Information processing method and device based on multi-agent system and storage medium

Abstract

This invention provides an information processing method, apparatus, and storage medium based on a multi-agent system. The method includes: performing semantic inference based on the table name and field information of a first data table to determine the information to be monitored; establishing monitoring tasks and monitoring rules based on the information to be monitored; collecting current data snapshots according to the monitoring rules, and determining whether to output monitoring prompts based on the comparison results between the current data snapshots and historical data snapshots; outputting monitoring prompts in a dialog box with the user; listening to the scheduling process of the first agent and the monitoring process of the second agent in an asynchronous non-blocking manner, extracting business knowledge, and writing it into a knowledge base; collecting application trajectory data of the user using multiple skills in a first time period, performing aggregation analysis on the application trajectory data in a second time period to generate a skill patch set, and determining whether to push for approval based on the confidence level of each skill in the skill patch set, and performing patch storage and installation in the skill patch set.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence (AI), and in particular to an information processing method, apparatus and storage medium based on a multi-agent system. Background Technology

[0002] With the rapid development of artificial intelligence and large-scale model technology, intelligent interaction systems based on multi-agent systems have been widely used. In multi-agent systems, different types of agents cooperate and divide tasks, respectively undertaking responsibilities such as task recognition, task scheduling, task execution, and knowledge accumulation.

[0003] In existing technologies, multi-agent systems mainly suffer from the following technical problems: In no-code platform scenarios, all initial data tables and fields are user-defined, and the system cannot predict which data metrics users should monitor. Existing technologies require users to manually configure monitoring rules, which is a high-barrier operation and cannot adapt to dynamically changing business needs.

[0004] Existing multi-agent systems often rely on analyzing user behavior to understand business changes when performing monitoring tasks. However, in no-code platforms, the system cannot know the specific context of user actions, resulting in insufficient monitoring accuracy.

[0005] In existing technologies, the skills used by intelligent agents lack an effective self-evolution mechanism. Once a skill is used, the system cannot know its effectiveness, nor can it learn and optimize the skill from the experiences of multiple users. Furthermore, there is a lack of effective verification mechanisms when updating skills, and direct updates may introduce incorrect knowledge.

[0006] Current technologies lack a unified security governance mechanism for the installation and use of intelligent agent skills. Skills may contain malicious scripts, and risks cannot be effectively identified before installation or isolated during runtime. Summary of the Invention

[0007] This invention provides an information processing method, apparatus, and storage medium based on a multi-agent system.

[0008] In a first aspect, embodiments of the present invention provide an information processing method based on a multi-agent system, the multi-agent system comprising a first agent, a second agent, and a third agent, the method comprising: when the first agent obtains a first data table from a user's dialogue, the first agent invokes the second agent to perform semantic inference based on the table name and field information of the first data table to determine the information to be monitored; the first agent establishes a monitoring task and monitoring rules based on the information to be monitored; the second agent, invoked by the first agent, collects a current data snapshot according to the monitoring rules, and determines whether to output a monitoring prompt based on the comparison result of the current data snapshot and historical data snapshots; if it is determined that a monitoring prompt should be output... The first intelligent agent outputs monitoring prompts in the dialog box with the user; the third intelligent agent listens to the scheduling process of the first intelligent agent and the monitoring process of the second intelligent agent in an asynchronous non-blocking manner, and extracts business knowledge and writes it into the knowledge base; the third intelligent agent collects application trajectory data of the user using multiple skills in the first time period, aggregates and analyzes the application trajectory data in the second time period to generate a skill patch set, and determines whether to push for approval based on the confidence level of each skill in the skill patch set, and completes the patch entry and installation of the skill patch set according to the confidence level and approval result; wherein, the frequency of the user using skills in the first time period is higher than the frequency of the user using skills in the second time period.

[0009] Secondly, embodiments of the present invention provide a skill accumulation system based on a multi-agent system. The multi-agent system includes a first agent, a second agent, and a third agent. The skill accumulation system includes: a calling module, used when the first agent obtains a first data table from a user's dialogue, the first agent calls the second agent to perform semantic inference based on the table name and field information of the first data table to determine the information to be monitored; an establishment module, used when the first agent establishes monitoring tasks and monitoring rules based on the information to be monitored; an output module, used when the second agent, under the call of the first agent, collects a current data snapshot according to the monitoring rules, and determines whether to output a monitoring prompt based on the comparison result of the current data snapshot and historical data snapshots; and a prompt module. The system includes a first intelligent agent that outputs monitoring prompts in a dialog box with the user when monitoring prompts are required; a writing module for the third intelligent agent to asynchronously and non-blockingly monitor the scheduling process of the first intelligent agent and the monitoring process of the second intelligent agent, and extract business knowledge to write into a knowledge base; and a patching module for the third intelligent agent to collect application trajectory data of the user using multiple skills in a first time period, aggregate and analyze the application trajectory data in a second time period to generate a skill patch set, determine whether to push for approval based on the confidence level of each skill in the skill patch set, and complete the patch entry and installation of the skill patch set based on the confidence level and approval result; wherein, the frequency of the user using skills in the first time period is higher than the frequency of the user using skills in the second time period.

[0010] Thirdly, embodiments of the present invention provide a computer storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the method provided in any embodiment of the present invention.

[0011] The information processing method, apparatus, and storage medium based on a multi-agent system provided in this invention have the following technical effects: First, it enables proactive monitoring with zero user pre-configuration. By performing semantic inference on the first data table mentioned in the user's session, it automatically identifies the business type and monitoring dimensions, eliminating the need for users to manually configure any monitoring rules and significantly lowering the barrier to entry.

[0012] Second, it enables objective change detection based on data snapshots. By abandoning reliance on user actions and instead identifying business anomalies by periodically collecting data snapshots and comparing them with historical snapshots, it fundamentally solves the technical challenge of lacking operational context in no-code platforms.

[0013] Third, achieve skill evolution through multi-source trajectory aggregation. By collecting and aggregating application trajectory data of skills in different time periods, common patterns can be extracted from users' successful experiences and lessons learned to generate skill patches, upgrading skill optimization from single-user experience to the accumulation of collective wisdom. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating an information processing method based on a multi-agent system according to an embodiment of the present invention. Figure 2 This is another flowchart illustrating an information processing method based on a multi-agent system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a task execution device based on an intelligent agent according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0016] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The terminology used herein includes, and / or encompasses, any and all combinations of one or more of the associated listed items.

[0017] For ease of understanding, the following explains some key terms in this embodiment: Intelligent agent: Generally refers to a software entity or system capable of sensing the environment, making decisions, and performing actions. In this embodiment, the intelligent agent is designed as an independent module with specific functions and responsibilities, such as software or programs based on AI models that are responsible for listening, scheduling, executing, or optimizing. Exemplarily, the intelligent agent includes, but is not limited to, various AI agents. Exemplarily, the intelligent agents in this embodiment can all operate based on user permissions (e.g., passwords).

[0018] The first intelligent agent, which can be a system-level intelligent agent, performs global task extraction, task parsing, and task scheduling for the first user. The first intelligent agent can be automatically set up by the system, without requiring manual creation by the user.

[0019] It is worth noting that the name, preferred style, working time and / or working mode of the first intelligent agent can be set by the first user.

[0020] The first intelligent agent is primarily responsible for recognizing dialogue information, determining target tasks, scheduling tasks, and updating the user's schedule. For example, the dialogue channel of the first intelligent agent remains open at all times and is not blocked by the execution of any target task. When the first intelligent agent schedules the second intelligent agent to execute a task, the first intelligent agent can release the current dialogue channel, allowing the first user to continue new dialogue interactions with the first intelligent agent. The second intelligent agent executes the task asynchronously in the background and reports the results to the first intelligent agent upon completion.

[0021] The second intelligent agent is a task-execution intelligent agent, which can be a system-level intelligent agent or a workspace-level intelligent agent, etc. The second intelligent agent can be installed or generated based on user instructions. For example, the second intelligent agent is a task-execution intelligent agent. The second intelligent agent is launched through the scheduling of the first intelligent agent, and its execution mode includes one of the following two: (1) Sub Agent: Shares the same skills, memories, tools, and user permissions as the first agent, but does not inherit the first agent's current dialogue context, focusing on the execution of a single task; for example, a Sub Agent can correspond to a sub-execution process: an asynchronous execution session started in the background by the first agent when performing complex tasks, not an independent agent entity, but an execution mode. The Sub Agent shares the first agent's skills, memories, tools, and user permissions, does not inherit the first agent's current dialogue context, and is transparent to the user (black box). The Sub Agent can be a clone of the first agent itself, or it can be the execution form of a specialized agent when it is scheduled. The Sub Agent exists during execution and reports the results to the first agent after completion.

[0022] (2) Specialized intelligent agents: Independent intelligent agents created by users with specific professional skills, at the same level as the first intelligent agent. From a physical implementation perspective, all intelligent agents are stateless configurations with context loading. The difference is that specialized intelligent agents load skills in specific fields to ensure accuracy.

[0023] Similarly, the name, preferred style, working time, and / or working mode of the second intelligent agent can be set by the first user. The second intelligent agent is primarily responsible for generating deliverables through the execution of target tasks, performing target tasks according to user instructions, and submitting or notifying the user of delivery status via dialog boxes.

[0024] In some embodiments, the second intelligent agent can be a task-execution intelligent agent responsible for performing specific tasks such as data snapshot collection and semantic inference. The second intelligent agent can be an avatar of the first intelligent agent or an independent intelligent agent created by the user with specific professional skills.

[0025] The third agent is a system-level agent that does not participate in task scheduling or execution. It is a secretary-type agent and is a built-in agent of the system. Similarly, the name, preferred style, working time, and / or working mode of the third agent can be set by the first user. The third agent is mainly responsible for asynchronously monitoring the entire task execution process, accumulating knowledge, and optimizing the task scheduling of the first agent and the task execution of the second agent.

[0026] For example, the third intelligent agent (represented in the system as the knowledge secretary "Xiaoyun") is a system-level intelligent agent that does not participate in task scheduling and execution. It is a secretary-type intelligent agent and is a built-in intelligent agent of the system. The third intelligent agent has an independent interaction entry point in the product, independent of the dialogue window of all business intelligent agents. It proactively interacts with users in a friendly secretary-like tone and style, proactively contacting users a maximum of 2 to 3 times a day, and taking priority over batch reporting over individual notifications.

[0027] First User: Refers to the direct user of the system service, who conducts daily communication and task interaction through the chat channel.

[0028] Chat channels refer to the means by which users access the intelligent agent system, including but not limited to instant messaging applications (WeChat Work, Lark, DingTalk, WeChat, Slack, Microsoft Teams), platform-built-in web chat windows, and platform-built-in mobile dialogue interfaces. Each channel corresponds to an independent session, with web and mobile sessions synchronized in real time. Third-party platform channels are not interoperable with the platform's own products.

[0029] The second user can be any user different from the first user, such as the first user's superior, peer, or other users who can help the first user grant permissions.

[0030] Dialogue information: refers to the various communication content participated in by the first user in the chat channel. This dialogue information can include human-computer (P2A) dialogue information, that is, the direct communication between the first user and the intelligent agent; human-to-human (P2P) dialogue information, that is, the point-to-point communication between the first user and other human users; and group dialogue information, that is, the communication content generated by the first user in a multi-person group chat.

[0031] First Data Table: Any table provided or referenced by the user in the user dialogue. For example, this first data table may include, but is not limited to, user-defined structured or unstructured first data tables. The first data table may contain a table name and multiple fields. In the no-code platform, the name of the first data table and the field names are defined by the user, and the structure of the first data table may vary significantly between different users.

[0032] Monitoring task: Refers to an automated work unit established by a first intelligent agent for periodically performing data acquisition and anomaly detection. One monitoring task corresponds to a set of monitoring requirements in a first data table, containing one or more monitoring rules for that first data table. The execution of a monitoring task is triggered by the first intelligent agent according to a preset time scheduling strategy (e.g., daily, weekly). When a task is triggered, the first intelligent agent calls a second intelligent agent to perform the specific snapshot acquisition operation. Monitoring tasks have independent lifecycles and can be created, started, paused, resumed, and deleted.

[0033] Monitoring rules: A set of rules used to define which data metrics to monitor, at what frequency to collect them, and when to trigger alerts when differences reach a certain threshold.

[0034] Monitoring information refers to a structured set of information generated by the second intelligent agent through semantic inference to assist the user in establishing monitoring tasks. The monitoring information must include at least the following two types of information: (i) Business type to be monitored: This refers to the business domain category to which the first data table belongs, inferred by the second intelligent agent. The business type helps the system understand the core purpose of the first data table, thereby recommending monitoring dimensions that match the business scenario. Business types include, but are not limited to: Customer Relationship Management (CRM), Sales Management, Project Management, Inventory Management, etc.

[0035] (ii) Business parameters to be monitored: These refer to the key business indicators and their constraints that the second intelligent agent identifies from the fields of the first data table and that have monitoring value. The specific content of the business parameters includes: The monitoring information is automatically generated by the second intelligent agent after semantic analysis of the table name and field name of the first data table using a large language model. The first intelligent agent displays the monitoring information to the user for confirmation. The user can adjust parameters such as suggested conditions and suggested frequency. The first intelligent agent then establishes formal monitoring tasks and monitoring rules based on the user's confirmation or adjustment of the parameters.

[0036] Application trajectory data: Records complete execution information when a skill is invoked, including user identifier, skill identifier, input content, output content, and a marker indicating whether the user corrected the agent's response.

[0037] Skill Patch: The smallest unit of modification to a skill. Each patch is independently labeled with the type of change, the content before and after the modification, the source of evidence, and the risk level.

[0038] The first time period and the second time period: The first time period is the high-frequency period when users use skills (such as during the day), and the second time period is the low-frequency period when users use skills (such as at night). Through time-sharing data collection and aggregation analysis, an offline evolutionary architecture is implemented that collects data during the day and processes it at night.

[0039] This embodiment provides an information processing method based on a multi-agent system, which can be applied to a cloud-native multi-agent collaborative work platform.

[0040] S101: When the first agent obtains the first data table from the user's dialogue, the first agent calls the second agent to perform semantic inference based on the table name and field information of the first data table to determine the information to be monitored.

[0041] In some embodiments, the user's conversation may be a conversation generated by the user in various conversation channels. Exemplarily, the conversation includes at least one of the following: human-computer P2A conversation, human-to-human P2P conversation, and group conversation; exemplarily, listening to each chat channel of the first user or listening to a chat channel specified by the first user.

[0042] In some embodiments, a user can upload a first data table file during a session or specify an existing first data table via a dialogue command. A first agent listens to the user's session and triggers a semantic inference process when it detects that the user mentions or uploads the first data table. The first agent sends the table name and all field names of the first data table to a second agent, which then analyzes the data using the semantic understanding capabilities of a large language model. For example, if a user uploads a first data table named "Customer Follow-up Records," which includes fields such as customer name, follow-up status, last follow-up date, expected contract signing date, and estimated amount, the second agent sends this information to the large language model. The large language model infers that the table belongs to the customer relationship management business type based on the table name, infers that overdue follow-up needs to be monitored based on the last follow-up date field, and infers that upcoming contract signings need to be monitored based on the expected contract signing date field.

[0043] In some embodiments, the business type indicates which business scenario the first data table belongs to, and the business parameters indicate the specific metrics that need to be monitored. The second agent returns the monitoring information to the first agent.

[0044] In some embodiments, the process by which the second intelligent agent performs semantic inference based on the table name and field information of the first data table includes the following sub-steps: The second intelligent agent infers the business type of the first data table based on the table name and field names of the first data table using a large language model. The business type includes at least one of customer relationship management, sales management, project management, and inventory management. The large language model determines which business scenario the first data table belongs to by performing semantic understanding on the table name.

[0045] The second agent classifies the fields based on the large language model, and the classification includes at least one of time fields, status fields, numerical fields, and personnel fields. The large language model classifies each field according to the field name and semantics. For example, the last follow-up date is classified as a time field, the follow-up status is classified as a status field, the estimated amount is classified as a numerical field, and the person in charge of follow-up is classified as a personnel field.

[0046] The second agent outputs the information to be monitored based on the LLM; the information to be monitored includes at least one recommended monitoring dimension, and the at least one monitoring dimension includes at least one of a monitoring name, an associated field, a recommended condition, and a recommended frequency. The large language model outputs monitoring dimensions that are deeply bound to business semantics based on the field classification results and business types. For example, for the time field of the last follow-up date, the monitoring dimension of customers with overdue follow-up is output, the recommended condition is more than a certain number of days from now, and the recommended frequency is daily.

[0047] When the inference confidence of the large language model is lower than the first threshold, the first agent asks the user about the business purpose of the first data table through a dialog box, and appends the user's answer to the prompt for re-inference. For example, when the table name and field names of the first data table are too ambiguous, such as Table A, Field A, the large language model cannot infer the business meaning from these meaningless names, and the inference confidence is lower than the threshold. The first agent asks the user about the business purpose of the first data table. After the user answers, the first agent appends the answer to the prompt and re-invokes the large language model for inference.

[0048] Through the above semantic inference process, the establishment of monitoring rules has changed from manual configuration by users to intelligent recommendation by the system plus one-click confirmation by users, reducing the user operation steps from dozens to two. At the same time, the monitoring rules have changed from technical indicators to business semantics, and users can understand the monitoring content without understanding the underlying data structure. For meaningless field names, through the human-machine collaborative interaction and inquiry mechanism, the accuracy of semantic inference is ensured.

[0049] S102: The first agent establishes a monitoring task and monitoring rules based on the information to be monitored.

[0050] In some embodiments, the first agent displays the information to be monitored to the user for confirmation. The display content includes: the inferred business type, the list of recommended monitoring dimensions, the recommended conditions and recommended frequencies for each dimension. The user can adjust the recommended conditions, for example, change more than 7 days to more than 3 days, or add monitoring dimensions that the user is concerned about.

[0051] In some embodiments, the first intelligent agent receives confirmation or adjustment instructions from the user and establishes a monitoring task based on the monitoring dimensions ultimately confirmed by the user. Each monitoring task corresponds to a first data table, which may contain multiple monitoring rules. Each monitoring rule independently defines the monitoring metrics, sampling frequency, and change threshold.

[0052] Through the user confirmation process, system suggestions are organically combined with users' personalized needs. Users can complete the establishment of monitoring tasks with just one click of confirmation or fine-tuning, realizing proactive monitoring configuration with zero presets.

[0053] In some embodiments, after the first agent determines the monitoring task, it will also set a monitoring identifier for each monitoring task. The monitoring identifier (ID) may include: (e.g., user ID), channel information (e.g., chat channel ID, dialog box ID), table ID of the first data table, etc., in sequence, so that the various agents in the subsequent intelligent agent system can track the source of the task, consumers, etc. based on the monitoring ID.

[0054] S103: The second agent, under the invocation of the first agent, collects the current data snapshot according to the monitoring rules, and determines whether to output a monitoring prompt based on the comparison result between the current data snapshot and the historical data snapshot.

[0055] The second agent, under the invocation of the first agent, collects a current data snapshot according to the monitoring rules, and determines whether to output a monitoring prompt based on the comparison result between the current data snapshot and the historical data snapshot.

[0056] In some embodiments, when the trigger time for a monitoring task arrives, the first agent invokes the second agent to perform snapshot collection.

[0057] In some embodiments, the second agent executes a query statement to collect current data based on the collection indicators in the monitoring rules. More specifically, the second agent collects indicators related to the monitoring rules and executes a query statement at a preset frequency to collect at least one of the following indicators: the number of records that meet the monitoring conditions, the aggregate value of the target field, and the count of each category of the status field.

[0058] In some embodiments, the second agent stores the collected current data snapshot in a snapshot storage module and reads the previously collected historical data snapshot from the snapshot storage module. The first agent compares the current snapshot with the historical snapshot and calculates change information such as absolute change and percentage change rate.

[0059] In some embodiments, when the absolute change exceeds a second threshold preset in the monitoring rules, and / or the percentage change rate exceeds a third threshold preset in the monitoring rules, the first agent determines that a monitoring prompt needs to be output.

[0060] In some embodiments, data anomalies are detected by snapshot comparison rather than analysis of user behavior, fundamentally solving the technical challenge of lacking user operation context in no-code platforms. Collecting only metrics relevant to monitoring rules, rather than scanning the entire table, significantly reduces computational overhead and storage costs. A dual threshold mechanism of absolute change and percentage change rate effectively filters out normal business fluctuations and reduces invalid monitoring alerts.

[0061] Step S104: If it is determined that a monitoring prompt should be output, the first agent outputs a monitoring prompt in the dialog box with the user.

[0062] In some embodiments, a first agent sends the discrepancy data to a large language model or a second agent for evaluation to generate insight text and action suggestions. Exemplarily, the insight text includes a description of data changes and / or business implications analysis, and the action may be presented as a button. The first agent displays the insight text and action suggestions as cards in a user dialog box.

[0063] In some embodiments, the method further includes: The first intelligent agent dynamically adjusts the prompt weight based on user feedback to the monitoring prompts. The first intelligent agent independently tracks the prompt effect for each monitoring rule. This prompt effect can record the total number of prompts, the number of user clicks, the number of times the user ignores the prompt, and / or the number of times the user closes the prompt. After each monitoring prompt is pushed, the first intelligent agent records the user's behavior: clicking the action suggestion button is recorded as a click, not taking any action is recorded as ignoring, and actively closing the prompt is recorded as closing.

[0064] The first intelligent agent updates the prompt weight in the monitoring rules based on the prompt effect, wherein the prompt weight is positively correlated with the number of times the user views the message and negatively correlated with the number of times the user ignores the message and the number of times the user closes the message. For example, the prompt weight is negatively correlated with the number of times the user ignores the message by a first coefficient; the prompt weight is negatively correlated with the number of times the user closes the message by a second coefficient; the second coefficient is greater than the first coefficient. In this way, the monitoring and / or prompt frequency is adaptively adjusted based on user feedback, improving the user experience.

[0065] When the prompt weight is lower than the fourth threshold, the first agent pauses the corresponding monitoring task or monitoring prompt.

[0066] When the prompt weight is higher than the fifth threshold, the first agent increases the monitoring frequency of the corresponding monitoring task or the push frequency of monitoring prompts, where the fifth threshold is greater than the fourth threshold.

[0067] The large language model translates changes in user behavior into business language that users can understand and actionable suggestions, achieving a leap from data monitoring to business insights. By recording user clicks, ignores, and closes, and dynamically adjusting the weight of prompts, the quality of prompts is calibrated. Push notifications with a weight below a threshold are automatically paused to avoid disturbing users, while push notifications with a weight above a threshold can be appropriately increased in frequency to ensure that high-value information reaches users more promptly.

[0068] Step S105: The third intelligent agent listens to the scheduling process of the first intelligent agent and the monitoring process of the second intelligent agent in an asynchronous non-blocking manner, and extracts business knowledge and writes it into the knowledge base.

[0069] In some embodiments, the third agent decouples itself from the first and second agents through an event bus and a message queue.

[0070] When the second agent performs a monitoring task, it publishes the execution steps, query statements, and collection results as events to the event bus. The third agent subscribes to the event bus and consumes the events asynchronously through a message queue.

[0071] The third agent employs a hierarchical sampling strategy for monitoring: the basic sampling layer collects execution state snapshots at low frequencies; the dynamic enhancement layer automatically switches to high-frequency mode and triggers full-dimensional monitoring when it receives an abnormal event; and the result locking layer locks the final sampling window to capture the execution result when it receives a task completion event.

[0072] The third intelligent agent extracts reusable business rules, user preferences, decision-making logic, and other knowledge from the monitored information, and performs hierarchical processing based on confidence level: high-confidence knowledge is directly written into the knowledge base, medium-confidence knowledge sends a confirmation request to the user through an independent channel, and low-confidence knowledge is temporarily stored to await more context.

[0073] The third agent implements asynchronous, non-blocking listening via an event bus and message queue, ensuring the smooth execution of the first and second agents without interfering with the main task flow. A tiered sampling strategy uses low-frequency sampling during normal conditions to conserve resources and high-frequency sampling during anomalies to capture details, achieving an optimal balance between listening integrity and system performance. A confidence-based tiered processing mechanism allows high-confidence knowledge to be silently accumulated, medium-confidence knowledge to request confirmation, and low-confidence knowledge to be observed, improving both the efficiency and accuracy of knowledge accumulation.

[0074] Step S106: The third intelligent agent collects application trajectory data of the user using multiple skills in the first time period, aggregates and analyzes the application trajectory data in the second time period to generate a skill patch set, and determines whether to push for approval based on the confidence level of each skill in the skill patch set. Based on the confidence level and the approval result, the patch in the skill patch set is added to the database and installed; wherein, the frequency of the user using skills in the first time period is higher than the frequency of the user using skills in the second time period.

[0075] In some embodiments, the first time period can be a high-frequency period when users use skills, such as during the day, and the second time period can be a low-frequency period when users use skills, such as at night. The third agent collects multiple application trajectory data during the day. Each application trajectory data includes a user identifier, a skill identifier, input content, output content, and a marker indicating whether the user corrected the agent's response. After the offline analyzer is activated at night, the third agent aggregates and analyzes all the trajectory data collected during the day.

[0076] More specifically, the third agent identifies operation sequences in multiple successful application trajectories of the user through a clustering algorithm. When the same operation sequence appears in more than a preset number of successful application trajectories, it is marked as a success mode candidate. The third agent performs sequence mining on the successful application trajectories, serializes the operation steps in each trajectory, and aggregates similar operation sequences together using a clustering algorithm. When the same operation sequence appears multiple times in successful application trajectories, the operation sequence is marked as a success mode candidate.

[0077] The third intelligent agent compares successful application trajectories and failed application trajectories in the same scenario using a sequence comparison algorithm, identifies operation steps that exist in successful application trajectories but are missing in failed application trajectories, and extracts the missing operation steps as skill content to be added.

[0078] The third intelligent agent aligns successful and failed application trajectories in the same scenario, and uses a sequence comparison algorithm to find operation steps that are present in the successful application trajectory but not in the failed application trajectory. These missing operation steps are the skill content to be added.

[0079] Based on the skill content to be added, a set of skill patches is generated. Each patch includes at least a patch type, original content, new content, and evidence source. Patch types include addition, modification, and deletion.

[0080] Thus, the architecture of collecting data in the first time period and aggregating it offline in the second time period fully utilizes the system's idle computing resources without affecting normal daytime use. Through multi-source trajectory aggregation, the successful experiences of multiple users are transformed into reusable organizational assets, realizing a leap from individual learning by a single user to collective wisdom by multiple users. Clustering algorithms identify successful pattern candidates, and sequence alignment extracts missing steps, providing clear evidence to support skill optimization and avoiding blind modifications. Patch-based updates rather than full rewrites reduce the risk exposure of each update.

[0081] Furthermore, before the skill patch is installed, a third agent verifies the new skill version. The third agent extracts real-world dialogues from a preset number of days as a test set. Each dialogue record in the test set includes the user's input query, the agent's generated response, and a marker indicating whether the user corrected the response. The third agent inputs each dialogue record from the test set into the new skill version to generate candidate responses. The third agent uses the user's actual acceptance as a benchmark; the user's actual acceptance is determined by using the original response when the user did not correct it, and the user's corrected response when the user corrects it. The third agent calculates the percentage by which the new skill version is superior to the current skill version; when this percentage reaches a preset threshold, the new skill is deemed to have passed verification.

[0082] Using real-world dialogues as the test set, rather than manually constructed test cases, reflects the distribution of real-world scenarios. Using user-accepted results as the benchmark, rather than self-evaluation by a large language model, makes the verification results more objective and reliable. The replay verification process does not interfere with online services, achieving lossless verification. Quantified superiority and inferiority ratios provide clear data support for upgrade decisions, solving the technical problems of being hesitant to update skills or fearing errors during updates.

[0083] Furthermore, the method also includes a third agent performing static code analysis on the files in the skill directory before the skill patch is installed, matching them against a preset risk rule base, and directly blocking the installation when a hard trigger rule is hit.

[0084] The third agent calculates the security score of the skill. When the security score falls below the sixth threshold, the first agent displays risk details on the installation interface and prompts the user to confirm each item before continuing the installation. The third agent executes the scripts within the skill in an independent sandbox container, detecting abnormal behavior in real time and interrupting the installation process upon detection of abnormal behavior. The third agent scans the installed skills at preset intervals, automatically notifying the skill owner when security risks are detected and generating a security audit report containing a list of risky skills and remediation suggestions.

[0085] In some embodiments, the abnormal behavior includes at least one of the following: non-compliant token pass-through, unauthorized data transfer, and unauthorized access attempts.

[0086] In this way, security protection is progressively strengthened through pre-installation scanning, installation verification, runtime isolation, and regular auditing, achieving full lifecycle security governance. The hard-triggered rule base specifically targets the unique risks of executable content in large language models (such as piped execution of remote scripts, reverse shells, etc.), filling a gap in traditional security solutions in this area. The security scoring mechanism quantifies risks into intuitive scores, facilitating user decision-making. Runtime sandbox isolation ensures that even if a script contains unknown vulnerabilities, it will not affect the host system.

[0087] Furthermore, the method also includes a knowledge retrieval step. This knowledge retrieval step may include: after the first agent receives the user's skill query instruction, it invokes the third agent to perform multi-path parallel retrieval.

[0088] The multi-path parallel retrieval includes at least two retrieval paths: a vector retrieval path, a keyword retrieval path, and a graph retrieval path. Vector retrieval converts the user query into a semantic vector and calculates its similarity with the vector index in the knowledge base. Keyword retrieval segments the user query and obtains highly relevant knowledge fragments by matching them with an inverted index. Graph retrieval constructs a knowledge graph from the entities and relationships in the knowledge base and retrieves knowledge fragments associated with the query entity. The third agent deduplicates and merges the knowledge fragments obtained from multiple paths, calls a reordering model to calculate a relevance score for each knowledge fragment, sorts them in descending order of relevance score, and returns the top-ranked knowledge fragments.

[0089] The three-pronged parallel recall approach (vector, keyword, and graph retrieval) complements each other. Vector retrieval excels in semantic understanding, keyword retrieval excels in exact matching, and graph retrieval excels in relational reasoning. The combination of these three approaches significantly improves both accuracy and recall. The re-ranking model performs a secondary ranking of the candidate set, further enhancing relevance accuracy. Source citations are included when returning results, enhancing traceability and facilitating user verification.

[0090] Through the above method, this invention achieves end-to-end intelligent management, from zero-preset monitoring establishment, objective data change detection, multi-source trajectory aggregation and evolution to layered security governance. The first intelligent agent is responsible for scheduling and monitoring alerts, the second intelligent agent is responsible for semantic inference and snapshot collection, and the third intelligent agent is responsible for asynchronous listening and knowledge accumulation. These three agents collaborate and perform their respective duties. This invention solves the core technical problems of existing technologies, such as monitoring rules relying on manual configuration, skills failing to evolve self-sufficiently, and a lack of security governance, achieving significant technological progress.

[0091] In some embodiments, the intelligent agent system has a hierarchical memory model that remembers user-participated dialogues in the form of sessions. A dialogue may include one or more sessions. Specifically, the method also includes at least one of the following: Based on the instantaneous memory layer of the hierarchical memory model, the immediate context of the memory session is remembered; Based on the hierarchical memory model, the short-term memory layer remembers chat information within a preset time period, a preset number of messages, or a preset number of sessions; Based on the intermediate memory layer of the hierarchical memory model, cross-session behavioral models and / or preference memory are performed; Based on the long-term memory layer of the hierarchical memory model, the user profile of the first user, the memory of personal business knowledge, the memory of the first user's work style, the team profile of the team to which the first user belongs, and the memory of team business knowledge are performed. Based on the persistent memory layer of the hierarchical memory model, the setting information of the first user is memorized; Based on the hierarchical memory model, the task memory layer stores dialogue information related to task execution and information required for task execution in the form of tasks.

[0092] Thus, a hierarchical memory model is a memory structure that organizes and stores information according to different time scales, levels of abstraction, or importance. Information can be divided into multiple layers, each responsible for storing specific types and time-sensitive information. This model can be implemented using a database system (e.g., relational databases, NoSQL databases) combined with a caching mechanism, where different layers correspond to different table structures or storage strategies. Alternatively, it can be implemented by combining in-memory databases with persistent storage; for example, the transient and short-term memory layers are stored in memory to improve access speed, while the medium-term, long-term, persistent, and task-oriented memory layers are stored on disk or in a distributed storage system.

[0093] For example, the transient memory layer is primarily used to remember the immediate context of the dialogue. This layer can be implemented using a circular buffer or a queue to store the latest rounds of messages from the ongoing dialogue, allowing the agent to quickly access and understand the context of the current conversation. Alternatively, it can be implemented using an attention mechanism based on the Transformer model, processing the input sequence of the current dialogue as immediate context without explicit storage.

[0094] For example, the short-term memory layer is used to store chat information within a preset time period, a preset number of messages, or a preset number of sessions. This layer can be managed using timestamps or message counters, for example, storing all messages from the past five minutes, the most recent fifty messages, or all messages from the most recent three sessions. Alternatively, it can be implemented using a sliding window mechanism, where older messages are moved out of the window as new messages arrive, ensuring that the memory always remains within the preset range.

[0095] For example, the intermediate memory layer is used for cross-session behavioral modeling and / or preference memorization. This layer can be implemented through user behavior log analysis and machine learning models (e.g., clustering algorithms, recommender systems) to identify recurring behavioral patterns or preferences of the first user in different conversations and abstract them into models. Alternatively, it can be implemented through rule engines and knowledge graphs to store and reason about the first user's choices, habits, etc., in specific contexts as rules or entity relationships.

[0096] For example, the long-term memory layer is used to create a user profile for the first user, remember personal business knowledge, and create a team profile and remember team business knowledge for the first user's team. This layer can store the first user's personal information, professional background, interests, etc., through a structured database, and store personal business documents and team shared knowledge bases through a document database. Alternatively, it can be implemented using embeddings and a vector database to transform user profiles and business knowledge into high-dimensional vectors, and then retrieve and match them using similarity search. For example, the first data representation established for the monitoring task is at least remembered by the long-term memory layer.

[0097] For example, the persistent memory layer is used to remember the settings information of the first user. For example, based on the explicit memory layer of the hierarchical memory model, it remembers the information actively set by the first user, including the first user's identity information, the role definition of the first intelligent agent, and work style preference settings. The information in the explicit memory layer is actively specified by the first user, has the highest priority, and does not decay over time. This persistent memory layer can be implemented through a user authentication system and a configuration management system, storing the first user's unique identifier, permission information, and their habitual operating procedures and communication styles during task execution. Alternatively, it can be implemented through blockchain technology or encrypted storage to ensure the security and immutability of the first user's identity and work style data. For example, the task memory layer remembers dialogue information related to task execution and information required for task execution in the form of tasks. This layer can be implemented through the database of a task management system or project management tool, storing the detailed description of each task, related dialogue records, attachments, deadlines, responsible persons, and other information in association. Furthermore, it can be implemented through event logs and state machine models to record each stage of the task from creation to completion, as well as the dialogues and resources involved in each stage. After the monitoring task has been continuously executed for a period of time, the first data table can be remembered by the persistent memory layer.

[0098] Thus, the solution in this application introduces a hierarchical memory model, enabling the first agent to store the chat information it listens to and acquires in the first user's chat channel into different memory layers according to their timeliness and importance. Specifically, the immediate conversation content is first captured by the transient memory layer for quick understanding of the current context. Subsequently, this information enters the short-term memory layer, remaining active within a preset time, message, or conversation range, allowing the agent to review recent communication history and maintain the coherence of the conversation. As the conversation progresses and the task is executed, the medium-term memory layer learns from multiple conversations and precipitates the first user's behavioral model and preferences, such as the tools or processing methods the user tends to choose in specific situations. The long-term memory layer is responsible for building and maintaining a more macro-level user profile and personal and team business knowledge, which is crucial for understanding the background of complex tasks and providing professional support. The persistent memory layer stores the first user's identity information and unique working style, which helps the agent better adapt to the user's personalized needs when performing tasks.

[0099] In some cases, the task memory layer is specifically task-centric, storing dialogue fragments related to task execution and all necessary resource information to ensure the integrity and traceability of task execution. Through this multi-layered, multi-dimensional collaborative memory, the first agent, when identifying deliverables and scheduling the second agent, no longer relies solely on the current dialogue but can comprehensively utilize information ranging from immediate context to long-term knowledge. For example, when identifying a potential task, the first agent can query the medium-term memory layer to understand the user's preferences for similar past tasks and the long-term memory layer to obtain relevant business knowledge, thereby more accurately determining the target task and formulating a more reasonable execution plan. When executing a task, the second agent can also utilize the detailed information stored in the task memory layer and the user's work style in the persistent memory layer to generate deliverables or content that better meet the user's expectations. This mechanism significantly improves the agent's ability to understand user intent, predict user needs, and optimize task execution, effectively solving the problem of context loss and incomplete understanding caused by relying solely on immediate dialogue information.

[0100] This application proposes an information processing method based on a multi-agent system, in which a third agent asynchronously monitors the execution process, results, feedback, and user interventions during the target task to accumulate knowledge and optimize the task scheduling of the first agent and the task execution of the second agent. However, without effective processing of this asynchronously monitored raw information, knowledge accumulation may be inefficient, making it difficult to extract reusable and valuable skills and optimization strategies, thus limiting the self-learning and evolution capabilities of the agent system.

[0101] The method may further include: The third intelligent agent constructs and asynchronously listens to information, and extracts skills from the constructed asynchronous listening information; A third intelligent agent filters skills and determines their value; When the value of a skill meets the demand, skill information is output; the skill information includes at least one of the following: the function targeted by the skill, the operation process, the tool calling rules, the user execution preferences, and the task processing mode; the task processing mode includes at least one of the following: task identification mode, task decomposition mode, task scheduling mode, and task execution mode. Based on user feedback regarding skill information, the skill information is added to the knowledge base.

[0102] In some embodiments, structured asynchronous listening information refers to the process of converting raw, unstructured, or semi-structured data acquired by a third-party agent during the listening process—such as log files, dialogue records, system events, and user operation sequences—into a well-defined data model or pattern. This may include, but is not limited to: parsing and mapping data using predefined JSON or XML patterns; using natural language processing (NLP) techniques to identify entities, relationships, and events from text and convert them into structured representations; or discretizing continuous system behavior into a structured event stream by defining event types and parameters.

[0103] Extracting skills from structured information refers to identifying and abstracting reusable sequences of operations, decision-making logic, or tool invocation patterns that have specific functions or solve specific problems. This can be achieved through rule matching, machine learning-based models (e.g., sequence labeling models, classification models) to identify recurring effective operation patterns, or through extraction using templates pre-set by expert systems.

[0104] In some embodiments, skill filtering refers to screening extracted skills according to preset criteria to remove redundant, inefficient, or unsuitable skills. Determining skill value refers to assessing the contribution of a skill to system performance improvement, user satisfaction, or task completion efficiency. This can be quantitatively evaluated based on the number of successful skill executions, task completion rate, explicit or implicit user feedback on the skill (e.g., likes, frequency of reuse), resources consumed during skill execution, or through A / B testing.

[0105] In some embodiments, the skill information is stored in a structured text format, including but not limited to Markdown format. This skill information is loaded into the agent's context to guide the agent's behavior, rather than being executed directly by code. The skill information includes a YAML-formatted metadata header containing skill name and skill description fields for system indexing and management.

[0106] In some embodiments, the skill information includes at least one of the following content formats: (1) Declarative content: business rules, operational preferences, decision-making logic, domain knowledge; (2) Procedural content: Multi-step operation methods with three or more steps, including step sequence and decision branches; (3) Collaborative content: the division of roles, collaborative execution order, and boundary constraints among multiple agents, which are used to guide the division of labor and cooperation among multiple shared agents in the same work area when handling similar tasks.

[0107] In some embodiments, the skill information is divided into three levels according to its visibility: Personal-level skills: Visible and usable only by the skill creator, and managed by the creator. Workspace-level skills: Visible and usable by workspace members, managed by workspace administrators. Upgrading personal-level skills to workspace-level skills requires approval from the workspace administrator. Enterprise-level skills: Visible and usable by all enterprise users, managed by enterprise administrators. Upgrading workspace-level skills to enterprise-level requires approval from the enterprise administrator.

[0108] The skill information triggers a security scan when it is created, modified, or upgraded and shared, and prevents the installation or upgrade of the skill when high-risk content is detected.

[0109] In some embodiments, the third intelligent agent is a globally resident knowledge secretary AI built into the system. After extracting skill-related information, it does not directly store it, but instead judges the value of the extracted skills according to a preset filtering framework to ensure that the stored skills have high availability, stability, and reusability. Specifically, this includes, but is not limited to, one or more of the following steps: First layer of filtering: Reusability determination. The third agent determines whether the skill is applicable to at least two similar task scenarios and whether it can be repeatedly invoked outside of the current single task. If it is only applicable to the current temporary scenario and does not have general execution value, it is determined to have low reusability and is filtered out.

[0110] The second layer of filtering: stability judgment. The third agent detects whether there are ambiguities, conflicts or unreproducible problems in the execution process, operation logic and tool calling method corresponding to the skill; if the skill depends on temporary data, occasional conditions or rules that are prone to failure, it is judged to be unstable and filtered out.

[0111] The third layer of filtering: source reliability judgment. The third intelligent agent verifies the source of skill extraction to determine whether it comes from explicit user instructions, effective intervention, confirmed operation, or high-frequency correct execution records; if it comes from invalid dialogue, test input, incorrect demonstration, or unconfirmed content, it is judged as unreliable and filtered out.

[0112] The fourth layer of filtering: Conflict detection. The third intelligent agent matches and compares the skills to be added with existing personal, work area and enterprise-level skills to determine whether there are duplicates, contradictions or logical conflicts; if there are conflicts and they cannot be automatically merged, they are marked as pending review and will not enter the value confirmation stage for the time being.

[0113] Fifth layer of filtering: Knowledge classification and applicability judgment. The third intelligent agent classifies skills according to business scenarios, function types, execution subjects, and scope of authority, determines whether they should belong to the personal level, work area level, or enterprise level, and confirms their matching degree with the corresponding intelligent agent; skills that do not match the current scope are isolated according to level and do not enter the general skill library.

[0114] The sixth layer of filtering: Collaboration pattern recognition. When a task involves two or more agents, the third agent additionally extracts the collaboration patterns between the agents, including: the division of roles among the agents, the execution order (serial or parallel), boundary constraints (what cannot be done, how to resolve conflicts), and reuse conditions (in what scenarios the same collaboration method should be used). When the same group of agents appears repeatedly at least three times on the same type of task, it is marked as a candidate collaboration pattern and it is recommended to include it in the collaborative content of skills.

[0115] For example, the third agent comprehensively scores and determines the skill value based on the above five-layer filtering results, classifying the skill value into three levels: high value, usable value, and low value. If all filtering criteria are met, the skill is determined to be of high value and enters the user confirmation process. If a skill meets some of the conditions and has no obvious defects, it is determined to be a usable skill and stored in the temporary skill pool. Skills that do not meet the specified filtering criteria are deemed low-value and are directly filtered out and discarded. The specified filtering criteria can be one or more of the aforementioned filtering criteria.

[0116] For high-value skills, the third agent proactively prompts the first user for confirmation. After user confirmation, the skill is stored in the corresponding level skill library in structured Markdown format to expand the agent's execution capabilities and effectively accumulate organizational experience.

[0117] In some embodiments, the triggering conditions for the third agent to extract skills include at least one of the following: (1) Each round of dialogue buffering reaches the preset number of rounds; (2) The business session ends; (3) After the second agent completes the task, it marks the method as worthwhile to accumulate. The mark does not directly create a skill. The triggering condition is only used as a priority prompt for the third agent to analyze. (4) The first user explicitly instructed the skill extraction.

[0118] In other embodiments, the value of a skill can be comprehensively calculated based on multiple dimensions such as its success rate in a specific task scenario, execution time, and user satisfaction rating. Skill information is structured data describing the detailed attributes and usage of a skill. It can include at least the function the skill targets, such as creating a meeting or sending an email; the operation process, such as a sequence of steps or decision branches; tool invocation rules, such as which API is called and what parameters are passed; user execution preferences, such as which operation method the user prefers in a specific context; and task processing mode.

[0119] In some embodiments, the output skill information may be provided to other intelligent agents for invocation via an API interface, or stored in a readable format in a skill base for system administrators or developers to access. User feedback refers to opinions, evaluations, or behavioral data provided by users after using or evaluating the skill information. This may include user satisfaction ratings for skill execution results, suggestions for modifying skill operation procedures, adjustments to tool invocation rules, or preferred choices for task processing modes. Based on this feedback, the system can revise, improve, or verify the skill information and formally store it in the knowledge base.

[0120] In some embodiments, a knowledge base is a structured information storage system used to permanently store and manage verified skill information, domain knowledge, user preferences, and so on. The process of knowledge accumulation is dynamic; through continuous user feedback and system learning, the skill information in the knowledge base is constantly updated and optimized, thereby enabling the continuous evolution of the intelligent agent system.

[0121] In some embodiments, the hierarchical memory model will also perform cross-layer transfer of memory, which includes at least one of the following: The migration from the transient memory layer to the short-term memory layer occurs when the conversation ends or a preset message limit is reached. The first agent scores the message sequence in the transient memory layer by attention, extracts high-attention information such as user intent keywords, task trigger phrases, and / or sentiment markers, compresses it into a conversation summary, and then transfers it to the short-term memory layer. For example, low-attention casual conversation content (such as greetings and transitional phrases) is filtered and discarded.

[0122] The migration from the short-term memory layer to the medium-term memory layer involves the first agent recognizing the repetitive behavior pattern when the frequency of a certain type of task in the short-term memory layer exceeds a preset threshold. The agent extracts behavioral features such as meeting time, minutes format, and / or distribution targets of user preferences, and generates a structured preference vector to be transferred to the medium-term memory layer. The original session details in the short-term memory layer are entered into the forgetting queue after the preset retention period (e.g., seven days) is met.

[0123] The migration from the intermediate memory layer to the long-term memory layer is as follows: when the first agent completes multiple task schedulings based on the preference vector of the intermediate memory layer and obtains positive feedback from the user (e.g., no user intervention is triggered, or the user actively confirms delivery), the preference vector is marked as a stable mode. Combined with the user profile dimensions of the first user (position, department, business domain) and the team's business knowledge (project background, collaboration specifications), it is merged into a long-term business knowledge graph node and migrated to the long-term memory layer; preference vectors that do not obtain positive feedback are downgraded or deleted.

[0124] The migration from the long-term memory layer to the persistent memory layer involves the first agent solidifying the user profile and work style in the long-term memory layer into a persistent identity identifier and style template, which are then written into the persistent memory layer, once the user profile and work style remain stable and are reused multiple times within a preset period. This migration process requires explicit user confirmation or system security audit to ensure the security and immutability of the identity and style data.

[0125] The task memory layer interacts bidirectionally with each other. It not only receives task-related dialogue fragments from the short-term memory layer, but also actively queries business knowledge in the long-term memory layer to supplement the task background. It also feeds back the task execution results (success / failure mode, optimized execution path) to the medium-term memory layer to update the behavior model.

[0126] In some embodiments, hierarchical forgetting (or hierarchical forgetting mechanisms) includes at least one of the following: For time-related forgetting, different data retention periods are preset for each memory layer: the transient memory layer retains data until the end of the current session, the short-term memory layer retains data for a preset number of days, the medium-term memory layer retains data for a preset number of months, and the long-term memory layer retains data for a preset number of years. Data that exceeds the retention period is automatically entered into the forgetting assessment process. For data entering the forgetting assessment process, the number of times it is accessed within the assessment window is counted. If the number of accesses is lower than a preset threshold, soft deletion (marking it as invalid but recoverable) or hard deletion (physical removal) is performed. If it is referenced by the memory layer of high-priority tasks, the retention period is extended and the access weight is increased. Conflict resolution and forgetting: When newly imported data has semantic conflicts with existing data (e.g., changes in user preferences), the first agent arbitrates the conflict based on timestamps and confidence levels: new data with high confidence levels overwrites old data, and the old data is migrated to the historical version library; new data with low confidence levels is temporarily stored in the verification area, and a decision is made on whether to replace it after user feedback or multiple verifications.

[0127] In some embodiments, the hierarchical memory model of this application does not simply encode the dialogue text into a unified vector for storage. Instead, it employs multi-dimensional layering based on the information's time decay curve, business abstraction level, and / or user personalization level, with each layer using a differentiated storage structure and indexing strategy. The transient memory layer uses a circular buffer for OI read / write; the short-term memory layer uses a time-series database to support time-range queries; the medium-term memory layer uses a graph database to support behavioral pattern associations; the long-term memory layer uses a vector database to support semantic similarity retrieval; the persistent memory layer uses encrypted storage to ensure identity security; and the task memory layer uses a relational database to support transaction consistency. This hierarchical architecture enables the first agent to accurately locate the optimal memory layer based on query requirements during task recognition, reducing the computational overhead of full vector retrieval and significantly improving the response speed and contextual understanding depth of task recognition.

[0128] Thus, through attention-weighted cross-layer transfer and multi-dimensional forgetting mechanisms, the hierarchical memory model achieves dynamic evaluation of information value and optimized allocation of storage resources. Instantaneous context is efficiently converted into long-term knowledge, redundant data is promptly cleaned up, and each memory layer maintains lightweight and high precision. When scheduling tasks, the first intelligent agent can quickly reconstruct user intent, match historical preferences, and invoke business knowledge based on multi-dimensional memory collaboration, fundamentally solving problems such as context loss, high retrieval noise, and insufficient personalization caused by single-vector memory.

[0129] In some of the embodiments described above in this application, this application further proposes a shadow mode A / B testing mechanism before skills are added to the database, as well as a skill decay and archiving mechanism based on usage frequency and scenario changes.

[0130] Specifically, shadow mode is a parallel verification mechanism where new skills run concurrently with old skills before officially taking effect. However, the output of the new skill is only used for comparative evaluation and does not actually affect task delivery. A / B testing involves randomly assigning task traffic to the old skill path (Group A) and the new skill path (Group B), comparing the performance differences between the two groups in dimensions such as execution success rate, user intervention rate, execution time, and user satisfaction. Skill decay refers to the continuous evaluation of the value of skills already in the database. When the usage frequency of a skill decreases or its applicable scenarios change, its retrieval priority is reduced or it is removed from the active knowledge base. The archiving mechanism involves transferring decayed skills to a historical archive, retaining traceability but no longer participating in regular scheduling.

[0131] For example, shadow mode A / B testing includes at least one of the following: A skill candidate pool is constructed, and the third agent puts the filtered high-value skills into the candidate pool and marks them as pending verification. The candidate skills retain complete extraction source, filtering score and / or applicable scenario label.

[0132] Shadow binding: When the first agent schedules target tasks of the same type as candidate skills, it allocates the tasks to two execution paths according to a preset split ratio (e.g., 90% for group A and 10% for group B): group A uses the old skill to execute and deliver normally, while group B uses the new skill to execute but the output enters the shadow buffer and is not submitted to the dialog box.

[0133] Evaluation metrics are collected by the third intelligent agent for groups A and B respectively: execution success rate (whether the task is completed), anomaly rate (whether anomaly alerts are triggered), user intervention rate (whether the user edits / rejects / re-executes), execution time (the time difference from scheduling to delivery), and / or implicit satisfaction (whether the user actively confirms / likes / forwards the deliverables). For significance determination, after the third agent runs the preset verification cycle (e.g., 7 days or 50 task executions), the evaluation indicators of groups A and B are statistically significant. If group B is significantly better than group A in the first type of indicator (e.g., reduced user intervention rate, shortened execution time) and the second type of indicator (indicators without side effects (e.g., increased abnormality rate)) deteriorates, the new skill is deemed to have passed the verification. If group B is significantly worse than group A, the new skill is deemed to have failed the verification, is discarded, and the reason for failure is recorded. If the difference is not significant, the verification cycle is extended or the proportion of group B is increased.

[0134] A full switch is performed. New skills that pass verification are updated by a third agent. The first and second agents subscribe to these updates and add them to the active skill library. Old skills are moved to the historical version library. New skills that fail verification are marked as obsolete, and their extraction source and reasons for failure are fed back to the third agent to optimize the filtering rules.

[0135] In some embodiments, the skill decay mechanism in the knowledge base includes at least one of the following: Time decay: The third agent sets an initial validity period (e.g., 6 months) for each skill in the database. If the skill is not executed or called after a preset period (e.g., 1 month), the time decay coefficient decreases. When the time decay coefficient is lower than a preset threshold (e.g., 0.3), the skill enters the decay evaluation state.

[0136] Frequency decay: The third agent counts the number of times a skill is called within a sliding window period (e.g., the last 3 months); if the number of calls is lower than a preset percentage of the average value of similar skills (e.g., 20%), the frequency decay coefficient decreases; the frequency decay coefficient of frequently called skills remains or increases.

[0137] Scene change decay: The third intelligent agent monitors the matching degree of the applicable scene tags of skills: When the enterprise business system is upgraded, the API interface is changed, or the user's organizational structure is adjusted, the tool calling rules or execution preferences on which the skill depends change, the scene matching degree decreases and the scene decay coefficient decreases; the third intelligent agent automatically identifies scene changes by comparing the tool version number, API signature, user role mapping table in the skill information with the current system status.

[0138] The third agent calculates a comprehensive skill decay score based on a weighted sum of time decay coefficient, frequency decay coefficient, and scene decay coefficient. Skills with scores above the activity threshold (e.g., 0.6) remain active. Skills with scores between the activity threshold and the archiving threshold (e.g., 0.3) are marked as low-frequency skills, their retrieval ranking weight is reduced, but they can still be invoked. Skills with scores below the archiving threshold are transferred to the historical archive and stop participating in regular retrieval and scheduling.

[0139] In some embodiments, skills in the historical archive support a revival mechanism: when a third agent identifies a requirement that is highly similar to the archived skill in a new task scenario (e.g., the vector similarity exceeds a preset threshold), or when the enterprise system rolls back to an old version, causing the old skill to be re-adapted, the third agent initiates a revival confirmation to the first user or system administrator; after confirmation, the archived skill is restored to an active state, the decay coefficient is reset, and it enters a new life cycle.

[0140] In some embodiments, both skill update events and decay events are broadcast asynchronously via an event bus. The first agent queries the real-time index of the active skill library during task scheduling to ensure that scheduling decisions are based on the latest skill status. The second agent checks the active status of the skills used before executing the task. If the skill has been archived, a skill missing exception is triggered, and the first agent reschedules a replacement skill or downgrades it to the basic execution mode.

[0141] Thus, through the shadow mode A / B testing mechanism, new skills undergo parallel verification under real task traffic before being officially deployed, reducing the systemic risks that may arise from directly replacing old skills with untested ones, and ensuring that every skill in the knowledge base has been tested in practice. The skill decay and archiving mechanism enables the elimination and updating of the knowledge base, ensuring that active skills always maintain high timeliness and high retrieval value, while low-frequency outdated skills are promptly cleaned up, reducing retrieval noise and storage waste caused by knowledge base expansion. The resurrection mechanism provides historical knowledge with traceability and reusability, preventing the permanent loss of skills due to temporary changes in scenarios. This closed-loop skill lifecycle management makes the knowledge accumulation of the third agent more intelligent, ensuring that the task scheduling of the first agent and the task execution of the second agent are always based on the optimal, latest, and most reliable knowledge assets, achieving continuous evolution and long-term stability of system capabilities.

[0142] In no-code platforms, user business data is often distributed in a network pattern. There are often no direct foreign keys between the first data table and the target second data table; instead, they are indirectly linked through one or more intermediate data tables. For example, the "Customer Basic Information" table and the "Payment Records" table may not have a direct foreign key, but they are indirectly linked through the customer ID and manager ID in the "Customer Manager Assignment" table. If only a single table is monitored or only directly related tables are searched, such cross-table business risks will not be identified. Therefore, this embodiment proposes a cross-table semantic inference and joint monitoring mechanism based on multi-hop association paths, such as... Figure 2 As shown, the method further includes: S201: The second agent searches for at least one second data table associated with the first data table, the association including indirect associations achieved through at least one intermediate data table.

[0143] In some embodiments, when the first agent obtains a first data table from a user dialogue, the first agent invokes a second agent to construct a Table Relation Graph (TRG) starting from the first data table, and performs multi-hop association path discovery. The nodes of the TRG are data tables within the platform, and the edges represent the relationships between tables, including physical foreign key relationships and semantic relationships. The multi-hop association path discovery includes the following sub-steps: The second agent queries the database metadata to identify all tables that have a direct foreign key reference relationship with the first data table, as candidate tables for one-hop joins. For example, the direct join table of the first data table "Customer Basic Information" includes "Customer Manager Assignment" (joined via a foreign key with customer ID).

[0144] The second agent uses the first data table as the root node and performs a restricted breadth-first search (BFS) on the TRG to search for all second data tables reachable through intermediate data tables. The search depth is limited to a preset number of hops (e.g., a maximum of 2 or 3 hops to avoid excessively long paths that dilute business semantics). For example, starting from "Customer Basic Information," the path can reach "Payment Records" (second data table) via "Customer Manager Assignment" (intermediate table), forming a 2-hop association path of Customer Basic Information → Customer Manager Assignment → Payment Records; similarly, "Contract Signing Records" can be reached, etc.

[0145] The second agent does not consider all physically reachable paths as valid business associations. Instead, it invokes a large language model to perform a semantic score on each multi-hop path. Specifically, the second agent concatenates the table names and key field names of all data tables along the path into a path description text (such as basic customer information, which may include, but is not limited to, customer ID, account manager assignment, customer ID, manager ID, payment records, etc.). The large language model then evaluates the path's rationality in the real business context and outputs a semantic coherence score between 0 and 1. Only when the semantic coherence score exceeds the eleventh threshold is the path retained as a valid association path.

[0146] The second agent returns the retained valid association paths to the first agent. Each path is clearly labeled with the first data table, intermediate data table (if any), second data table, and path hop count. The first agent displays the association paths in the form of a topology diagram in the user dialog box (e.g., customer basic information → account manager assignment → payment record) for the user to confirm or add / delete.

[0147] In other embodiments, the second agent directly selects the Top-M paths with the highest semantic coherence scores as the determined paths.

[0148] S202: When the second agent analyzes the table name and table fields of the first data table, it analyzes the table name and table fields of the at least one second data table, determines the semantic mapping relationship of the table fields between the first data table and the at least one second data table, and determines whether there is a cross-table business type.

[0149] For example, while analyzing the table names and fields of the first data table, the second intelligent agent simultaneously analyzes the table names and fields of the intermediate data table and the second data table to determine the semantic mapping relationship of the table fields and to determine whether there is a cross-table business type based on the complete path semantics. The intermediate data table can be a table that associates the first and second data tables, and it can belong to one of the second data tables. Of course, the first data table can also be directly associated with the second data table.

[0150] In some embodiments, the second agent does not analyze each data table in isolation, but instead takes the metadata of all data tables along the entire association path as input and calls a large language model to perform chained semantic inference. The chained semantic inference includes the following sub-steps: Intermediate tables serve as semantic bridges in multi-hop paths, and their table names and field names often directly reveal the nature of the business relationship between the first and second tables. For example, the intermediate table "Customer Manager Assignment" contains fields for assignment date and service status. Based on this, the large language model recognizes that the intermediate table represents a service assignment relationship between customers and managers, rather than a simple ID mapping. The second agent labels these semantic features of the intermediate table as bridging semantic tags (such as service assignment, attribution mapping, and time-related associations).

[0151] The second agent constructs a semantic mapping matrix covering the entire path, including not only direct field alignments between adjacent tables but also indirect semantic mappings across intermediate tables. For example: Direct mapping layer: Customer Basic Information.Customer ID ↔ Customer Manager Assignment.Customer ID (Primary Key-Foreign Key Alignment); Customer Manager Assignment.Manager ID ↔ Payment Record.Manager ID (Bridge Alignment).

[0152] Indirect mapping layer: Customer Basic Information.Customer ID ⇝ Payment Record.Manager ID (Business semantic mapping through intermediate table, indicating the payment collections handled by the manager of this customer).

[0153] The elements of the semantic mapping matrix include: source field, target field, mapping type (direct / indirect), bridging table (if any), semantic similarity score, and business semantic explanation.

[0154] The large language model infers the cross-table business type expressed by the entire path based on the single-table business type of the first data table (e.g., customer relationship management), the single-table business type of the second data table (e.g., financial management), and the bridging semantic tags of the intermediate tables (e.g., service assignment). The cross-table business type includes, but is not limited to, at least one of the following: Services assigned but output delayed (Customer Relationship Management + Service Assignment + Financial Management): Customers have been assigned account managers, but there has been no payment for a long time; Ownership change not synchronized (inventory management + ownership mapping + sales management): Product location has been transferred, but sales ownership records have not been updated; Break in the timeline (project management + time-related issues + procurement management): Project milestones have expired, but procurement orders have not been followed up.

[0155] When the inference confidence level is lower than the twelfth threshold, the first agent asks the user through a dialog box about the business meaning of the associated path (e.g., you mentioned customers and payments, do you want to monitor the situation where the assigned manager is assigned but the payments are delayed?), and appends the user's answer to the prompt to re-perform the chain inference.

[0156] S203: When the cross-table business type exists, the first intelligent agent establishes a cross-table monitoring task and generates cross-table monitoring rules based on the foreign key relationship or semantic mapping between the first data table and the at least one second data table.

[0157] In some embodiments, the first agent displays the inferred cross-table business type, related path topology, and monitoring dimensions to the user for confirmation. The monitoring dimensions include, but are not limited to, at least one of the following: Cross-table monitoring name: such as "Number of customers who have been assigned an account manager but have not made payments within 30 days"; Related path topology: such as "customer basic information → account manager assignment → payment record"; Chained JOIN conditions: Multi-table JOIN statements automatically generated based on foreign key chains or semantic mapping chains. For example, if a physical foreign key chain exists, it will be automatically generated as follows: If there is no physical foreign key between the first data table and the second data table, but there is a high-confidence indirect semantic mapping (such as the semantic similarity score of the bridge through the intermediate table exceeding the thirteenth threshold), the second agent automatically generates a JOIN condition based on the semantic mapping and marks the JOIN as a semantic bridging JOIN (distinguished from a physical foreign key JOIN) in the monitoring rules.

[0158] Cross-table aggregation metrics: such as "number of customer records meeting the criteria", "number of managers involved", "estimated total annual revenue", etc. Recommended data collection frequency and threshold: If data is collected daily, a prompt should be triggered when the number of customers increases by ≥3 compared to the previous day.

[0159] After user confirmation or fine-tuning, the first intelligent agent establishes a cross-table monitoring task. The cross-table monitoring task has an independent lifecycle, and the query statements in its monitoring rules are generated cascaded based on the foreign key relationships or semantic mapping relationships between adjacent tables on the association path, automatically processing the bridging fields of intermediate data tables.

[0160] S204: When performing cross-table monitoring tasks, the second agent uses a distributed snapshot algorithm to ensure the time consistency of data snapshots across multiple tables, and triggers monitoring alerts based on changes in cross-table aggregated metrics.

[0161] In some embodiments, cross-table monitoring involves a first data table, at least one intermediate data table, and a second data table. If each table takes snapshots independently, the logically identical business status may be fragmented due to the time difference in the collection (for example, the intermediate table has updated the service status, but the second table has not yet synchronized the new repayment record, resulting in a misjudgment of "allocated but no repayment"). To address this, the second intelligent agent adopts the following multi-table cascaded distributed snapshot consistency mechanism.

[0162] In some embodiments, before performing cross-table snapshot collection, the second agent requests a Global Snapshot Version (GSV) from the platform's unified transaction coordinator or obtains a Logical Timestamp (LT) based on the global clock. The GSV or LT serves as the consistency anchor point for this cross-table snapshot, and all participating tables use this anchor point as the collection benchmark.

[0163] In some embodiments, the second agent collects snapshots of each table sequentially according to the association path topology (e.g., customer basic information → account manager assignment → payment record) in the cross-table monitoring rules, either in the direction of reverse dependency or in the topological sorting order. Specifically, it first collects snapshots of tables without upstream dependencies (e.g., "Customer Basic Information"), then collects snapshots of tables that depend on upstream dependencies (e.g., "Account Manager Assignment"), and finally collects snapshots of end tables (e.g., "Payment Record"). Each table query is accompanied by a version constraint to ensure that the committed data state at the GSV / LT anchor point is read.

[0164] In some embodiments, the second agent packages and stores data snapshots of the first and second data tables into a Cascaded Snapshot Group (CSG) of the snapshot storage module, using the same GSV / LT as the index key. The CSG is marked as an atomic unit at the storage level: if snapshot collection of any table fails or version verification fails, the entire CSG is marked as invalid, and the current monitoring task skips the comparison or executes based on the last valid CSG.

[0165] The second agent reads the last valid CSG and compares it with the current CSG. Since both snapshot groups are collected based on strict time consistency anchors, their cross-table aggregated metrics (such as the number of customers assigned to account managers but with no payments within 30 days) are logically comparable. The second agent calculates the absolute change and / or percentage change rate of the cross-table aggregated metrics. When the change exceeds the fourteenth threshold preset in the cross-table monitoring rules, the first agent determines to output a cross-table monitoring prompt.

[0166] In some embodiments, when the first agent outputs monitoring alerts, it not only alerts to changes in cross-table aggregated metrics, but also performs differential attribution based on cascaded data snapshots in the CSG. For example, when the number of "assigned customers with no repayments" increases, the first agent calls a large language model to analyze which table's change drove the metric change (e.g., because 5 new customer manager assignment records were added yesterday, but the corresponding repayment records were not generated synchronously), and outputs the attribution conclusion along with the monitoring alert.

[0167] For example, in a conversation with the first intelligent agent, a user might mention: "The clients I recently assigned an account manager for don't seem to be making good payments. Could you keep an eye on them for me?" The first intelligent agent recognizes that the core entities the user is concerned about are clients and payments, and then invokes the second intelligent agent to perform multi-hop association path discovery.

[0168] The second agent constructs a TRG and executes BFS: The first data table "Customer Basic Information" has no direct foreign key relationship to "Payment Records", but it can reach the intermediate table "Customer Manager Assignment" (via Customer ID) via one hop, and then reach the second data table "Payment Records" via another hop (via a composite association of "Manager ID" and "Customer ID") via one hop, forming a two-hop path. The large language model performs semantic scoring on this path, identifying the semantic coherence of the table names (Customer → Manager Assignment → Payment), with a coherence score of 0.92, exceeding the threshold, and is retained as a valid path.

[0169] The second agent performs chained semantic inference: the allocation date and service status of the intermediate table "Account Manager Assignment" are marked as service assignment bridging semantics; the indirect mapping "Customer Basic Information.Customer ID" → "Payment Record.Manager ID" is identified as "the customer is associated with payment through its assigned manager"; the cross-table business type is inferred as "the account manager has been assigned but the payment is delayed".

[0170] The first intelligent agent displays the suggested cross-table monitoring rules: "Monitor the number of customers who have been assigned an account manager for more than 30 days but have not made payments in the last 30 days," with the associated path being "Customer Basic Information → Account Manager Assignment → Payment Records." The suggested threshold is to prompt an alert when the number of customers increases by ≥3. Users can confirm with one click.

[0171] The first agent establishes a cross-table monitoring task, with the task ID containing the user ID, channel ID, and path hash identifier. The query statements in the monitoring rules automatically include two-level JOINs and intermediate table bridging conditions.

[0172] The next day, when the task was triggered, the second agent requested GSV=20260524080000. It collected snapshots of three tables in this version according to the topology: first, "Customer Basic Information," then "Customer Manager Assignment," and finally "Payment Records," packaging them into a CSG and storing it in the snapshot storage module. Comparing the CSG from the previous day, it was found that the number of customers meeting the criteria increased from 8 to 15, an absolute change of 7, exceeding the threshold of 3. The first agent generated a cross-table monitoring prompt: "7 new customers with assigned customer managers but no payments within 30 days have been added compared to yesterday, involving 5 customer managers. It is recommended to prioritize investigating the following customers..." along with an attribution analysis that the increase was mainly due to 5 new assignment records added to the "Customer Manager Assignment" table yesterday that have not yet generated corresponding payments.

[0173] This embodiment overcomes the technical limitations of searching only directly related tables by introducing a data table association graph and restricted multi-hop path search, enabling cross-table monitoring to cover network data relationships. Through semantic bridging identification of intermediate data tables and chain-like semantic inference, it solves the problem of business semantic breaks in indirect association scenarios, extending cross-table business type inference from "direct connection between two tables" to "multi-table cascading." Through cascading snapshot groups and global consistency anchors, it ensures strict consistency of multi-table snapshots, including intermediate tables, in the time dimension, eliminating distortion of cross-table aggregation indicators caused by differences in collection timing. Ultimately, it achieves accurate identification and attribution warning of risks of "local normality and global anomaly" in complex business networks.

[0174] like Figure 3 As shown, a skill accumulation system based on a multi-agent system is disclosed. The multi-agent system includes a first agent, a second agent, and a third agent. The skill accumulation system based on the multi-agent system includes: The module 301 is used to call the second agent to perform semantic inference based on the table name and field information of the first data table when the first agent obtains the first data table from the user's dialogue, so as to determine the information to be monitored. Module 302 is used by the first intelligent agent to establish monitoring tasks and monitoring rules based on the information to be monitored. The output module 303 is used to collect the current data snapshot according to the monitoring rules when the second intelligent agent is called by the first intelligent agent, and to determine whether to output a monitoring prompt based on the comparison result of the current data snapshot and the historical data snapshot. The prompt module 304 is used to output a monitoring prompt in a dialog box with the user when a monitoring prompt needs to be output. The writing module 305 is used for the third intelligent agent to listen to the scheduling process of the first intelligent agent and the monitoring process of the second intelligent agent in an asynchronous non-blocking manner, and to extract business knowledge and write it into the knowledge base; The patch module 306 is used by the third intelligent agent to collect application trajectory data of the user using multiple skills in a first time period, aggregate and analyze the application trajectory data in a second time period to generate a skill patch set, and determine whether to push for approval based on the confidence level of each skill in the skill patch set. Based on the confidence level and the approval result, the patch in the skill patch set is added to the database and installed. In this case, the frequency of the user using skills in the first time period is higher than the frequency of the user using skills in the second time period.

[0175] In some embodiments, the system can execute any of the aforementioned agent-based task processing methods, which will not be listed here; and the technical effects are the same as those in the method embodiments, which will not be described again here.

[0176] To achieve the above objectives, embodiments of the present invention also provide an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 501 and a memory 503 connected to the processor 501 via a communication bus 502; wherein, the memory 503 is used for executing programs; the processor 501, through the execution of the executable program, can implement the information processing method based on a multi-agent system provided in any of the foregoing embodiments.

[0177] Optionally, the processor 501 may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. Here, the program executed by the processor 501 may be stored in a memory 503 connected to the processor 501 via a communication bus 502. The memory 503 may be volatile memory or non-volatile memory, or may include both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); magnetic surface memory can be disk storage or magnetic tape storage. Volatile memory can be random access memory (RAM), which is used as an external cache.By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), Sync Link Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM). The memory 503 described in this embodiment of the invention is intended to include, but is not limited to, these and any other suitable types of memory 503. The memory 503 in this embodiment of the invention is used to store various types of data to support the operation of the processor 501. Examples of this data include: any computer programs operated by the processor 501, such as operating systems and applications; contact data; phonebook data; messages; pictures; videos, etc. The operating system contains various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks.

[0178] In some embodiments of the present invention, the memory 503 may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memory 503 of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0179] The processor 501 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 501 or by software instructions. The processor 501 can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory 503, and the processor 501 reads the information in memory 503 and, in conjunction with its hardware, completes the steps of the above method. In some embodiments, the embodiments described herein can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described herein, or combinations thereof.

[0180] For software implementation, the techniques described herein can be achieved through modules (e.g., procedures, functions, etc.) that perform the functions described herein. The software code can be stored in memory and executed by a processor. The memory can be implemented within the processor or externally.

[0181] Another embodiment of the present invention provides a computer storage medium storing an executable program. When executed by a processor 501, the executable program can implement the steps of an information processing method based on a multi-agent system applied to the computing device. For example, such as... Figures 1-3 One or more of the methods shown.

[0182] In some embodiments, the computer storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0183] It should be noted that the technical solutions described in the embodiments of the present invention can be combined arbitrarily without conflict.

[0184] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. An information processing method based on a multi-agent system, characterized in that, The multi-agent system includes a first agent, a second agent, and a third agent, and the method includes: When the first intelligent agent obtains the first data table from the user's dialogue, the first intelligent agent calls the second intelligent agent to perform semantic inference based on the table name and field information of the first data table in order to determine the information to be monitored. The first intelligent agent establishes monitoring tasks and monitoring rules based on the information to be monitored; The second intelligent agent, under the invocation of the first intelligent agent, collects the current data snapshot according to the monitoring rules, and determines whether to output a monitoring prompt based on the comparison result between the current data snapshot and the historical data snapshot; If it is determined that a monitoring prompt should be output, the first intelligent agent outputs the monitoring prompt in the dialog box with the user. The third intelligent agent listens to the scheduling process of the first intelligent agent and the monitoring process of the second intelligent agent in an asynchronous non-blocking manner, and extracts business knowledge and writes it into the knowledge base; The third intelligent agent collects application trajectory data of the user using multiple skills in the first time period, aggregates and analyzes the application trajectory data in the second time period to generate a skill patch set, and determines whether to push for approval based on the confidence level of each skill in the skill patch set. Based on the confidence level and the approval result, the agent completes the storage and installation of the patches in the skill patch set. In this case, the frequency of the user using skills in the first time period is higher than the frequency of the user using skills in the second time period.

2. The method according to claim 1, characterized in that, The second intelligent agent performs semantic inference based on the table name and field information of the first data table, including: The second intelligent agent infers the business type of the first data table based on the table name and field name of the first data table using a large language model (LLM). The business type includes at least one of customer relationship management, sales management, project management, and inventory management. The second intelligent agent classifies the fields based on the LLM, and the classification includes at least one of time-based fields, status-based fields, numerical fields, and personnel-based fields; The second agent outputs the monitoring information based on the LLM; the monitoring information includes at least one suggested monitoring dimension, and the at least one monitoring dimension includes at least one of monitoring name, associated field, suggestion condition and suggestion frequency; When the confidence level of the LLM inference is lower than the first threshold, the first agent asks the user through a dialog box about the business purpose of the first data table, and adds the user's answer to the prompt words to re-infer.

3. The method according to claim 1 or 2, characterized in that, The second agent determines whether to output monitoring prompts based on the comparison results, including: The second intelligent agent collects indicators related to the monitoring rules and executes query statements at a preset frequency to collect at least one of the following indicators: the number of records that meet the monitoring conditions, the aggregate value of the target field, and the count of each category of the status field. The current data snapshot is compared with the previous historical data snapshot to determine the change information; When the change information exceeds the second threshold preset in the monitoring rules, the first agent determines that the monitoring prompt needs to be output.

4. The method according to claim 1 or 2, characterized in that, The method further includes: The first intelligent agent records the effect of the monitoring prompts, wherein the effect includes at least: recording the total number of prompts and user feedback; the user feedback includes at least one of the number of times the user viewed the prompts, the number of times the user ignored the prompts, and the number of times the user closed the prompts. The first intelligent agent updates the prompt weight in the monitoring rules according to the prompt effect, wherein the prompt weight is positively correlated with the number of times the user views it, and negatively correlated with the number of times the user ignores it and the number of times the user closes it. When the prompt weight is lower than the third threshold, the first agent pauses the corresponding monitoring task or monitoring prompt. When the prompt weight is higher than the fourth threshold, the first agent increases the monitoring frequency of the corresponding monitoring task or the prompt and frequency of the monitoring prompt, where the fourth threshold is greater than the third threshold.

5. The method according to claim 1 or 2, characterized in that, The application trajectory data includes at least one of the following: user identifier, skill identifier, input content, output content, and a marker indicating whether the user corrected the agent's response; The process of aggregating and analyzing the application trajectory data to generate a skill patch set during the second time period includes: The third intelligent agent identifies the operation sequence in multiple successful application trajectories of the user through a clustering algorithm. When the same operation sequence appears in more than a preset number of successful application trajectories, it is marked as a successful mode candidate. The third intelligent agent compares successful application trajectories and failed application trajectories in the same scenario using a sequence alignment algorithm, identifies operation steps that exist in successful application trajectories but are missing in failed application trajectories, and extracts the missing operation steps as skill content to be added. Based on the skill content to be added, generate the skill patch set.

6. The method according to claim 1 or 2, characterized in that, The method further includes: Before installing the skill patch, the third intelligent agent performs static code analysis on the files in the skill directory and matches them against a preset risk rule base. When a hard trigger rule is hit, the installation is directly blocked. When the security score of the third intelligent agent's computing skills is lower than the fifth threshold, the first intelligent agent displays risk details on the installation interface and prompts the user to confirm each item before continuing the installation. The third agent executes the execution script in the skill in an independent sandbox container. The third agent detects abnormal behavior in real time and interrupts the installation execution when abnormal behavior is detected. The abnormal behavior includes at least one of the following: non-compliant token pass-through, unauthorized data transmission behavior, and unauthorized access attempt. The third intelligent agent scans the installed skills at a preset period. When a security risk is detected, it automatically notifies the skill owner and generates a security audit report containing a list of risky skills and remediation suggestions.

7. The method according to claim 1 or 2, characterized in that, The method further includes: After receiving the user's skill query instruction, the first agent invokes the third agent to perform a multi-path parallel search; the multi-path parallel search includes at least two of the following: The user query information is augmented to obtain a semantic vector. The semantic vector is then compared with the vector index in the knowledge base to calculate the similarity and obtain the first number of knowledge fragments with the highest similarity. The user query information is segmented into words, and the second most relevant knowledge fragments are obtained by matching the inverted index. The entities and relationships in the knowledge base are constructed into a knowledge graph, and a third number of knowledge fragments are retrieved that are associated with the entities in the user query information. The third agent deduplicates and merges the knowledge fragments acquired from multiple sources, then calls the rearrangement model to calculate the relevance score for each knowledge fragment, sorts them in descending order of relevance score, and returns the top K knowledge fragments.

8. The method according to claim 1 or 2, characterized in that, The method further includes: The second agent searches for at least one second data table associated with the first data table; The second intelligent agent, after analyzing the table name and table fields of the first data table, analyzes the table name and table fields of the at least one second data table, determines the semantic mapping relationship of the table fields between the first data table and the at least one second data table, and determines whether there is a cross-table business type. When the cross-table business type exists, the first intelligent agent establishes a cross-table monitoring task and generates cross-table monitoring rules based on the foreign key relationship or semantic mapping between the first data table and the at least one second data table. When performing the cross-table monitoring task, the second agent uses a distributed snapshot algorithm to ensure the temporal consistency of data snapshots across multiple tables, thereby aggregating changes in metrics across tables.

9. A skill accumulation system based on a multi-agent system, characterized in that, The multi-agent system includes a first agent, a second agent, and a third agent, and the skill accumulation system based on the multi-agent system includes: The calling module is used to call the second agent to perform semantic inference based on the table name and field information of the first data table when the first agent obtains the first data table from the user's dialogue, so as to determine the information to be monitored. A module is established for the first intelligent agent to establish monitoring tasks and monitoring rules based on the information to be monitored. The output module is used to collect the current data snapshot according to the monitoring rules when the second agent is called by the first agent, and to determine whether to output a monitoring prompt based on the comparison result of the current data snapshot and the historical data snapshot. The prompt module is used to enable the first intelligent agent to output a monitoring prompt in a dialog box with the user when a monitoring prompt is required. The writing module is used by the third intelligent agent to listen to the scheduling process of the first intelligent agent and the monitoring process of the second intelligent agent in an asynchronous non-blocking manner, and to extract business knowledge and write it into the knowledge base; The patch module is used by the third intelligent agent to collect application trajectory data of the user using multiple skills in a first time period, aggregate and analyze the application trajectory data in a second time period to generate a skill patch set, and determine whether to push for approval based on the confidence level of each skill in the skill patch set. Based on the confidence level and the approval result, the patch in the skill patch set is added to the database and installed. In this case, the frequency of the user using the skills in the first time period is higher than the frequency of the user using the skills in the second time period.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, characterized in that the computer program, when executed by a processor, implements the method according to any one of claims 1 to 8.

Images