the first agent such as "edit the video clip of player A scoring a goal" or an instruction representing this semantic meaning. This utilizes the first model's existing video editing capabilities, guiding it to re-edit the video clip of player A scoring a goal from the video material as a highlight segment for output. This implementation does not require modification of the first model's original structure and model parameters.

[0088] In this application, the prompt information obtained when the verification fails varies depending on the different steps to be verified. That is, the prompt information is determined based on the verification result of the actual task steps executed. By dynamically determining the prompt information, the corresponding reasoning process of the first model can be intervened in real time and accurately, so that the model reasoning output results are stable and controllable. This solves the problem of the first model executing task steps based on static prompt words, improves the success rate of task execution and the accuracy of information, and helps to improve the overall processing efficiency of the task.

[0089] Therefore, the step-by-step preventive processing method proposed in this application can promptly detect and correct task processing deviations, prevent error accumulation, greatly improve task processing efficiency and resource utilization, and ensure that the final task processing result after the completion of the entire execution process meets the user's personalized needs. It eliminates the need to wait for the entire execution process to be completed before performing personalized checks on the final result, thereby avoiding the need for retrying the entire execution process or serious consequences due to errors in the final result, reducing rework costs. Especially in high-risk business scenarios such as finance, healthcare, and content generation, it significantly improves task execution efficiency, success rate, and personalization, thereby enhancing the user experience.

[0090] This application introduces an independent second intelligent agent to achieve personalized verification, without needing to reconstruct the core logic of the first intelligent agent, thus improving the compatibility of the system architecture. By having two intelligent agents focus on different processing operations of the task processing method proposed in this application—the first intelligent agent focusing on task execution capabilities and the second intelligent agent focusing on personalized verification capabilities—the accuracy and reliability of their respective processing results are improved, solving the problem of a single intelligent agent being unable to accommodate multiple targets. Furthermore, the embodiment of this application achieves the task processing method through the collaboration of two intelligent agents, not through simple sequential invocation. Instead, the two intelligent agents establish a dynamic, closed-loop feedback control mechanism through verification guidance information, embedding personalized verification into the verification steps in the execution flow and using it as the core architecture for execution flow control. This significantly improves the personalization compliance and first-time success rate of complex task processing, while saving computational resources.

[0091] Reference Figure 2 This is a schematic diagram of the signaling flow of the task processing method proposed in Embodiment 2 of this application. This embodiment describes one possible implementation method of how the second intelligent agent generates verification guidance instructions in the task processing method described above. The processing steps of the first intelligent agent can be referred to the description in the corresponding part of the context, and will not be detailed in this embodiment. Based on this, as... Figure 2 As shown, the method for the second intelligent agent to generate verification guidance instructions may include:

[0092] Step S21: Based on the attributes of the step to be verified currently being executed by the first agent, determine the target verification logic among multiple candidate verification logics;

[0093] In this embodiment, the step to be verified is a task step in the execution flow. Its attributes may include at least one of the following: task step type, output result attributes, and business context. This directly determines which verification logic to use, enabling personalized verification of the output result of the task step. Alternatively, through a selector or a mapping relationship between pre-configured attributes (such as task step type) and verification logic, the most suitable candidate verification logic can be dynamically determined from multiple pre-configured candidate verification logics as the target verification logic for the current step to be verified. For example, for the task step "video filtering", the target verification logic selected is different depending on whether the video type is "sports" or "food". The personalized verification of the two task steps "analyzing video style" and "video synthesis" uses different target verification logics, which helps to improve the accuracy and reliability of personalized verification results.

[0094] The aforementioned candidate verification logic may include, but is not limited to, at least one of the following: matching score (preference weight calculation) logic (used to calculate the matching score between the output result and the personalized information), classification logic (used to classify the output result as "compliant with personalized information" or "incompatible with personalized information"), and rule verification logic (used to verify whether the output result meets the specific rules set by the user in the personalized information). This application does not limit the type of verification logic or its configuration implementation method; it can be dynamically updated according to the verification requirements of the business scenario.

[0095] Subsequently, the second intelligent agent can, based on the user's personalized information, perform personalized verification on the output results generated by the first intelligent agent in executing the verification steps through target verification logic, and generate corresponding verification guidance instructions. As described above, different candidate verification logics select different target verification logics, and there may be multiple execution actions to implement personalized verification here. This application will not go into detail about each one, but will only illustrate one optional verification process described in steps S22-S24 below, but is not limited to this.

[0096] Step S22: Determine the object to be verified included in the output result of the first agent performing the step to be verified;

[0097] The object to be verified refers to an object that needs to conform to the user's personalized information or reflect personalized differences. It is usually a specific element / unit in the output result that can be quantified and verified, with clear verification dimensions and measurable attributes. As a bridge connecting the user's personalized information and the target verification logic, it is strongly related to the user's personalized information. Depending on the type of output result, the object to be verified can be at least one of the following categories: data (such as quantifiable data such as numerical values ​​and text), structure (such as structured features such as format, hierarchy, arrangement or style), status (such as business status, review results), and rule (such as business rule constraints that need to be met). In different business scenarios, the output result of the same type of task step may contain different objects to be verified, and there is no restriction on this.

[0098] Therefore, during the implementation of step S22, the second intelligent agent can determine at least one object to be verified in the current output result based on the user's personalized information. For example, it can pre-configure a list of objects to be verified based on the user's personalized information, compare the output result with this list, and determine the objects to be verified contained in the output result; or, it can infer and analyze the objects to be verified contained in the output result based on the user's personalized information, etc. It can also be determined according to the actual business scenario as needed.

[0099] For example, the output of the verification step "Analyze the editing style of the reference video": The first agent analyzes the editing style of the reference video (such as fast forward, hard cut, filter usage, etc., at least one of them), and extracts key style parameters (such as pacing, transition type, filter preference, etc., which are consistent with the editing style). The objects to be verified can include the editing style and style parameters. The output of the verification step "Analyze the highlight segments of the edited material": The first agent extracts highlight segments from the original material (recording video clips of types such as goals, assists, exciting matches, etc.). The objects to be verified can include the type of highlight segment, the focus of the content, and the featured player.

[0100] Step S23: Through the target verification logic, evaluate the degree of conformity of the object to be verified relative to the user's personalized information;

[0101] Step S24: Based on the degree of compliance, generate verification guidance instructions for the currently executed task steps.

[0102] As analyzed above, the representation of the degree of conformity obtained from different target verification logics can vary, such as matching score, classification result instruction, or result instruction on whether it conforms to the rules. Since this application can determine the target verification logic based on the attributes of the output result, it includes selecting the target verification logic based on the type of the object to be verified, in order to evaluate the degree of conformity of different content. For example, if the object to be verified is editing style and style parameters, the matching score logic can be used to calculate the similarity between the style of the generated content and the user's preferred style to obtain the corresponding matching score as the degree of conformity. This application does not restrict the implementation method of this similarity calculation. If the object to be verified is a star player and a focus object in an edited video clip, the rule verification logic can be used to verify whether the user's list of favorite star players includes the main character (i.e., the focus object) in the generated content, to obtain two conclusive degrees of conformity (confidence level of conformity to the user's personalized needs), namely "included" or "not included."

[0103] Specifically, for the degree of conformity, such as matching score / weight, it can be compared with the corresponding threshold to determine whether the personalized verification of the output result has passed or failed. If the degree of conformity is greater than the threshold, the personalized verification is determined to have passed, and a verification pass instruction / execution process recovery instruction is generated as a verification guidance instruction to trigger the first intelligent agent to continue to execute the next task step. If the degree of conformity is less than the threshold, the personalized verification is determined to have failed. At this time, based on the target object information corresponding to the object to be verified in the personalized information, correction guidance information for the current task step can be generated as a verification guidance instruction and fed back to the first intelligent agent. The first intelligent agent is then triggered to re-execute the task step based on the correction guidance information to obtain an output result that conforms to the user's personalized information, and then execute the next task step.

[0104] In one possible implementation, after the first agent re-executes the task steps, the output result can still be further personalized and verified by the second agent. Only after confirming that it conforms to the personalized information is the first agent triggered to execute the next task step, thereby improving the reliability of task step correction. Of course, to save resources and improve task processing efficiency, since the output result generated by the first agent re-executing the task step based on the corrected guidance information usually conforms to the user's personalized information, further personalized verification by the second agent is not required. The next task step can be executed directly based on the re-obtained output result, without any restrictions.

[0105] Reference Figure 3 This is a schematic diagram of the signaling flow of the task processing method proposed in Embodiment 2 of this application. This embodiment may be a possible optimized implementation of the task processing method described above, such as... Figure 3 As shown, the optimized implementation method proposed in this embodiment may include:

[0106] Step S31: The first intelligent agent responds to the user's input task processing request and generates an initial execution flow containing multiple task steps.

[0107] Step S32: The second intelligent agent calibrates the initial execution process based on the user's personalized information and generates an execution process that conforms to the personalized information, which is then fed back to the first intelligent agent.

[0108] In this embodiment, during the task planning process for a user-input task processing request, if it is determined that the second intelligent agent has obtained and stored the user's personalized information, the first and second intelligent agents can jointly perform task planning, that is, jointly formulate an execution plan to implement the user's requested task, so as to obtain an execution flow that conforms to the user's personalized information. In other words, this execution flow is collaboratively generated by the first and second intelligent agents based on the user's personalized information (memory). This collaborative generation process can adopt, but is not limited to, the implementation methods described in steps S31 and S32, so that the execution direction of each task step included in the execution flow conforms to the user's personalized information. For example, in a video editing task, the task planning process can explicitly focus on the user's favorite sports star (such as star C) and preferred editing style (such as fast forward, hard cut, etc.).

[0109] In one possible implementation, the calibration of the initial execution flow by the second intelligent agent can include, but is not limited to, adjusting the priority of task steps, adding or deleting specific task steps, or injecting personalized parameters into task steps. During the calibration process, the second intelligent agent uses the user's personalized information as the core, the task objective as the guide, and the constraints of the actual business scenario as the basis to achieve dynamic optimization of the initial execution flow. This makes the calibrated execution flow more closely aligned with the actual business execution scenario (such as resource constraints, real-time data changes, etc.) and the user's personalized information (such as preferences, compliance requirements, output format, and other personalized needs).

[0110] Among the calibration methods, the priority adjustment method for task steps can be based on information such as resource allocation, time constraints, and business urgency in the actual business scenario. This involves reordering the execution order of task steps, such as changing sequential task steps (e.g., those with resource conflicts) to parallel execution, or adjusting the execution order of task steps and prioritizing tasks with high urgency, thereby improving process efficiency and ensuring the achievement of key objectives. The addition or deletion of specific task steps can be based on changes in requirements, constraints, and personalized requirements in the actual business scenario, adding necessary steps or deleting redundant steps to ensure the integrity and simplicity of the process. The injection of personalized parameters into task steps can inject appropriate parameters based on the task step type, such as at least one of quantitative constraint parameters, qualitative requirement parameters, output format parameters, and resource allocation parameters.

[0111] Therefore, the second intelligent agent, in addition to basing its calibration on user-specific information, can also dynamically select appropriate calibration methods based on multiple dimensions, including constraints of actual business scenarios, real-time data feedback, and historical calibration data. This ensures that the initial execution process is calibrated reasonably and meets the user's personalized needs. The selection of the calibration method can be achieved through the first model or the third model, and can also be combined with a rule engine as needed; this application does not impose any restrictions on this. Optionally, after completing the calibration, the second intelligent agent can also verify the rationality and feasibility of the obtained execution process, such as sending it to the first intelligent agent for verification or providing feedback to reviewers for verification; this is also not restricted.

[0112] Step S33: The first intelligent agent executes the task steps sequentially based on the received execution flow and obtains the corresponding output results;

[0113] Step S34: The first agent determines whether the currently executed task step is a step to be verified. If not, proceed to step S35; if yes, proceed to step S36.

[0114] Step S35: The first agent executes the next task step based on the output result;

[0115] Step S36: The first agent sends the output result to the second agent;

[0116] It should be noted that the execution process of the first intelligent agent after receiving the execution flow that conforms to the personalized information can refer to the task processing process described in the above method embodiment, including but not limited to the processing method described in this embodiment, such as the judgment process of the above-mentioned verification step, which can also be implemented by the second intelligent agent. The implementation process will not be described in detail in this embodiment.

[0117] Step S37: The second intelligent agent performs personalized verification of the output result based on the user's personalized information and generates corresponding verification guidance instructions to be fed back to the first intelligent agent.

[0118] Step S38: The first intelligent agent executes subsequent processing steps for the step to be verified based on the verification guidance instruction, so as to control the execution flow to continue.

[0119] In step S39, the first intelligent agent determines that all task steps in the execution process have been completed and outputs the final task processing result that conforms to the personalized information.

[0120] In this process, the first intelligent agent determines, based on the verification guidance instruction, that the current output result matches the user's personalized information. If the output result is consistent with the user's personalized preferences, it can continue to execute the next task step, obtain its output result, and return to step S34 to continue execution. Alternatively, if it determines that the current output result does not match the user's personalized information, such as the output result containing a low-confidence point that conflicts with the user's personalized preferences (which belongs to a type of object to be verified as described above), then the verification guidance instruction is to correct the guidance information. For example, in the video editing scenario described above, the second intelligent agent provides verification guidance instructions such as "Please re-filter the highlight clips containing star player C" or "Please adjust the editing rhythm to fast forward." The first intelligent agent will guide the first model to re-execute the current verification step based on this, obtain a new output result, execute the next task step based on this, obtain its output result, and return to step S34 to continue execution, or send the new output result to the second intelligent agent and return to step S37 to continue execution.

[0121] This approach accurately identifies and corrects outputs that conflict with user preferences across multiple task steps, ensuring that the final task processing result meets both business objectives and user preferences. This improves the success rate and accuracy of task execution, requires minimal user intervention, and enhances the user experience.

[0122] In the task processing methods described in the embodiments above, if the user does not directly express the various task steps required for task execution in the input task processing request, the first intelligent agent can directly generate, or collaboratively generate with the second intelligent agent, an execution flow containing multiple task steps by understanding the intent contained in the task processing request. That is, the execution flow or initial execution flow in the above embodiments (corresponding to the embodiments collaboratively generated by two intelligent agents) is generated by the first intelligent agent's understanding of the intent of the task processing request. Therefore, referring to... Figure 4 The signaling flow diagram of the task processing method proposed in Embodiment 4 of this application is shown. The generation process of the execution flow described in the above embodiments, such as the generation process by the first intelligent agent itself or the generation process in collaboration with the second intelligent agent, may include, but is not limited to:

[0123] Step S41: The first intelligent agent receives a task processing request input by the user;

[0124] Step S42: The first agent performs intent understanding on the task processing request through the first model and sends the obtained intent understanding result to the second agent.

[0125] In this application, the first model can be a general AI (Artificial Intelligence) model, which may employ, but is not limited to, the Transformer or its architectural variants (such as using only an encoder-only / decoder-only architecture, an encoder-decoder architecture, or MoE (Mixture of Experts, a neural network architecture), etc.) or other infrastructures. It learns the features and patterns of natural language by training on large amounts of diverse data, thereby enabling it to understand and generate natural language. Typically, it has hundreds of millions to trillions of model parameters (model parameters are variables that control the behavior of the target model) and is capable of capturing complex relationships and patterns in natural language. In this case, the first model can be a primary task processing model deployed in the cloud.

[0126] The AI ​​models may include, but are not limited to, generative models and generative language models (GLMs). For example, one or more of the following: large language model (LLM), GPT (Generative Pre-trained Transformer) series models, T5 (Text to Text Transfer Transformer) models, large visual models, and multimodal large models.

[0127] Depending on actual needs, the first model in this application embodiment can also be an expert large model fine-tuned from business data in actual business scenarios, making it applicable to specific business scenarios and improving its task processing accuracy in those scenarios. Furthermore, the first model can also be a lightweight model (which can be referred to as the second task processing model, with a much smaller number of model parameters than the first task processing model) compressed from the general AI model or expert large model using lightweight methods such as quantization, knowledge distillation, or pruning. This lightweight model, meeting the deployment requirements of edge devices with limited computing resources, can then be deployed on local electronic devices. The model type of the first model in this application is not limited and can be determined based on actual application needs.

[0128] Therefore, it can be seen that in the implementation process of step S42, referring to Figure 5 The flowchart shown illustrates that, when the local electronic device has sufficient resources, the first intelligent agent can perform task planning through the second task processing model (local expert model) to generate an execution flow containing multiple task steps. When the local electronic device has insufficient resources, the first intelligent agent can perform task planning through a cloud model (which could be the first task processing model), generate an execution flow, and then feed the execution flow back to the electronic device. The first intelligent agent can then execute the task steps through the local expert model, or it can continue to have the cloud model execute the steps; there are no restrictions on this. The generation of the execution flow (including the collaborative generation process with the second intelligent agent) and the execution process of the task steps can be found in the descriptions of the corresponding sections of the above embodiments.

[0129] Step S43: The second intelligent agent performs personalized verification of the intent understanding result based on the user's personalized information through the second model, and generates corresponding verification guidance instructions to be fed back to the first intelligent agent.

[0130] The personalized verification and correction process of the second intelligent agent's understanding of intent is similar to the personalized verification and correction process of the output results of the task steps described in the previous embodiment, and this embodiment will not describe the process in detail.

[0131] In this application, the second model can be a personalized model trained based on the user's personalized information, such as a similarity calculation model, classification model, regression model, deep neural network model, or other machine learning model. This personalized information can be deployed on the user's electronic device. The training process of the personalized model can include: acquiring a training sample set, where each training sample includes the input of a sample task step, the sample output result of the corresponding sample task step, and the user's feedback label for the sample output result; then, using the sample task step input and sample output result as sample input features, and the corresponding feedback label as the training target, training the initial model to obtain the personalized model.

[0132] Optionally, this application may also record the output results of the task steps generated by the first intelligent agent and the verification guidance instructions generated by the second intelligent agent; use the output results of the task steps as new sample input features, and use the verification guidance instructions or user feedback data on the final task processing results as supervision signals to incrementally update the personalized model, so as to quickly realize the dynamic update of the personalized model and improve the reliability of subsequent task processing, but is not limited thereto.

[0133] In some embodiments, to enhance the personalized verification capability of the second intelligent agent, the second model it invokes may include the aforementioned personalized model and a data processing model deployed in the cloud. This data processing model can be the general AI model or expert large model described above. In this case, the second intelligent agent can leverage the powerful data processing capabilities of the cloud model and the security of the local personalized model to implement the personalized verification process described above. Thus, only relevant task data during task processing, such as output results, needs to be sent to the cloud; personalized information of the user that requires protection will not be sent to the cloud, thereby avoiding the risk of leakage of user personal information and improving data security.

[0134] Based on this, the second intelligent agent can send the output of the step to be verified to the cloud, process the output through a data processing model, and feed back the feature data of the generated output (such as feature vectors obtained from feature extraction of the output, which can be directly used for further personalized verification of the personalized model) to the second intelligent agent. The second intelligent agent inputs the feature data into the personalized model and performs personalized verification based on the stored user's personalized information. The implementation process can be referred to the description in the corresponding part of the above embodiment. Therefore, in this situation... Figure 5 The cloud model in the model can be a data processing model, the cloud processing flow is the process of obtaining feature data, and the correction task steps are implemented based on the local personalized model.

[0135] Therefore, compared to implementing the task processing method proposed in this application through a cloud model, which involves calling stored personalized information and other memories to improve the contextual understanding and personalized response capabilities of the cloud model, task processing using memories in cloud models like Manus is more about reusing task flows and generating automated scripts. It is weak in personalized verification and dynamic correction, and struggles to accurately identify and correct step-level deviations. Cloud models like Lovart mainly rely on static user preference tags and lack real-time verification mechanisms during dynamic reasoning, making it difficult to reliably resolve step-level deviations in complex tasks. Cloud models like GPT primarily use memory retrieval to maintain dialogue coherence and contextual understanding. In actual task execution, they lack fine-grained verification and proactive correction capabilities for user personalized information, especially prone to deviation accumulation in multi-step tasks, reducing the success rate and reliability of task execution.

[0136] To address this, this application introduces a second intelligent agent during the task execution process of the first intelligent agent. Through the aforementioned step-level verification process, the second agent provides step-level verification guidance instructions to the first intelligent agent, proactively and in real-time correcting task steps that do not meet the user's personalized needs. This suppresses the illusion problem of cloud models at the system architecture level, avoiding the failure of the final task processing result due to the accumulation of deviations in complex tasks, and significantly improving the success rate and accuracy of task execution. Optionally, this application can also externalize the personalized verification results in the form of the degree of conformity between the output results and personalized information (confidence level), and request clarification from the user when necessary, making the entire task processing process more controllable and transparent.

[0137] In step S44, the first intelligent agent inputs the verification guidance instruction as prompt information into the first model, guiding the first model to re-understand the intent of the task processing request and obtain an intent understanding result that conforms to the personalized information.

[0138] In step S45, the first agent generates an execution flow containing multiple task steps based on the re-obtained intent understanding results using the first model.

[0139] In summary, this application introduces an independent second model to invoke a personalized model. Based on memory, it proactively verifies the output results at the step level (i.e., the intermediate output results of the entire task) in real time (i.e., the personalized verification process described above). This focuses on user personalized preferences rather than general facts or objective logic, solving the problem in traditional task processing methods based on large models that cannot identify conflicts between intermediate results and user preferences during the inference process. During the personalized verification process, the output results of each step to be verified described above can be verified in real time. This includes proactively intervening in key aspects of the entire task execution process, or personally verifying the output results of each task step included in the execution flow. This allows for timely identification of low-confidence points in the output results that conflict with user personalized preferences, avoiding the direct delivery of erroneous results to the user and reducing rework costs.

[0140] The second intelligent agent not only relies on static user preference tags (such as pre-stored personalized information like the aforementioned preference settings), but also continuously optimizes its understanding of user personalized information through a dynamic verification process, forming a closed loop of "memory-verification-correction" to enhance long-term personalized adaptation capabilities. Furthermore, the task processing method proposed in this application is not only applicable to video editing tasks as exemplified above, but also to other content creation tasks such as copywriting generation, tool calls that recommend parameters matching user style, decision support for investment advice that aligns with user risk preferences, and business scenarios that require multiple steps and cater to user subjective preferences (fine-grained personalized adaptation). Therefore, it has broad applicability and promotional value.

[0141] In conjunction with the task processing methods executed by the electronic devices described in the above embodiments, in order to clarify the task processing methods executed by each intelligent agent (first intelligent agent or second intelligent agent), the task processing methods executed by each intelligent agent will be described below from the perspective of each intelligent agent.

[0142] Reference Figure 6 This is a flowchart illustrating the task processing method proposed in Embodiment Six of this application. The method proposed in this embodiment can be executed by a first intelligent agent in an electronic device, and may include, but is not limited to, the following steps:

[0143] Step S61: In response to the user's input task processing request, generate an execution flow containing multiple task steps;

[0144] Step S62: Execute the task steps and generate the corresponding output results;

[0145] Step S63: If the task step is identified as a step to be verified, receive a verification guidance instruction for the output result of the step to be verified; the verification guidance instruction is generated based on the user's personalized information verification.

[0146] Step S64: Based on the verification guidance instruction, execute subsequent processing steps for the step to be verified to control the execution process to continue until a task processing result that matches the personalized information is obtained.

[0147] As described in the above embodiments, the task steps identified as the implementation process of the steps to be verified can be completed by the first intelligent agent or the second intelligent agent, or specified by the user, etc., without limitation. Regarding the generation of verification guidance instructions, they can be implemented by the second intelligent agent according to the method described above, or by the first intelligent agent calling a personalized model, etc., without limitation in this application.

[0148] Therefore, it can be seen that during the process of the first intelligent agent executing each task step in the generated execution flow according to the task rules, or after obtaining its output results, the method described above can be directly adopted, but not limited to, to verify whether the output results meet the user's personalized needs based on a dynamic personalized verification mechanism. If they do, the next task step will be executed. If they do not, the current task step will be modified to make its output results meet the user's personalized needs before executing the next task step. This achieves fine-grained (task step level) verification, avoids error accumulation, and ensures that the final task processing result meets the user's personalized needs. Moreover, this closed-loop processing method of "memory-verification-correction" improves the long-term personalized adaptation capability.

[0149] Reference Figure 7 This is a flowchart illustrating the task processing method proposed in Embodiment 7 of this application. The method proposed in this embodiment can be executed by a second intelligent agent in an electronic device, and may include, but is not limited to, the following steps:

[0150] Step S71: In response to the execution of the step to be verified, obtain the output result generated by the execution of the step to be verified; the step to be verified is a task step included in the execution flow generated based on the user's input task processing request.

[0151] Step S72: Based on the user's personalized information, perform personalized verification on the output results and generate corresponding verification guidance instructions;

[0152] Step S73: Output a verification guidance instruction, which is used to control subsequent processing steps for the step to be verified.

[0153] In this embodiment, the second intelligent agent can receive a personalized verification instruction for the currently executed task step, parse the personalized verification instruction, obtain the output result to be verified, and implement step-level personalized verification according to, but not limited to, the dynamic personalized verification mechanism described above. The verification guidance instruction generated therefrom is used as the trigger condition for subsequent processing steps of the task step to determine whether to correct the task step or execute the next step, so as to ensure that the final result meets the user's personalized needs. There is no need to repeat the entire task process, which improves the success rate and personalization of task execution.

[0154] Based on the task processing methods described in the above embodiments, this application will use a video editing business scenario as an example for illustration. Assume the first intelligent agent is a business agent designed for this business scenario, and the second intelligent agent is a personalized agent that stores user personalized information (which can be stored in a knowledge base), such as... Figure 8 The flowchart shown illustrates that after a user initiates a task processing request to the electronic device's system to "edit match highlights," in order to formulate a suitable task flow, the business agent first performs intent understanding on the user's task request to determine the intent understanding results such as the theme, content scope, and target of the video the user wants to edit, such as a football match highlights compilation. At this point, the personalized agent first verifies whether the intent understanding results meet the user's personalized needs, such as checking whether it focuses on the user's preferred star player or content type. If a deviation is found, a correction process is triggered, such as providing correction guidance information (i.e., a verification guidance instruction) to the business agent, such as asking to focus on star player C (who is the user's preferred star player), to request the business agent to readjust the intent understanding direction to obtain an intent understanding result that focuses on star player C (such as the task intent of editing a match highlights compilation of star player C).

[0155] Subsequently, the business agent performs task planning based on the intent understanding results that conform to the user's personalized needs, in order to generate tasks including, Figure 8 The execution flow of the three task steps shown is as follows: If all three task steps are to be verified, after executing task step 1 and obtaining the editing style of the reference video, the personalized agent verifies the style, that is, whether it matches the user's personalized preferences. For example, it checks whether the editing rhythm preferred by the user is used, and whether the transition type is consistent with the user's preferences. If a style mismatch is found, such as the reference video using a slow-paced lyrical editing style, while the user prefers a fast-paced and energetic style, a verification guidance instruction such as "Please use a fast-paced style" can be provided to trigger the business agent to adjust the style parameters of the reference video editing.

[0156] Similarly, after the business agent executes task step 2, the personalization agent can verify whether the extracted highlight clips meet the user's personalized needs. This includes checking if the extracted highlight clips contain the user's preferred sports star, and whether the type of highlight clip matches the user's preferred content focus. If the highlight clips are found to be inconsistent with user preferences, such as not containing player C or containing too many defensive clips, a correction process is triggered, such as requesting the addition of a shot of player C in task step 2. Finally, after the business agent executes task step 3 to obtain the synthesized video, it verifies the video and ultimately delivers the highest quality synthesized video to the user. However, this method is not limited to this one. It should be noted that the task processing method proposed in this application is also applicable to other business scenarios, including but not limited to video editing. The task processing method proposed in this application can still be reused in other business scenarios to achieve efficient, personalized, reliable, and low-cost task processing.

[0157] Reference Figure 9 This is a schematic diagram of the task processing device proposed in Embodiment 1 of this application. The task processing device proposed in this embodiment can be applied to electronic devices, such as... Figure 9 As shown, the task processing device proposed in this embodiment may include:

[0158] The first task processing module 91 is used to generate and execute an execution flow containing multiple task steps in response to a task processing request input by the user through the first intelligent agent.

[0159] The personalized verification module 92 is used to respond to the first intelligent agent's current task step as a step to be verified, and based on the user's personalized information, to perform personalized verification on the output result generated by the first intelligent agent in executing the step to be verified, and generate corresponding verification guidance instructions to be fed back to the first intelligent agent.

[0160] The second task processing module 93 is used to execute subsequent processing steps for the step to be verified based on the verification guidance instructions by the first intelligent agent, so as to control the execution process to continue until the task processing result that conforms to the personalized information is obtained.

[0161] Optionally, the personalized verification module 92 may include any of the following units:

[0162] The first triggering unit is used for the first intelligent agent to continue executing the next task step in response to the verification guidance instruction indicating that the personalized verification has passed.

[0163] The first input unit is used by the first intelligent agent to input the verification guidance instruction as a prompt message into the first model in response to the verification guidance instruction indicating that the personalized verification has failed, so as to guide the first model to re-execute the current task steps.

[0164] Optionally, the personalized verification module 92 may also include:

[0165] The target verification logic determination unit is used to determine the target verification logic among multiple candidate verification logics based on the attributes of the step to be verified.

[0166] The first verification unit is used to perform personalized verification on the output results of the first intelligent agent executing the steps to be verified based on the user's personalized information and through target verification logic, and generate corresponding verification guidance instructions.

[0167] Optionally, the first verification unit may include:

[0168] The object to be verified determination unit is used to determine the object to be verified contained in the output result of the first intelligent agent performing the step to be verified;

[0169] The evaluation unit is used to evaluate the degree to which the object to be verified conforms to the user's personalized information through the target verification logic;

[0170] The first generation unit is used to generate verification guidance instructions for the currently executed task steps based on the degree of compliance.

[0171] In some embodiments, the above execution flow is generated based on the first intelligent agent's understanding of the intent of the task processing request, and the task processing device may further include:

[0172] The second generation unit is used to perform personalized verification of the intent understanding results generated by the first intelligent agent based on the user's personalized information before generating the execution process, and generate corresponding verification guidance instructions to be fed back to the first intelligent agent.

[0173] The second input unit is used by the first intelligent agent to input the verification guidance instruction as prompt information into the first model, guiding the first model to re-understand the intent of the task processing request, and generate an execution flow based on the intent understanding result that conforms to the personalized information.

[0174] Optionally, the above execution process can also be collaboratively generated by the first intelligent agent and the second intelligent agent based on the user's personalized information. The collaborative generation module used to implement this collaborative generation process may include:

[0175] The third generation unit is used by the first intelligent agent to generate an initial execution flow containing multiple task steps in response to a task processing request input by the user.

[0176] The fourth generation unit is used by the second intelligent agent to calibrate the initial execution process based on the user's personalized information, and generate an execution process that conforms to the personalized information and feeds it back to the first intelligent agent.

[0177] The personalized verification process described above is implemented by the second intelligent agent through the second model. Therefore, the personalized verification module 92 may also include:

[0178] The first processing unit is used to send the output results to the cloud by the second intelligent agent, process the output results through the data processing model, and feed back the feature data of the generated output results to the second intelligent agent.

[0179] The second verification unit is used by the second intelligent agent to input feature data into the personalized model and perform personalized verification of the feature data based on the stored personalized information of the user.

[0180] The second model can be a personalized model deployed on an electronic device used by the user, or it can include a personalized model and a data processing model deployed in the cloud.

[0181] Reference Figure 10 This is a schematic diagram of the task processing device proposed in Embodiment 2 of this application. The task processing device proposed in this embodiment can be executed by a first intelligent agent in an electronic device, such as... Figure 10 As shown, the task processing device executed by the first intelligent agent may include:

[0182] The execution flow generation module 101 is used to generate an execution flow containing multiple task steps in response to a user-inputted task processing request.

[0183] The task step execution module 102 is used to execute task steps and generate corresponding output results.

[0184] The verification guidance instruction receiving module 103 is used to receive a verification guidance instruction for the output result of the step to be verified if the task step is identified as a step to be verified; the verification guidance instruction is generated based on the user's personalized information.

[0185] The execution flow control module 104 is used to execute subsequent processing steps for the step to be verified based on the verification guidance instructions, so as to control the execution flow to continue until the task processing result that matches the personalized information is obtained.

[0186] Reference Figure 11 This is a schematic diagram of the task processing device proposed in Embodiment 3 of this application. The task processing device proposed in this embodiment can be executed by a second intelligent agent in an electronic device, such as... Figure 11 As shown, the task processing device executed by the second intelligent agent may include:

[0187] The output result acquisition module 111 is used to obtain the output result generated by the execution of the step to be verified in response to the execution of the step to be verified; the step to be verified is the task step contained in the execution flow generated based on the user input task processing request.

[0188] The verification guidance instruction generation module 112 is used to perform personalized verification of the output results based on the user's personalized information and generate corresponding verification guidance instructions.

[0189] The verification guidance instruction output module 113 is used to output verification guidance instructions; wherein, the verification guidance instructions are used to control the subsequent processing steps for the step to be verified.

[0190] This application also provides a computer program product including computer-readable instructions. When the computer-readable instructions are executed on a computer device, the computer device implements any of the task processing methods provided in this application. The computer program product can be stored in a readable storage medium, such as a computer floppy disk, USB flash drive, external hard drive, ROM (Read-Only Memory), RAM (Random Access Memory), magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training device, or network device, etc.) to execute the task processing methods described in the various embodiments of this application.

[0191] This application also provides a computer-readable storage medium carrying one or more computer programs. When these programs are executed by a computer device, the computer device can implement any of the task processing methods provided in this application. This application does not limit the product form of the computer-readable storage medium.

[0192] Reference Figure 12 This is a schematic diagram of the hardware structure of an electronic device proposed in an embodiment of this application. The electronic device may include at least one memory 121, a computer program 122 stored in the memory 121, and at least one processor 124 capable of running an intelligent agent 123.

[0193] The intelligent agent can execute a computer program through the processor 124 to implement the task processing method described in the corresponding embodiments of this application. The intelligent agent may include a first intelligent agent and a second intelligent agent, or act as the first intelligent agent or the second intelligent agent, to implement the steps of the task processing method described from the corresponding side in the above embodiments. The implementation process will not be described in detail in this application.

[0194] For example, processor 124 can be configured to implement a first intelligent agent and a second intelligent agent. The first intelligent agent is configured to: generate an execution flow containing multiple task steps in response to a user-input task processing request; execute the task steps and generate corresponding output results; receive a verification guidance instruction for the output results of a task step identified as a step to be verified; and, based on the verification guidance instruction, execute subsequent processing steps for the step to be verified to control the execution flow to continue until a task processing result conforming to personalized information is obtained. The second intelligent agent is configured to: respond to the first intelligent agent's currently executed task step being a step to be verified; perform personalized verification on the output results generated by the first intelligent agent executing the step to be verified based on personalized information; generate a verification guidance instruction; and send the verification guidance instruction to the first intelligent agent. The processing steps of each intelligent agent can be referred to the description in the corresponding part of the above embodiments.

[0195] Furthermore, the electronic device may also include at least one input component 125, such as an audio acquisition device, stylus, mouse and keyboard, or joystick, for the user to operate the input component 125 to input task processing requests to the electronic device. This application does not describe in detail the type of input component 125 or its input principle. The electronic device may also include at least one output component 126, such as an audio player or display screen, for outputting task processing results, such as visual charts or reports, etc., without limitation.

[0196] It should be understood that, Figure 12 The structure of the electronic device shown does not constitute a limitation on the computer device in the embodiments of this application. In practical applications, the computer device may include more than Figure 12 The application does not provide detailed examples of the more or fewer components shown, or combinations of certain components, such as gyroscopes, accelerometers, and gravity sensors used to obtain sensing parameters, power management modules, antennas, or other communication elements.

[0197] Finally, it should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0198] In the above embodiments, the invention can be implemented entirely or partially by software, hardware, firmware, or any combination thereof. Through the description of the above embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware, or it can be implemented using dedicated hardware including dedicated integrated circuits, dedicated CPUs, dedicated memory, dedicated components, etc. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The various embodiments in this specification are described in a progressive or parallel manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to mutually. For the apparatus and computer equipment disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

Claims

1. A task processing method, the method comprising: The first intelligent agent responds to the user's input task processing request, generates and executes an execution flow that includes multiple task steps; The second agent responds to the fact that the task step currently being executed by the first agent is a step to be verified. Based on the user's personalized information, the second agent performs personalized verification on the output result generated by the first agent in executing the step to be verified, and generates a corresponding verification guidance instruction to be fed back to the first agent. Based on the verification guidance instruction, the first intelligent agent executes subsequent processing steps for the step to be verified, so as to control the execution process to continue until a task processing result that conforms to the personalized information is obtained.

2. The method according to claim 1, wherein the first intelligent agent, based on the verification guidance instruction, executes subsequent processing steps for the step to be verified, including any one of the following: The first intelligent agent, in response to the verification guidance instruction indicating that the personalized verification has passed, continues to execute the next task step; In response to the verification guidance instruction indicating that the personalization verification failed, the first intelligent agent inputs the verification guidance instruction as a prompt message into the first model, guiding the first model to re-execute the current task steps.

3. The method according to claim 1, wherein the step of performing personalized verification on the output result generated by the first intelligent agent performing the verification step based on the user's personalized information, and generating corresponding verification guidance instructions, includes: Based on the attributes of the step to be verified, the target verification logic is determined among multiple candidate verification logics. Based on the user's personalized information, the target verification logic is used to perform personalized verification on the output results generated by the first intelligent agent in performing the verification step, and generate corresponding verification guidance instructions.

4. The method according to claim 3, wherein the step of performing personalized verification on the output result of the first intelligent agent executing the step to be verified based on the user's personalized information and through the target verification logic, and generating a corresponding verification guidance instruction, includes: Determine the object to be verified included in the output result of the first intelligent agent executing the step to be verified; The target verification logic is used to evaluate the degree of conformity of the object to be verified relative to the user's personalized information. Based on the degree of compliance, a verification guidance instruction is generated for the currently executed task step.

5. The method according to any one of claims 1-3, wherein the execution flow is generated based on the first intelligent agent's understanding of the intent of the task processing request, and further comprises, before generating the execution flow: The second intelligent agent performs personalized verification on the intent understanding results generated by the first intelligent agent based on the user's personalized information, and generates corresponding verification guidance instructions to be fed back to the first intelligent agent. The first intelligent agent inputs the verification guidance instruction as a prompt into the first model, guiding the first model to re-understand the intent of the task processing request, and generates the execution flow based on the intent understanding result that conforms to the personalized information.

6. The method according to any one of claims 1-3, wherein the execution flow is collaboratively generated by the first intelligent agent and the second intelligent agent based on the user's personalized information, and the collaborative generation process includes: The first intelligent agent responds to the user's input task processing request and generates an initial execution flow containing multiple task steps; The second intelligent agent calibrates the initial execution process based on the user's personalized information and generates an execution process that conforms to the personalized information, which is then fed back to the first intelligent agent.

7. The method according to any one of claims 1-3, wherein, The personalized verification process is implemented by the second intelligent agent through the second model, and includes: The second intelligent agent sends the output result to the cloud, processes the output result through a data processing model, and feeds back the feature data of the generated output result to the second intelligent agent; The second intelligent agent inputs the feature data into the personalization model and performs personalized verification on the feature data based on the stored personalized information of the user. The second model is the personalized model deployed on the electronic device used by the user, or it includes the personalized model and the data processing model deployed in the cloud.

8. A task processing method, the method being executed by a first intelligent agent in an electronic device, comprising: In response to a user's task processing request, an execution flow containing multiple task steps is generated; Perform the task steps described above to generate the corresponding output results; If the task step is identified as a step to be verified, a verification guidance instruction for the output result of the step to be verified is received; the verification guidance instruction is generated based on the user's personalized information verification. Based on the verification guidance instruction, subsequent processing steps for the step to be verified are executed to control the execution process to continue until a task processing result that conforms to the personalized information is obtained.

9. A task processing method, the method being executed by a second intelligent agent in an electronic device, comprising: In response to the execution of the step to be verified, obtain the output result generated by the execution of the step to be verified; The step to be verified is a task step included in the execution flow generated based on the user-input task processing request. Based on the user's personalized information, the output results are personalized for verification, and corresponding verification guidance instructions are generated. Output the verification guidance command; The verification guidance instruction is used to control subsequent processing steps for the step to be verified.

10. An electronic device, the electronic device comprising: At least one memory, and a computer program stored in the memory; At least one processor capable of running intelligent agents; The intelligent agent can execute the computer program through the processor to implement the steps of the task processing method according to claim 1, 8 or 9.

Images