Multi-agent cooperative data processing method and system, and storage medium

Abstract

Embodiments of the present application provide a multi-agent collaborative data processing method and system, and a storage medium, which belong to but are not limited to the technical field of data analysis and processing. The method comprises the following steps: after obtaining a target state corresponding to a target data processing task, a central scheduling agent calls a state machine control layer to check the target state; and after the checking result indicates that the check is passed, a subtask corresponding to the target state is published to enable each professional agent to actively collect the matched target subtask and execute it. The checking result includes a state flow compliance checking result and a general business precondition checking result. The state flow compliance checking result is obtained by checking the target state according to a state machine topology and a current state. The general business precondition checking result is obtained by checking the target state according to a target general business precondition in a precondition constraint set. Embodiments of the present application can safely complete a cross-domain data processing task under a safety boundary by using a remote control agent.

Description

Technical Field

[0001] This application relates to the field of data analysis and processing technology, and in particular to a multi-agent collaborative data processing method, system and storage medium. Background Technology

[0002] With the development of Large Language Model (LLM) technology, LLM-based AI agents are widely used in various automated task processing scenarios, especially for enterprise data processing. However, enterprise data processing often involves silos of data from multiple different sources. Furthermore, because core business data within enterprises must adhere to strict data security policies and compliance requirements, multi-agent collaborative operations can only process necessary metadata or anonymized data. Moreover, enterprise data processing is often performed in remote execution environments. Therefore, existing multi-agent collaborative data processing methods for enterprise data further increase the uncertainty of the process, resulting in poor data processing security. Therefore, how to securely complete cross-domain data processing tasks using remotely controlled agents while ensuring security boundaries for internal enterprise data is a pressing technical problem that needs to be solved. Summary of the Invention

[0003] The main objective of this application is to propose a multi-agent collaborative data processing method, system, and storage medium that can securely complete cross-domain data processing tasks using remotely controlled agents while ensuring that security boundaries are met for internal enterprise data.

[0004] To achieve the above objectives, a first aspect of this application proposes a multi-agent cooperative data processing method, the method comprising: In response to obtaining the target state for the target data processing task, the central scheduling agent calls the state machine control layer to perform a precondition check on the target state and obtain the check result. The state machine control layer is configured with a state machine topology and a set of preconditions. The state machine topology includes multiple task states, and the set of preconditions includes at least one general business precondition. At least some of the general business preconditions are set to correspond to the task states. If the inspection result indicates that the verification has passed, the central scheduling agent issues a sub-task corresponding to the target state; Each specialized intelligent agent selects a matching target subtask from the published subtasks and executes the target subtask. The inspection results include state transition compliance inspection results and general business pre-inspection results. The state transition compliance inspection results are obtained by verifying the target state based on the state machine topology and the current state. The general business pre-inspection results are obtained by verifying the target state based on the target general business pre-conditions in the pre-constraint set that match the target state. The current state of the target data processing task is updated to the target state when the inspection results indicate that the state transition compliance inspection and the operation constraint inspection have passed.

[0005] To achieve the above objectives, a second aspect of this application proposes an intelligent agent cooperative data processing system, comprising: An application interaction layer is used to obtain data processing requests from users. A multi-agent collaboration layer, comprising a central scheduling agent and multiple specialized agents, wherein the central scheduling agent is used to create a target data processing task based on the data processing request; A state machine control layer is used to store the state machine topology and a set of pre-constraint conditions. The state machine control layer responds to the verification request from the central scheduling agent for a target state, performs pre-constraint checks on the target state, and sends the check results to the central scheduling agent. The check results include state transition compliance check results and general business pre-constraint check results. The state transition compliance check results are obtained by verifying the target state based on the state machine topology and the current state. The general business pre-constraint check results are obtained by verifying the target state based on target general business pre-constraint conditions that match the target state in the pre-constraint set. The current state of the target data processing task is updated to the target state when the check results indicate that both the state transition compliance check and the operation constraint check have passed. Data access call layer; Wherein, if the inspection result indicates that the verification is passed, the central scheduling agent issues a sub-task corresponding to the target state; each of the specialized agents selects a matching target sub-task from the issued sub-tasks and executes the target sub-task; The data access and invocation layer is used to provide each of the specialized intelligent agents with the data sources and tool resources required to execute the corresponding target sub-tasks.

[0006] To achieve the above objectives, a third aspect of the present application provides a computer-readable storage medium storing computer-executable instructions, which are configured to be executed by a processor as described in any one of the first aspects of the multi-agent cooperative data processing method.

[0007] The multi-agent collaborative data processing method, system, and storage medium proposed in this application impose hard constraints on the task flow of the central scheduling agent in executing the target data processing task through state transition rules defined in the state machine topology. Simultaneously, it verifies the operational constraints before the transition of the target state through a set of pre-constraint conditions, further confirming the acceptance of the state of the next node in the target data processing task. Furthermore, since the central scheduling agent only issues sub-tasks based on the verified target state and does not involve the allocation of sub-tasks to various specialized agents, the entire data processing flow of the central scheduling agent and various specialized agents collaboratively completing the target data processing task is completely controllable, and the flow from each node to the next in the data processing flow is accurate. In the event of anomalies, the anomalies can be quickly identified based on the current state. Therefore, compared with related technologies, the embodiments of this application can securely complete cross-domain data processing tasks using remotely controlled agents, while ensuring that security boundaries are met, for internal enterprise data. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the multi-agent collaborative data processing method provided in this application; Figure 2 This is a flowchart illustrating the target state verification process in one embodiment of the multi-agent collaborative data processing method provided in this application. Figure 3 This is a schematic diagram of the state topology in one embodiment of the multi-agent cooperative data processing method provided in this application; Figure 4 This is a schematic diagram of the application process of precondition checking in one embodiment of the multi-agent collaborative data processing method provided in this application; Figure 5 This is a schematic diagram illustrating the application of a security configuration policy file in one embodiment of the multi-agent collaborative data processing method provided in this application; Figure 6 This is a schematic diagram of the system structure in one embodiment of the multi-agent collaborative data processing system provided in this application; Figure 7 This is a schematic diagram of the hardware structure of the electronic device used in the application of the multi-agent collaborative data processing method provided in this application. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0010] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0011] 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0012] With the development of Large Language Model (LLM) technology, LLM-based AI agents are widely used in various automated task processing scenarios. Existing technologies typically employ two types of multi-agent scheduling: distributed scheduling and centralized scheduling. In distributed scheduling, agents autonomously allocate task resources through mutual negotiation and consensus to jointly achieve the overall goal, representing autonomous decision-making based on LLM. However, an LLM-based agent is essentially a probabilistic model, and its output is uncertain. Therefore, using distributed scheduling can lead to uncontrollable data processing flows (such as skipping risks, loop risks, forgetting risks, or security risks). In centralized scheduling, a central agent is typically responsible for global task decomposition, resource allocation, and task assignment; other working agents receive and execute the assigned tasks, and these working agents generally do not autonomously negotiate global decisions. Therefore, in data processing-related task scenarios, to achieve unified scheduling driven by the central agent, existing solutions typically combine workflow engines for task allocation. However, traditional workflow engines rely on pre-defined process topologies and can only execute flows according to preset steps. They cannot dynamically adjust subtasks, agent call relationships, and collaboration logic based on the real-time context of task execution, thus failing to cope with changes in actual business. Therefore, while existing technologies have proposed combining centralized scheduling with frameworks like LangGraph to leverage their flexible orchestration capabilities to adapt to business flexibility needs, the constraints of frameworks like LangGraph are "programmatic." They rely on developers correctly writing node logic, and if the logic has vulnerabilities or is bypassed by malicious prompts, the data processing flow may still lose control. This is especially true for enterprise data, which often involves siloed data from various sources (such as ERP, CRM, SCM, etc.), thus requiring greater flexibility in business adaptation. Furthermore, because core business data in enterprise systems must adhere to strict data security policies and compliance requirements, multi-agent collaborative cross-domain data operations can only process necessary metadata or anonymized data. This makes it difficult for the scheduling system to ensure the correct progress of the data processing flow without knowing the specific content of the data. Furthermore, enterprise data processing is often based on remote execution environments. However, remote execution environments are subject to uncertainties such as network latency, connection interruptions, and execution timeouts, making it impossible for the system to verify task status by directly viewing data. Therefore, existing multi-agent collaborative data processing methods for enterprise data struggle to ensure both business adaptability and the determinism and accuracy of the data processing flow, resulting in poor data processing security. Therefore, how to securely complete cross-domain data processing tasks using remotely controlled agents while ensuring security boundaries for internal enterprise data is a pressing technical problem that needs to be solved.Based on this, embodiments of this application provide a multi-agent collaborative data processing method, system, and storage medium, which can securely complete cross-domain data processing tasks using remotely controlled agents while ensuring that security boundaries are met for internal enterprise data.

[0013] Reference Figure 1 As shown, the multi-agent cooperative data processing method provided in the embodiments of this application includes: Step S100: In response to obtaining the target state for the target data processing task, the central scheduling agent calls the state machine control layer to perform a precondition check on the target state and obtain the check result. The state machine control layer is configured with a state machine topology and a set of preconditions. The state machine topology includes multiple task states, and the set of preconditions includes at least one general business precondition. At least some of the general business preconditions are set to correspond to the task states. Step S200: If the inspection result indicates that the verification has passed, the central scheduling agent issues a sub-task corresponding to the target state; Step S300: Select the matching target subtask from the published subtasks through each professional intelligent agent and execute the target subtask; The inspection results include state transition compliance inspection results and general business pre-inspection results. The state transition compliance inspection results are obtained by verifying the target state based on the state machine topology and the current state. The general business pre-inspection results are obtained by verifying the target state based on the target general business pre-conditions that match the target state in the pre-constraint set. The current state of the target data processing task is updated to the target state when the inspection results indicate that the state transition compliance inspection and the operation constraint inspection have passed.

[0014] By defining state transition rules in the state machine topology, the task flow of the central scheduling agent in executing the target data processing task is subject to hard constraints. Simultaneously, the operational constraints before the transition of the target state are verified through a set of pre-constraint conditions, further confirming the acceptance of the state of the next node in the target data processing task. Since the central scheduling agent only issues sub-tasks based on the verified target state and does not involve the allocation of sub-tasks to various specialized agents, the entire data processing flow of the central scheduling agent and various specialized agents collaboratively completing the target data processing task is completely controllable, and the transition from each node to the next in the data processing flow is accurate. In case of anomalies, the anomaly can be quickly identified based on the current state. Therefore, compared with existing centralized scheduling, this embodiment can transform the control of execution steps into control of the target, state, and result. Furthermore, compared with existing autonomous decision-making that only controls the target, this embodiment can reduce the probability of uncontrollable paths by controlling the state (including result acceptance). Therefore, this embodiment can securely complete cross-domain data processing tasks using remote control agents for internal enterprise data, while ensuring that security boundaries are met.

[0015] The target state is the execution state that is expected to be entered in the next stage, determined by the acceptance results of all subtasks that are broken down from the current state of the target data processing task.

[0016] This application does not limit how the central scheduling agent obtains the target state. In some embodiments, the target state in step S100 can be obtained by the user based on information fed back by the central scheduling agent. For example, if the feedback information indicates that the user confirms entering the survey state, then the target state can be obtained as the survey in progress state. In other embodiments, the target state can come from the various third-party specialized agents that collaboratively complete the target data processing task. For example, if the third-party specialized agent, Surveyor, actively proposes "we can enter SURVEY_CONFIRMING" after completing the survey, then the target state is SURVEY_CONFIRMING. In other embodiments, the target state can also be determined by the central scheduling agent itself. For example, if the central scheduling agent determines "we should enter DESIGNING" based on the dialogue context, then the target state is DESIGNING. In other embodiments, it can also come from a system abnormality, such as a network outage, then the target state is CONNECTION_LOST.

[0017] The multi-agent collaborative data processing method of this application can be applied to at least the processing of target data processing tasks such as data integration. In some embodiments, it can also be used for data research, data design, etc. Details regarding these aspects will not be elaborated upon in the embodiments of this application.

[0018] This application embodiment does not limit the number of target data processing tasks being executed on the central scheduling agent, nor does it limit whether the types of the executed target data processing tasks are the same. For example, if the central scheduling agent is executing two target data processing tasks, both of which are for data integration, where target data processing task 1 is used to integrate data from data sources 1 to 3, and target data processing task 2 is used to integrate data from data sources 2 and 7, this application embodiment will not elaborate on these details.

[0019] The central scheduling intelligent agent and the state machine control layer are independent applications that can be deployed on the same hardware device or on different hardware devices. This application will not elaborate on these aspects in its embodiments.

[0020] In this embodiment, each task state of the state machine topology is a discrete state obtained by dividing the life cycle of the target data processing task. The state machine topology is used to define the target state that each predefined discrete state is allowed to transition to.

[0021] Subtasks are tasks determined by the central scheduling agent to complete the target data processing task under the target state, and are dynamically generated by the central scheduling agent in real time. There can be multiple subtasks. This application embodiment does not limit how the central scheduling agent determines subtasks; those skilled in the art can selectively set them by referring to existing agent task splitting methods, which will not be elaborated upon in this application embodiment. After the central scheduling agent generates subtasks, each subtask is set with a task completion status. The task completion status indicates the acceptance and acceptance status. Each specialized agent can choose to accept subtasks with a completion status of "not accepted" or "not accepted" based on its own resources. For example, if the initial status of a subtask 1 is "accepted," after it is accepted by specialized agent 1, its task completion status is updated to "accepted," and other specialized agents cannot repeatedly accept subtask 1. When a specialized intelligent agent times out due to a sudden stagnation, unresponsiveness, repeated errors, or inability to continue during autonomous task execution, the task status of subtask 1 is updated to "acquired, but timed out." Other specialized intelligent agents then have the right to acquire subtask 1. Since different specialized intelligent agents have different execution methods, the probability of subtask 1 being blocked when executed by other subtask agents is greatly reduced. In other embodiments, if specialized intelligent agent 2 delivers subtask 2 but fails acceptance, the task status of subtask 2 is updated to "completed, but failed acceptance for [reason]." In this case, the central scheduling intelligent agent can generate an automatic repair task bound to subtask 2 for this failure. Other specialized intelligent agents can simultaneously acquire the repair task when acquiring subtask 2. This allows subtask 2 to be flexibly handled according to the actual situation. In other embodiments, taking a research-focused intelligent agent as an example, when investigating the cause of a failed execution, it discovers that key fields in the original data source indeed prevent the generation of the research report. The central scheduling agent then generates a new task based on the research-focused agent's cause analysis, which is then executed by other agents. At this point, if the research-focused agent determines that the requirements for receiving the task are met, it will re-receive the task for execution. Compared to centralized scheduling, since anomalies occurring during the execution of the target data processing task are unpredictable, they cannot be pre-configured in the workflow engine. Using a workflow engine in this scenario would lead to the failure of the entire task and downstream tasks until the next round of finding a new data source, modifying the process, and restarting the task to complete. However, in this embodiment, the central scheduling agent can issue sub-tasks, and each specialized intelligent agent automatically receives the task. Furthermore, the central scheduling agent dynamically adjusts the issued sub-tasks based on the task acceptance status of each specialized intelligent agent, enabling collaborative processing between different specialized intelligent agents. This embodiment reduces the probability of restarting the target data processing task, making the target data processing process more flexible and efficient.

[0022] Specialized intelligent agents are deployed agents used to actually perform tasks. By having each specialized intelligent agent proactively select and execute matching target subtasks from the published subtasks, the central scheduling agent only needs to verify the task status and the publication of subtasks. Subtasks can be processed asynchronously and collaboratively among the specialized intelligent agents, thus better adapting to target data processing tasks with different requirements.

[0023] The "Check Result Passed" indicates that all precondition checks have passed. For example, if the precondition checks include status transition compliance checks and general business precondition checks, then both status transition compliance checks and general business precondition checks have passed.

[0024] The state transition compliance check result is the same as the state transition compliance verification result, and the general business precondition check result is the same as the general business precondition check result. State transition compliance verification and general business precondition checks can be performed independently and in parallel. In some embodiments, the general business precondition check can also be performed only after the state transition compliance verification passes. This application does not impose restrictions on this. If either the state transition compliance verification or the general business precondition check fails, updating the current state to the target state will be prohibited. In some embodiments, in the event of a verification failure, a rollback or retry can be performed based on the state machine topology. This application will not elaborate on these aspects further.

[0025] This application embodiment does not restrict where the current state of the target data processing task is stored; it can be stored in the state machine control layer or in the central scheduling agent layer.

[0026] The target general business prerequisites are used to support general business prerequisites that match the target state.

[0027] This application does not limit the specific types of task states. In some embodiments, the state machine topology includes at least the states shown in Table 1 below:

[0028] Table 1 In some embodiments, the state machine topology also includes states related to the remote control environment, as shown in Table 2 below:

[0029] Table 2 This application does not limit the form of the state machine topology in practical applications. In some embodiments, it can be... Figure 3The state machine topology of this application embodiment is stored in the manner shown. In other embodiments, it can be stored in the form of a state transition whitelist. For example, referring to the task state examples shown in Table 1, the state machine topology transition rules of this application embodiment are explained. The whitelist storage is specifically as follows: (1) CREATED (201)→SURVEYING (202), CANCELLED (213); (2) SURVEYING (202)→SURVEY_CONFIRMING (203), FAILED (212); (3) SURVEY_CONFIRMING (203)→DESIGNING (204), CANCELLED (213); (4) DESIGNING (204)→DESIGN_CONFIRMING (205), FAILED (212); (5) DESIGN_CONFIRMING (205)→IMPLEMENTING (206), DESIGNING (204), CANCELLED (213); (6) IMPLEMENTING (206)→QA_TESTING (207), FAILED (212); (7) QA_TESTING (207)→TESTING (208), IMPLEMENTING (206), FAILED (212); (8) TESTING (208)→DEPLOYING (209), IMPLEMENTING (206), FAILED (212); (9) DEPLOYING (209)→OPERATING (210), FAILED (212); (10) OPERATING (210)→COMPLETED (211), FAILED (212); (11) FAILED (212)→SURVEYING (202), IMPLEMENTING (206).

[0030] For example, referring to the task state examples in Table 2, the transition rules of the state machine topology in this application embodiment are explained as follows: (12) CONNECTING (214)→AUTHENTICATING (215), FAILED (212); (13)AUTHENTICATING (215)→SURVEYING (202), FAILED (212); (14) Any state → CONNECTION_LOST (217) (Automatically triggered when connection heartbeat timeout); (15) CONNECTION_LOST (217)→CONNECTING (214) (reconnection), FAILED (212); (16) SURVEYING (202) → SANDBOX_EXECUTING (216) (perform data sampling), SURVEY_CONFIRMING (203); Any state can transition to CANCELLED (213). Specifically, DESIGNING (204) in rule (5), IMPLEMENTING (206) in rule (7), and IMPLEMENTING (206) in rule (8) are states entered when the fallback condition is met. Corresponding constraint rules can be configured in the general business preconditions. The state in rule (11) indicates a state where retrying is allowed. CONNECTING (214) in rule (15) is the state entered when reconnection is allowed.

[0031] This application embodiment does not restrict whether a general business prerequisite is set for each task state, or whether only one general business prerequisite is set for each task state that requires a general business prerequisite. Those skilled in the art can selectively set the prerequisite according to the actual situation.

[0032] For example, taking the task status in Table 2 as an example, the general business preconditions set for the task status in this application embodiment can be shown in Table 3 below:

[0033] Table 3 For example, in some embodiments, when the task status includes both the statuses in Table 1 and the statuses in Table 2, the general business preconditions of this application embodiment also include those shown in Table 4 below:

[0034] Table 4 Understandably, the inspection results also include the flow threshold inspection results, which are determined through the following steps: If the operation constraint verification passes, obtain the cumulative number of times the target data processing task enters the same target state during execution; If the cumulative number of occurrences is less than or equal to the preset trigger threshold, the flow threshold check result is determined to be verified as passed.

[0035] By increasing the verification of cumulative counts, a hard constraint on the data processing process can be added, forming a three-layer progressive constraint control of topological constraints (state flow whitelist) + data constraints (precondition hard checks) + frequency constraints (loop detection). This further ensures that the method of this application embodiment can securely complete cross-domain data integration tasks using remote control intelligent agents for internal enterprise data, while ensuring that security boundaries are met.

[0036] Understandably, the method also includes: Upon receiving a confirmation request from the first specialized intelligent agent for target data operation of the first target subtask, the central scheduling intelligent agent obtains the target depth defense inspection conditions corresponding to the target data operation by matching from the set of pre-constraint conditions through the state machine control layer; the first target subtask is a subtask that the first specialized intelligent agent actively retrieves from the central scheduling intelligent agent. Based on the target's defense-in-depth inspection conditions, access verification is performed on target data operations.

[0037] Specifically, if the access verification results in a data constraint check indicating that the verification fails, the first professional intelligent agent will stop executing the target data operation.

[0038] For example, taking the target data operations as including data access operations, cross-data source operations, and data write operations, the set of preconditions also includes the general business preconditions shown in Table 5 below:

[0039] Table 5 For example, referring to Table 5, assuming a remotely controlled first professional intelligent agent requests access to a data source, based on the set of pre-constraints, the following checks are executed sequentially: F470 (connection check) → F471 (authentication check) → F473 (data masking check) → F474 (scope check) → F475 (audit check). If any check fails, the entire call chain is blocked, and the data access request is rejected. Thus, the five checks, constructed based on state machine topology and general business pre-constraints and data constraints, together constitute a defense-in-depth system in the "usable but invisible" scenario.

[0040] The target depth defense check conditions and access verification of target data operations can be performed by the central scheduling agent calling the corresponding interface of the state machine control layer, or the central scheduling agent can perform access verification of target data operations after obtaining the target depth defense check conditions by calling the state machine control layer. In this application embodiment, the specific details are not elaborated. Those skilled in the art can select the appropriate execution subject according to the actual situation.

[0041] Understandably, based on the target's defense-in-depth inspection criteria, access verification is performed on target data operations, including: Obtain the target security policy corresponding to the target subtask from the security policy matrix. The security policy matrix includes various data access boundary policies; each data access boundary policy corresponds to a different policy type. Access verification is performed on target data operations through target security policies and target defense-in-depth inspection conditions.

[0042] This application does not restrict the source of the security policy matrix, and those skilled in the art can selectively set it according to actual needs.

[0043] The data access boundary policy is used to limit the data that the first professional intelligent agent can access, the data that cannot be accessed, and the conditions for accessing the data. The target defense-in-depth check condition is used to limit the state of data when the task state transitions. For example, if the target data operation is a data access operation, the target defense-in-depth check condition restricts data flow to only data that has undergone de-identification processing. When the data access boundary policy restricts access to data in range A, and data of type B in range A cannot be accessed except after de-identification processing, then when the access verification for the target data operation passes, it means that the accessed data is data in range A whose data type is not B, and the data obtained after de-identification processing is in a de-identified state. This application's embodiments will not elaborate further on this.

[0044] Policy types are used to distinguish the purpose of data operations. In some embodiments, policy types include at least one such as access type (e.g., for data integration), data authentication type, and data auditing type. Different policy types are configured with different boundary policies depending on the application scenario. For example, taking access type as the policy type, access policies can be defined for three different application scenarios: production environment, testing environment, and development environment. This application does not impose limitations on this aspect.

[0045] Understandably, the method also includes: In response to the fulfillment of preset configuration loading conditions, a preset declarative configuration file is read from a preset directory; The declarative configuration file is processed to obtain configuration constraint data, which includes at least one of the following: security policy matrix, state machine topology, and precondition set.

[0046] The configuration loading conditions include at least one of the following: The declarative configuration file has been updated; Command-triggered (e.g., user-triggered, or timed-triggered); System initialization.

[0047] This application does not limit the file format of the declarative configuration file. For example, in some embodiments, the YAML file format is used.

[0048] This application embodiment does not restrict whether the security policy matrix, state machine topology, and pre-constraint condition set are all configured using declarative configuration files. There can be only one, two, or all of them. In actual applications, the settings can be selected according to actual needs.

[0049] This application does not restrict whether the security policy matrix, state machine topology, and pre-constraint set are all configured in the same declarative configuration file. In practical applications, they can be selectively set according to the frequency of modification. For example, if the modification frequency is low, they can be set in the same declarative configuration file. In other embodiments, the security policy matrix, state machine topology, and pre-constraint set can each be set in a separate declarative configuration file.

[0050] For example, taking the declarative configuration file corresponding to the security policy matrix as an example, the declarative configuration file defines a policy type field, a data operation scope, and data operation rules. In some embodiments, the value of the policy type field includes access type, data authentication type, and data audit type. The data operation scope includes the adaptation task state. For example, the target data operation corresponding to the access type can only be accessed in three states: SURVEYING, SANDBOX_EXECUTING, and IMPLEMENTING. The data operation rules include the application environment, operation type (e.g., allowed operation types include full read, prohibited operation types include copy, discard, etc.), and data access conditions. The data access conditions may include requirements such as running in a sandbox, a maximum query quantity of 1000, and the need to perform de-identification operations, etc. These will not be elaborated on in this embodiment.

[0051] Understandably, the method also includes at least one of the following: The state transition events generated during the execution of the target data processing task are processed according to a preset data structure and recorded in the log file; The append-only method is used to record the state transition events generated during the execution of the target data processing task to the preset audit log, which is configured to be write-only. Upon receiving user query parameters, the unified query interface is invoked to query the state transition event based on the user query parameters. The user query parameters include at least one of the following: task identifier, time period parameter, trigger source parameter, and data source parameter.

[0052] State transition events can include normal state transition events and abnormal state transition events. Normal state transition events are used to indicate events where the state transition is successful, while abnormal state transition events are used to indicate the reasons for the failure of the state transition. Abnormal state transition events include events caused by the failure of target security policy verification, events caused by the failure of target general business precondition verification, events caused by the failure of state machine topology to perform state transition compliance verification, and events caused by the cumulative number of times exceeding the trigger threshold.

[0053] This application does not limit the format of the data structure. In some embodiments, it includes a task identifier field, an agent identifier field, a structured result field for precondition checks (such as the check result of the target defense-in-depth check condition), a global sequence number field for the target data task, an accessed data source field (used to record the accessed data source), an operation type field, a data range verification code field, a sandbox instance field, a remote session field (used to record which session accessed the data source), a structured result field for the compliance check of the target security policy, and a task context verification field. In some embodiments, other fields may also be included, which will not be elaborated on in this application. This data structure enables rapid troubleshooting of various anomalies during data processing.

[0054] By appending all abnormal state transition events to the same audit log without supporting modification or deletion, it is possible to ensure that all times are arranged according to a unified timeline, achieving end-to-end temporal consistency. Furthermore, the complete state transition sequence can be retrieved sequentially based on the task identifier, enabling cross-stage correlation queries. Moreover, when using data structure storage, the task context can be verified to ensure that the task context has not been tampered with, thus verifying data integrity.

[0055] This application embodiment does not limit the executing entity for processing the aforementioned state transition events. The central scheduling agent can acquire the corresponding event and then schedule it, or the central scheduling agent can record the event itself during the process of calling the state machine control layer to perform precondition checks on the target state. This application embodiment will not elaborate on these points further.

[0056] A complete state transition chain can be queried using a single task identifier.

[0057] By using a single time period parameter for querying, you can query the distribution of all task statuses within a specific time period.

[0058] By using a single trigger source parameter for querying, you can retrieve all operation records of a specific triggerer.

[0059] By using a single data source parameter, you can query all access records of a specific data source.

[0060] This application does not restrict whether the user query parameters input using the unified query interface are a single parameter or a combination of multiple parameters (such as a combination of task identifier and time period parameters, or a combination of task identifier and trigger source parameters, data source parameters, etc.). Those skilled in the art can selectively set these parameters according to the actual situation. This application will not elaborate on these points in detail.

[0061] Therefore, compared with the problems in the prior art where different types of verification storage have different log formats and scattered storage, requiring manual cross-file piecing for tracing, the embodiments of this application can realize automatic post-event tracing and analysis of the data processing process through a unified data structure, the same review log, and a unified query interface.

[0062] The same audit log refers to logs of the same type located at the same path. For example, Audit1.dog and Audit2.dog belong to different time periods of the same audit log, thus allowing for the splitting of audit files when the amount of data recorded in the audit log is large. In some embodiments, a single physical file can also be used. When the accumulated abnormal state transition events in this physical file exceed a preset number, the records of tasks that do not need to be traced within the historical time period are stored in other files. This application will not elaborate on these details in its embodiments.

[0063] Understandably, the method also includes: In response to receiving a data processing request from a user, the request is matched for sensitive words using a preset pre-defined rule engine filter. If no matching sensitive words are found in the data processing request, the central scheduling agent is notified to create the target data processing task corresponding to the data processing request and to initialize the current state of the target data processing task.

[0064] By performing hard code checks through a pre-defined rule engine filter, the impact of prompt word injection attacks can be reduced, as well as the risk of LLM "forgetting" or "misjudging".

[0065] For example, suppose the system receives user-inputted command text via an interactive interface. The pre-processing rule engine then filters the command text against a pre-defined sensitive word library (such as "investment advice," "risk prediction," and "credit rating"). If a match is found, the command is rejected and a pre-defined response is returned; the command is not passed to the central scheduling agent for processing. If no match is found, the command is passed to the central scheduling agent for processing. This ensures that even if the LLM is induced by prompt words, malicious commands cannot reach the agent, further enhancing the security of the data processing process.

[0066] Understandably, based on the state machine topology and the current state, a state transition compliance check is performed on the target state, including at least one of the following: If the current status is "task creation completed", and the target status is "under investigation" or "cancelled", then the compliance verification of the status transition is confirmed to be passed. When the current status is "under investigation", and the target status is "investigation result confirmed" or "failed", the compliance verification of the confirmed status flow is passed. If the current status is the survey results confirmation status, and the target status is the data architecture design status or the cancelled status, the status transition compliance verification is passed. When the current state is the data architecture design state, and the target state is the architecture design confirmation state or the failure state, the compliance verification of the state transition is confirmed to be passed. If the current status is architecture design confirmation, and the target status is data integration, data architecture design, or cancelled, the compliance verification of the status transition is confirmed to be passed. If the current state is data integration, and the target state is quality verification or failure, then the compliance verification of the status flow is confirmed to be successful. If the current status is quality verification status, and the target status is testing status, data integration status, or failure status, then the status transition compliance verification is confirmed to be successful. If the current state is test state, and the target state is deployment state, data integration state, or failure state, the compliance verification of the state transition is confirmed to be passed. If the current state is deployment, and the target state is running or failed, the compliance verification of the state transition is confirmed to be successful. If the current state is running, and the target state is either task completed or failed, the compliance verification of the state transition is confirmed to be successful. If the current status is a failure, and the target status is either in the investigation stage or the data integration stage, then the compliance verification of the status transition is confirmed to be successful.

[0067] This application does not limit the combination of the above-mentioned state flow compliance verification methods. In some embodiments, the state flow compliance verification of the target state can be some of the above-mentioned verifications or all of them. This application will not elaborate on these points one by one, and can selectively set them according to the actual situation.

[0068] The following reference Figures 2 to 5 The multi-agent collaborative data processing method described in this application is illustrated by an example of performing an internal enterprise data integration task within a multi-agent collaborative data processing system, as follows: Example 1: Financial institution 1 needs to integrate data from its investment management system A (production environment) and investment management system B (production environment) to generate a unified "golden copy" dataset. According to the company's data security policy, production environment data must not leave the intranet. Therefore, each agent can only view anonymized data samples remotely in a sandbox environment; the original data is never exposed. The specific process is as follows: Step 1: Task creation and remote connection establishment: Users send data processing requests via interactive applications such as WeChat: "Please help me investigate the data status of the two investment management systems on the internal network." The data processing request first passes through the pre-rule engine filter for sensitive word matching. When no sensitive words in the preset sensitive word library are matched, it is passed to the central scheduling agent. The central scheduling agent creates the target data processing task and initializes the current state of the target data processing task to CREATED, and determines the target state to CONNECTING. The central scheduling agent invokes the state machine control layer to verify CONNECTING. At this point, the state machine control layer's process is as follows: Figure 2 As shown: S1. Get the current state; S2. Check the flow rules: Specifically, it checks whether CONNECTING is in the allowed set based on the state machine topology, using the state machine topology... Figure 3 Taking the form shown as an example, then refer to... Figure 3 As shown, CONNECTING is the state in the flow direction of CREATED. Therefore, if the state flow compliance check result is passed, jump to S3; otherwise, jump to S4. S3. Perform precondition check: Taking CONNECTING as an example where no general service precondition is set, the general service precondition check result passes and jumps to S5; otherwise, the blocking reason is returned to the central scheduling agent. S4: Rejection. At this point, the central scheduling agent can either return an abnormal state event transition to the application layer and log it, or the state machine control layer can record it. S5, Execution State Transition: Among them, with Figure 3Taking the state machine topology shown as an example, after determining that the current state of the target data processing task is CONNECTING, a remote connection is triggered to establish a remote session. After the remote session connection is successful, transition(task_id, AUTHENTICATING, triggered_by="System") is automatically executed, calling the state machine control layer to perform an F471 check (authentication token valid). If successful, the current state is changed to AUTHENTICATING. After successful authentication, the state transitions to SURVEYING.

[0069] Step 2: Survey Execution ("Available but Invisible" Mode): The central scheduling agent determines the current state as SURVEYING through the state machine control layer, and issues survey sub-tasks. Different Surveyor agents select matching survey tasks from the issued survey sub-tasks.

[0070] In this context, each Surveyor agent, acting as the first specialized agent, invokes the state machine control layer to perform a security depth check before accessing any data. In some embodiments, such as... Figure 4 As shown, the security depth check includes items F470 to F475. In practical applications, some items from F470 to F475 can be combined to construct a security depth check. For example, using... Figure 4 Taking the checks in the example, in some embodiments, the Surveyor agent sequentially executes F470, F471, F473, F474, and F475. Passing all checks indicates that the security depth check has passed. Specifically, F470 verifies that the remote connection is still alive, F471 verifies that the security token is still valid, F473 verifies that the target data has been anonymized, F474 verifies that the current operation (such as READ_SCHEMA) is within the authorized scope, and F475 verifies that the audit session is active. At this time, such as... Figure 4 As shown, each intelligent agent with data access operations will undergo two types of precondition checks: general business precondition checks and security defense-in-depth checks.

[0071] After passing the security defense-in-depth check, the Surveyor agent creates a sandbox instance, performs data sampling and table structure probing within the sandbox, and generates a survey report. Throughout the entire process, the original sensitive data remains within the production environment.

[0072] Step 3: Wait for customer confirmation that the Surveyor agent has completed the task. The central scheduling agent determines the target state to be SURVEY_CONFIRMING and then calls transition(task_id, SURVEY_CONFIRMING, triggered_by="Surveyor"). The state machine control layer checks and passes the check, and the current state changes to SURVEY_CONFIRMING. The system then sends the anonymized survey report to the user via WeChat, awaiting confirmation.

[0073] Step 4: After client confirmation, proceed to the design phase: The user confirms the survey report in WeChat, and the target state is set to DESIGNING. The central scheduling agent calls the state machine control layer to verify the target state DESIGNING. At this point, the state transition compliance check result is passed, and the pre-check constraint verification is executed. The pre-check constraint verification includes F410 and data constraint checks F470-F475. If F410 and F470-F475 all pass, the current state is updated to DESIGNING. F410 is used to check whether the survey report has been confirmed by the user. Accordingly, the central scheduling agent triggers the Architect agent to design a solution.

[0074] The Architect agent follows the steps outlined in the Surveyor agent, undergoing state transition compliance checks and hard checks of preconditions (including operational constraint checks and data constraint checks) at each stage. Furthermore, any operation involving data access during the data constraint check must pass five security checks, F470-F475.

[0075] In step 4, if the user has not yet confirmed the survey report, and the central scheduling agent mistakenly calls the state machine control layer to verify the target state, the following procedure will be followed: The state machine control layer will inevitably fail the matching check function F410 and return the error message to the state machine control layer: "Precondition not met: The survey report has not been confirmed by the customer. Please confirm the survey report before continuing." At this time, the current state of the target data processing task is still SURVEY_CONFIRMING(203).

[0076] In the above steps, assuming a data source has not yet completed its data masking process, and the Surveyor agent attempts to access this data source, when the state machine control layer executes F473, it checks that the data masking status is "PENDING," returns a failure message, and displays the error message: "Data source XXX has not yet completed data masking. Please retry after data masking is complete. Current status: Pending data masking." The data access request is intercepted, and the original data is not exposed.

[0077] Example 2: Taking the automatic handling of remote connection interruptions as an example, suppose that during the IMPLEMENTING phase, the remote connection is unexpectedly interrupted due to network problems. The specific details are as follows: The connection heartbeat detection mechanism of the state machine control layer detects a connection interruption and automatically performs verification of the target state CONNECTION_LOST. Upon successful verification, the current state is changed to CONNECTION_LOST, and an audit log is recorded. When reconnection conditions are met (e.g., network recovery), a reconnection process is triggered. At this point, verification based on the state machine topology is performed, and upon successful verification, the state is updated to CONNECTING, and a reconnection attempt is made. If the reconnection is successful, the current state is updated to AUTHENTICATING, and authentication is performed. Upon successful authentication, the current state is updated to IMPLEMENTING, at which point the agent can resume execution from the point of interruption. In some embodiments, if the number of failed reconnections exceeds a threshold (e.g., the threshold is set to 3 times), the state machine control layer automatically sets the current state to FAILED. In some embodiments, when the current state is FAILED, maintenance personnel are notified for manual intervention.

[0078] Using the above method, the task status of each target data processing task is clear and traceable, which can avoid the task waiting indefinitely in a "suspended" state.

[0079] Example 3: Taking the support of cyclic detection protection for precondition checks as an example, the specific example is as follows: Suppose that during the quality verification phase, the implementing agent repeatedly enters the QA_TESTING state due to multiple execution anomalies. The state machine control layer records the number of transitions (i.e., the cumulative number) of this task entering the QA_TESTING state. Taking a trigger threshold of 5 as an example, when attempting to transition from IMPLEMENTING to QA_TESTING for the 6th time, if the number of transitions exceeds the trigger threshold, it is determined to be a loop anomaly. The state machine control layer then automatically executes the following: 1. Force the task state to FAILED (212); 2. Record the audit log (136), marking it as "loop detection triggered"; 3. Notify the user via WeChat: "The task has been detected to loop more than 5 times in the QA_TESTING phase and has been automatically terminated. Please contact the operations team for processing."

[0080] Example 4: Taking the security policy configured using a declarative configuration file as an example, the example is as follows: Suppose an enterprise security compliance team needs to add a new security policy: for production environment data sources, prohibit agents from performing full data reads (READ_FULL), and only allow reading anonymized samples (READ_SAMPLE). Then, as follows... Figure 5 As shown: Phase 1: Declarative Definition of Security Policy In the security policy configuration file, move the "Full Data Read" option from the "Allowed Data Operations" category to the "Disallowed Data Operations" category for the production environment data source, and save the configuration.

[0081] Phase Two: Loading and Compiling Execute the hot reload command. The state machine control layer will then reload the configuration file and update the security policy matrix in memory.

[0082] Phase 3: Runtime hard constraint execution: Taking the above modifications as an example of hard constraint verification in F474, such as... Figure 5 As shown, when an agent attempts to request the READ_FULL operation, F474 queries the security policy matrix, checks that the operation is not in the allowed list, immediately rejects the request, returns the error message "Operation is outside the authorized scope" and records it in the audit log.

[0083] Phase 4: Write to the audit log. Record configuration changes and inspection results in the audit log to form a closed-loop management of security policies, ensuring the auditability of security policies and further ensuring that any configuration changes and their runtime effects can be fully traced.

[0084] In summary, the methods of the above embodiments of this application have the following beneficial effects: (1) Improve the reliability of task scheduling. By imposing hard constraints on the flow rules and hard checks on the preconditions, the uncontrollable risks such as skipping, looping, and forgetting caused by the autonomous decision-making of LLM are reduced. All flows must comply with the predefined rules and preconditions to ensure that tasks proceed in the correct order.

[0085] (2) Enhance system security by using a pre-rule engine filter and a state machine control layer that run independently of the LLM, thus avoiding the impact of prompt injection attacks. Even if the LLM is maliciously manipulated, the state machine will still prevent transitions that do not conform to the rules, and the pre-filter will still intercept out-of-bounds instructions. Furthermore, a defense-in-depth system for "usable but invisible" scenarios is constructed through check functions such as F473-F475, further enhancing the system's security.

[0086] (3) Enhance safety constraints: By making the state machine control layer an independent control entity from the LLM, the occurrence of state transitions is determined by the code layer rather than by the LLM, thereby enhancing safety constraints.

[0087] (4) Improve the maintainability and auditability of the system: The unified and immutable audit log records the complete information of each state transition. When the task execution is abnormal, the audit log can be quickly searched to locate the problem link and cause, making the analysis of task execution more efficient.

[0088] (5) Balancing flexibility and standardization: While maintaining hard constraints on state transition rules, a fallback path (such as...) can be used to... Figure 2 (dashed arrow) and retry path (FAILED to SURVEYING / IMPLEMENTING), supporting flexible handling of abnormal situations and avoiding the inefficiency of fixed workflow engines that "have to start all over again once an error occurs".

[0089] (6) By employing state machine hard constraints, hard checks of preconditions (such as data anonymization checks in F473, operation scope checks in F474, etc.), and a combination of security policy configuration files and immutable audit logs, a complete compliance link from operation to audit can be provided for "usable but invisible" data integration scenarios. This allows enterprises to securely utilize large-scale model agents to break down internal data silos and achieve compliant release of data value without exposing the original data.

[0090] (7) Reduce the operational risks of remote data integration: By explicitly managing the remote connection status (CONNECTING / AUTHENTICATING / CONNECTION_LOST) and sandbox execution status (SANDBOX_EXECUTING), the uncertainty of the remote execution environment is brought into the controllable management scope of the state machine control layer. When the remote connection is interrupted, the state machine control layer automatically triggers the reconnection or failure process to avoid the task waiting indefinitely in a "suspended" state.

[0091] (8) Reduce system development and maintenance costs. By defining state transition rules, precondition checks, and security policies through declarative configuration files, no rewriting of scheduling logic is required when adding or modifying task flows; only configuration adjustments are needed. The coded expression of security policies makes security boundaries version-managed and audit-traceable.

[0092] It is understood that the multi-agent cooperative data processing system provided according to the embodiments of this application includes: The application interaction layer is used to obtain data processing requests from users. The multi-agent collaboration layer includes a central scheduling agent and multiple specialized agents. The central scheduling agent is used to create target data processing tasks based on data processing requests. The state machine control layer stores the state machine topology and the set of pre-constraint conditions. It responds to verification requests from the central scheduling agent for the target state, verifies the target state, and sends the verification results to the central scheduling agent. The verification results include state transition compliance verification results and general business pre-constraint verification results. The state transition compliance verification results are obtained by verifying the target state based on the state machine topology and the current state. The general business pre-constraint verification results are obtained by verifying the target state based on the target general business pre-constraint conditions that match the target state in the pre-constraint set. The current state of the target data processing task is updated to the target state when the verification results indicate that both the state transition compliance verification and the operation constraint verification have passed. Data access call layer; When the inspection results indicate that the verification has passed, the central scheduling agent issues a sub-task corresponding to the target state; each specialized agent selects the matching target sub-task from the issued sub-tasks and executes the target sub-task. The data access and invocation layer is used to provide each specialized intelligent agent with the data sources and tool resources required to execute the corresponding target sub-tasks.

[0093] This application embodiment does not restrict whether the application interaction layer, multi-agent collaboration layer, and state machine control layer are on the same physical device. They can be deployed separately on a single physical device (such as a terminal device, server, etc.) or partially deployed on the same device. This application embodiment will not elaborate on these points further.

[0094] For example, refer to Figure 6 As shown, the multi-agent cooperative data processing in this application embodiment includes the following components: User interaction layer: Used to receive user instructions and provide feedback on task execution status. This application does not limit the form of the user interaction layer. In some embodiments, the user interaction layer can be built based on instant messaging applications such as WeChat, DingTalk, or Lark. The user interaction layer is configured with a pre-defined rule engine filter for sensitive word matching. If a sensitive word in a preset database is matched, the request is rejected directly; otherwise, the instruction is passed to the multi-agent collaboration layer.

[0095] Multi-Agent Collaboration Layer: This layer comprises a central scheduling agent and multiple specialized agents with varying capabilities. The types and number of specialized agents can be selectively configured based on data processing needs. In some embodiments, the types of specialized agents include at least: Surveyor, Architect, Engineer, QA, Test, and Ops. Each agent is LLM-driven and responsible for executing specific types of tasks. In some embodiments, the multi-agent collaboration layer also includes a shared storage unit for storing the task execution results of each agent (such as survey reports, architecture schemes, implementation logs, etc.) for downstream agents to access.

[0096] State Machine Control Layer: A scheduling and control module that operates independently of the LLM. In some embodiments, the state machine control layer includes: (1) State definition and flow rule storage unit, used to store the predefined set of task states and allowed flow rules (i.e., state machine topology). (2) Precondition checking unit, configured with multiple checking functions (132A, 132B, 132C...). Each checking function corresponds to a precondition in the precondition set and is used to perform code-level hard checks before conversion. All checking functions do not involve any LLM calls. (3) A cyclic detection unit is used to record the number of transitions for each state and trigger protection; (4) The execution unit is responsible for updating the execution status; (5) Audit log unit, used to write transformation events to an immutable audit log.

[0097] The data access and invocation layer includes the MCP gateway and sandbox manager, as well as various data source adapters (such as database adapters, document adapters, API adapters, RPA adapters, etc.). The sandbox manager creates an isolated execution environment for each data integration task, ensuring that data integration operations are performed within the sandbox and that raw data is not leaked outside the sandbox. The MCP gateway is used to establish remote sessions.

[0098] Understandably, an electronic device provided according to embodiments of this application includes: At least one processor; At least one memory is provided for storing the programs of the central scheduling agent and the state machine control layer, respectively. The above-described multi-agent collaborative data processing method is implemented when the program of the central scheduling agent is executed by at least one processor.

[0099] For example, please refer to Figure 7 , Figure 7 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 701 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 702 can be a NAND flash. The relevant code of the central scheduling agent and the state machine control layer is stored in the memory 702 and is called by the processor 701 to execute the multi-agent cooperative data processing method of the present application embodiment. The input / output interface 703 is used to implement information input and output; The communication interface 704 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 705 transmits information between various components of the device (e.g., processor 701, memory 702, input / output interface 703, and communication interface 704); The processor 701, memory 702, input / output interface 703, and communication interface 704 are connected to each other within the device via bus 705.

[0100] This application also provides a computer-readable storage medium that stores a computer program. When the computer program is executed by a processor, it implements the control method of the above-described multi-agent cooperative data processing method.

[0101] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0102] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0103] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0104] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0105] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0106] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0107] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0108] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0109] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0110] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0111] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0112] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A multi-agent collaborative data processing method, characterized in that, The method includes: In response to obtaining the target state for the target data processing task, the central scheduling agent calls the state machine control layer to perform a precondition check on the target state and obtain the check result. The state machine control layer is configured with a state machine topology and a set of preconditions. The state machine topology includes multiple task states, and the set of preconditions includes at least one general business precondition. At least some of the general business preconditions are set to correspond to the task states. If the inspection result indicates that the verification has passed, the central scheduling agent issues a sub-task corresponding to the target state; Each specialized intelligent agent selects a matching target subtask from the published subtasks and executes the target subtask. The inspection results include state transition compliance inspection results and general business pre-check results. The state transition compliance inspection results are obtained by verifying the target state based on the state machine topology and the current state. The general business pre-check results are obtained by verifying the target state based on the target general business pre-conditions in the pre-constraint set that match the target state. The current state of the target data processing task is updated to the target state when the inspection results indicate that the state transition compliance inspection and the operation constraint inspection have passed.

2. The multi-agent cooperative data processing method according to claim 1, characterized in that, The inspection results also include the flow threshold inspection results, which are determined through the following steps: If the operation constraint verification passes, obtain the cumulative number of times the target data processing task enters the same target state during execution; If the cumulative number of times is less than or equal to a preset trigger threshold, the flow threshold check result is determined to be a successful check.

3. The multi-agent collaborative data processing method according to claim 1, characterized in that, The method further includes: Upon receiving a confirmation request from the first specialized intelligent agent for target data operation of the first target sub-task, the central scheduling intelligent agent obtains the target defense-in-depth inspection conditions corresponding to the target data operation by matching from the set of pre-constraint conditions through the state machine control layer; the first target sub-task is a sub-task actively retrieved by the first specialized intelligent agent from the central scheduling intelligent agent. Based on the target-in-depth defense inspection conditions, access verification is performed on the target data operations.

4. The multi-agent cooperative data processing method according to claim 3, characterized in that, The step of performing access verification on the target data operation based on the target defense-in-depth inspection conditions includes: Obtain the target security policy corresponding to the target subtask from the security policy matrix, which includes various data access boundary policies; each data access boundary policy corresponds to a different policy type. Access verification is performed on the target data operations using the target security policy and the target defense-in-depth inspection conditions.

5. The multi-agent cooperative data processing method according to claim 4, characterized in that, The method further includes: In response to the fulfillment of preset configuration loading conditions, a preset declarative configuration file is read from a preset directory; The declarative configuration file is processed to obtain configuration constraint data, which includes at least one of the following: a security policy matrix, the state machine topology, and a set of pre-constraint conditions.

6. The multi-agent cooperative data processing method according to claim 1, characterized in that, The method further includes at least one of the following: The state transition events generated during the execution of the target data processing task are processed according to a preset data structure and recorded in a log file; The state transition events generated during the execution of the target data processing task are recorded in a preset audit log using an append-only write method. Upon receiving user query parameters, the unified query interface is invoked to query the state transition event based on the user query parameters. The user query parameters include at least one of the following: task identifier, time period parameter, trigger source parameter, and data source parameter.

7. The multi-agent cooperative data processing method according to claim 1, characterized in that, The method further includes: In response to receiving a data processing request from a user, the data processing request is matched for sensitive words using a preset pre-defined rule engine filter; If it is determined that no matching sensitive words exist in the data processing request, the central scheduling agent is notified to create the target data processing task corresponding to the data processing request and to initialize the current state of the target data processing task.

8. The multi-agent cooperative data processing method according to claim 1, characterized in that, The step of performing state transition compliance verification on the target state based on the state machine topology and the current state includes at least one of the following: If the current status is "task creation completed", and the target status is "under investigation" or "cancelled", then the compliance verification of the status transition is confirmed to be passed. When the current status is "under investigation", and the target status is "investigation result confirmed" or "failed", the compliance verification of the status transition is determined to be successful. If the current state is the survey result determination state, and the target state is the data architecture design state or the cancelled state, then the compliance verification of the state transition is determined to be passed. If the current state is the data architecture design state, and the target state is the architecture design confirmation state or the failure state, then the compliance verification of the state transition is determined to be successful. If the current state is the architecture design confirmation state, and the target state is the data integration state, data architecture design state, or cancelled state, then the compliance verification of the state transition is determined to be passed. If the current state is data integration state, and the target state is quality verification state or failure state, it is determined that the state transition compliance verification has passed. If the current state is a quality verification state, and the target state is a test state, data integration state, or failure state, then the compliance verification of the state transition is determined to be passed. If the current state is a test state, and the target state is a deployment state, data integration state, or failure state, then the compliance verification of the state transition is determined to be passed. If the current state is the deployment state, and the target state is the running state or the failure state, then the compliance verification of the state transition is determined to be passed. If the current state is running, and the target state is task completion or failure, then the compliance verification of the state transition is determined to be passed. If the current state is a failure state, and the target state is a survey state or a data integration state, then the compliance verification of the state transition is determined to be successful.

9. A multi-agent collaborative data processing system, characterized in that, include: An application interaction layer is used to obtain data processing requests from users. A multi-agent collaboration layer, comprising a central scheduling agent and multiple specialized agents, wherein the central scheduling agent is used to create a target data processing task based on the data processing request; A state machine control layer is used to store the state machine topology and a set of pre-constraint conditions. The state machine control layer responds to the verification request from the central scheduling agent for a target state, performs pre-constraint checks on the target state, and sends the check results to the central scheduling agent. The check results include state transition compliance check results and general business pre-constraint check results. The state transition compliance check results are obtained by verifying the target state based on the state machine topology and the current state. The general business pre-constraint check results are obtained by verifying the target state based on target general business pre-constraint conditions that match the target state in the pre-constraint set. The current state of the target data processing task is updated to the target state if the check results indicate that both the state transition compliance check and the operation constraint check have passed. Data access call layer; Wherein, if the inspection result indicates that the verification is passed, the central scheduling agent issues a sub-task corresponding to the target state; each of the specialized agents selects a matching target sub-task from the issued sub-tasks and executes the target sub-task; The data access and invocation layer is used to provide each of the specialized intelligent agents with the data sources and tool resources required to execute the corresponding target sub-tasks.

10. A computer-readable storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are used by a processor to execute the multi-agent cooperative data processing method as described in any one of claims 1 to 8.

Images