A multi-model intelligent agent collaboration method and system

Abstract

The present invention relates to the field of artificial intelligence technology, specifically a multi-model intelligent agent collaboration method and system, comprising the following steps: obtaining task execution plans and actual status, identifying deviation nodes, generating a synchronization offset list, analyzing task trends, marking progress labels, identifying tool-affected sections, screening unaffected nodes, judging distribution efficiency, extracting abnormal fluctuations, analyzing resource and task cycle differences, and generating monitoring structure indicators. In the present invention, by extracting task plan numbers and time intervals and comparing the real-time status of intelligent agents, task deviation identification is achieved, the accuracy of plan execution monitoring is improved, and the perception of progress changes is enhanced. By cross-comparing fluctuating tasks with tool call data, interference sections are identified, operation logs and dispatch data are mapped, task distribution efficiency is evaluated, and progress monitoring indicators are extracted based on resource allocation and task fluctuation differences, thereby improving collaborative efficiency and system control capabilities.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multi-model intelligent agent collaboration method and system. Background Art

[0002] The field of artificial intelligence technology includes multiple branches such as intelligent agent technology, machine learning, deep learning, natural language processing, computer vision, etc. The core content of this technology field is to design autonomous learning, judgment and decision-making by simulating human intelligence. Intelligent agent refers to an automated system that can perceive the environment, make decisions, execute tasks and feedback results. With the improvement of computing power and the enrichment of data resources, artificial intelligence technology has been widely used, covering multiple application scenarios such as automated control, image recognition, speech processing, natural language understanding, etc. The development of this field has not only promoted the progress of science and technology, but also brought profound changes to various industries, including education, transportation and many other fields.

[0003] Among them, the multi-model intelligent agent collaboration method refers to achieving more efficient task completion through the collaboration of multiple intelligent agents. The theme proposes an optimized collaboration method to address the problems of task allocation, resource coordination, information sharing, etc. in the collaborative work of multiple intelligent agents. This method designs a collaboration mechanism between multi-model intelligent agents, enabling each intelligent agent to flexibly collaborate and cooperate according to the requirements of different tasks. In this method, intelligent agents share knowledge, data, and resources to achieve information transmission and joint completion of tasks. Each intelligent agent performs tasks independently or collaboratively based on its own characteristics and capabilities, thereby improving the efficiency and intelligence level of the overall system. This method also optimizes the process and efficiency of task execution by adjusting the collaboration strategy between intelligent agents.

[0004] Existing technologies for task status identification mostly rely on static execution processes, lacking the ability to promptly capture actual operational deviations. This makes it difficult to detect progress deviations during task execution. For example, when multiple tasks are executed in parallel, a delay at a task node can cause the entire chain of subsequent tasks to lag behind. However, the lack of a real-time monitoring mechanism prevents early warning of delay risks. Task trend assessments are often based on fixed templates rather than the evolution of actual operational data, making it difficult to accurately reflect task progress and hindering resource allocation and intervention strategies. The lack of tool call impact analysis prevents the assessment of conflicts between frequent calls of key tools and task node execution, leading to resource congestion in high-concurrency scenarios. The lack of a mechanism to compare operational logs with task dispatch data makes it difficult to detect inefficient task distributions, resulting in wasted resources. Existing technologies for progress monitoring primarily rely on coarse-grained statistics, neglecting fine-grained segmented resource matching and task completion efficiency analysis. This makes them unable to adapt to dynamic control requirements. This problem is particularly prominent in multi-model agent collaboration, directly impacting overall coordination and task completion quality. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a multi-model intelligent agent collaboration method and system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-model agent collaboration method, comprising the following steps:

[0007] S1: Obtain the knowledge base documents, tool interface configuration information, and external API call records in the task scenario, extract the execution plan number and time interval of the agent task node, compare the agent's real-time execution status with the plan node identifier, identify the task node number that deviates from the plan, and generate a task node synchronization offset list;

[0008] S2: Based on the task node synchronization offset list, select an agent-type intelligent assistant, extract the task decomposition start and end time in combination with the task decomposition table, calculate the difference with the planned time point and classify the trend, and generate a task progress trend label group;

[0009] S3: Call the task progress trend tag group, extract the fluctuation task partition number, identify the regional tool call frequency data, compare the tool call period with the fluctuation task time period, record the number of overlapping time periods, and generate a list of task node segments affected by the tool call;

[0010] S4: Based on the list of sections of the task node affected by the tool call, filter the unaffected partition node tasks, extract the agent segment task records and operation log data, map the daily task assignment volume and the agent operation period, determine whether there is an inefficient distribution of tasks, and obtain the agent node task completion fluctuation group.

[0011] As a further solution of the present invention, the task node synchronization offset list includes the task offset number, execution status label, time offset, and task node classification; the task progress trend label group includes the progress offset level, trend change type, plan comparison result, and task association number; the task node tool call affected segment list includes the tool call type, affected time segment, number of overlapping time periods, and affected task number; the intelligent node task completion fluctuation group includes the task uneven distribution number, running time record, task completion deviation, and intelligent agent operation matching degree.

[0012] As a further solution of the present invention, the steps for obtaining the task node synchronization offset list are specifically as follows:

[0013] S111: Obtain the knowledge base documents, tool interface configuration information, and external API call records in the task scenario, extract the agent task node plan time and number, match the knowledge base document update time and coordinates, compare the matching document time range with the node plan time, and generate the partition node document record period;

[0014] S112: Based on the partition node document record period, the overlap with the task node planned time interval is judged, the ratio of the overlap time to the total duration is extracted, the node numbers with a ratio lower than the benchmark value are screened, and the node document status label is updated to obtain the partition node progress coverage deviation rate;

[0015] S113: According to the partition node progress coverage deviation rate, the deviation status of the task node number is determined, the task node number with a deviation rate higher than the node synchronization threshold is identified, the node number, document coverage information and deviation rate value are integrated, and a task node synchronization offset list is generated.

[0016] As a further solution of the present invention, the steps for obtaining the task progress trend tag group are specifically as follows:

[0017] S211: Based on the task node synchronization offset list, identify an agent-type intelligent assistant with reasoning-planning-execution capabilities and a task decomposition table, extract the real-time task decomposition start and completion time, calculate the start and end time difference of the task segment, and compare it with the planned task node time difference to obtain the task time offset value;

[0018] S212: Call the task time offset value, combine the segment distribution, offset trend and adjustment frequency, collect task node offset data, identify and calculate the trend offset degree according to the segment number, determine the trend direction according to the adjustment frequency, and obtain the task progress trend label group.

[0019] As a further solution of the present invention, the steps for obtaining the list of sections of the task node affected by the tool call are specifically as follows:

[0020] S311: calling the task progress trend tag group, filtering the fluctuation offset task partition number, extracting the task time period according to the node association table, processing the segment task time period according to the time dimension, identifying the task time period index table, and obtaining the fluctuation task activity time period set;

[0021] S312: Based on the set of fluctuating task activity periods, collect tool call frequencies in the same period, identify the impact table, determine interference based on the interference threshold, match it with the fluctuating task construction period, analyze tool intervention intensity, determine abnormal correlations in progress monitoring, and generate a list of sections of task nodes affected by tool calls.

[0022] As a further solution of the present invention, the steps for obtaining the agent node task completion fluctuation group are specifically as follows:

[0023] S411: Based on the list of sections of the task node affected by the tool call, filter the unmarked partition nodes, extract and record the job task list of the node, including the task number, planned process, start time and end time, and obtain the task set of the node not affected by the tool call;

[0024] S412: Calling the node task set that is not interfered with by the tool call, matching the agent's daily segmented task records with the operation log data, extracting the planned task amount by task number, counting the running time and the real-time running time period, and generating an agent task execution matching data set;

[0025] S413: Based on the intelligent agent task execution matching data set, determine the matching degree between the task execution efficiency and the operation distribution, identify the efficiency fluctuation nodes and evaluate the task deviation degree, analyze the task efficiency fluctuation degree of the task nodes, mark the tasks whose fluctuation degree exceeds the set benchmark value as abnormal task nodes, and obtain the intelligent agent node task completion fluctuation group.

[0026] As a further embodiment of the present invention, the method further comprises step S5:

[0027] S5: Call the task completion fluctuation group of the agent node, extract the abnormal fluctuation task group, calculate the ratio of resource input to task amount in the knowledge base, record the difference distribution between resource input cycle and task execution cycle, and generate task progress monitoring structure indicators;

[0028] The task progress monitoring structure indicators include resource allocation ratio, task intensity level, execution cycle deviation, and task efficiency index.

[0029] As a further solution of the present invention, the steps for obtaining the task progress monitoring structure indicator are specifically as follows:

[0030] S511: Call the node number and the corresponding operation time interval in the task completion fluctuation group of the agent node, identify the task volume difference between adjacent tasks, filter out the nodes that exceed the threshold, record the time interval and task volume change range, and obtain the progress fluctuation abnormality identification set;

[0031] S512: Based on the knowledge base resources corresponding to the nodes in the progress fluctuation abnormality identification set, identify the node resource input ratio sequence, extract the abnormal distribution interval and compare the critical coefficient, record the ratio deviation direction and node number, and form a task resource matching deviation indicator group;

[0032] S513: According to the task resource matching offset indicator group, extract the knowledge base resource delivery and task execution time period, analyze the difference between the resource delivery cycle and the operation cycle, identify the task time synchronization offset distribution table, and sort and mark the difference according to the progress benchmark to generate the task progress monitoring structure indicator.

[0033] The multi-model agent collaboration system is used to execute the multi-model agent collaboration method, and the system includes:

[0034] The task status extraction module obtains the knowledge base documents, tool interface configuration information and external API call records in the task scenario, extracts the agent task node number and time interval, compares the real-time execution status with the task node identifier, marks the node tasks with inconsistent time, and generates a node status offset label group;

[0035] The node trend classification module locates the agent-type intelligent assistant with reasoning-planning-execution capabilities based on the node state offset label group, extracts the task decomposition table and planned time, identifies the difference between the real-time task time and the planned time, classifies and labels the difference type, and generates a task node progress trend identification group;

[0036] The tool call identification module selects the fluctuating and lagging task nodes based on the task node progress trend identification group, locates the corresponding partitions, extracts the tool call frequency interference period, determines the overlap with the task time period, selects the frequent interference sections, and generates a task progress tool call interference mapping set;

[0037] The task efficiency diagnosis module calls the interference mapping set based on the task progress tool, eliminates the interfering segment tasks, extracts the agent task records and operation logs, matches the task assignments with the operation time periods, calculates the operation volume and task ratio, identifies the dense but inefficient node tasks, and obtains the node task execution deviation set;

[0038] The resource allocation analysis module locates the agent resource input record based on the node task execution deviation set, extracts the task volume and resource allocation ratio, compares the execution cycle and input cycle difference, maps resource usage and task progress status, and forms a task progress monitoring structure indicator.

[0039] Compared with the prior art, the advantages and positive effects of the present invention are:

[0040] In the present invention, by dynamically extracting the execution plan number and time interval of the task node and comparing the real-time status of the intelligent agent, task deviation identification is achieved, and the dynamic monitoring accuracy of the plan execution is improved. In the task trend analysis stage, an intelligent assistant with reasoning and planning capabilities is integrated, and based on the task decomposition time and plan node evaluation difference changes, trend labels are effectively generated to enhance the insight into the changes in task progress. Further, through the cross-analysis of the fluctuating task time period and the tool call data, the time overlap is recorded, so that the sections of the task affected by external factors can be accurately identified, and the undisturbed task partitions are mapped to the dispatch amount and the running time period through the intelligent agent operation log to achieve a quantitative evaluation of the intelligent agent resource and task distribution efficiency. Combined with the task completion fluctuation and the difference in resource delivery cycle, the task completion efficiency and resource allocation matching degree are refined to form a progress monitoring indicator system with timeliness and accuracy. The entire processing logic reflects a full-process closed loop from task execution status identification, trend evolution analysis, tool call impact analysis, resource distribution rationality assessment to indicator structure generation, ensuring that task execution and resource allocation are simultaneously optimized in multiple dimensions, thereby effectively improving overall operational efficiency and responsiveness, and enhancing the flexibility and adaptability of task control in complex multi-agent collaboration scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the workflow of the present invention;

[0042] Figure 2 This is a flowchart for obtaining the task node synchronization offset list in the present invention;

[0043] Figure 3 This is a flowchart for obtaining the task progress trend tag group in the present invention;

[0044] Figure 4 A flowchart for obtaining a list of sections affected by tool calls on a task node in the present invention;

[0045] Figure 5 This is a flow chart for obtaining the fluctuation group of the intelligent node task completion in the present invention;

[0046] Figure 6 This is a flowchart for obtaining task progress monitoring structure indicators in the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, 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 only used to explain the present invention and are not intended to limit the present invention.

[0048] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0049] Example 1: Please refer to Figure 1 The present invention provides a technical solution: a multi-model intelligent agent collaboration method, comprising the following steps:

[0050] S1: Obtain the knowledge base documents, tool interface configuration information, and external API call records in the task scenario, extract the execution plan number and time interval of the agent task node, compare the agent's real-time execution status with the plan node identifier, identify the task node number that deviates from the plan, and generate a task node synchronization offset list;

[0051] S2: Based on the task node synchronization offset list, select agent-type intelligent assistants with reasoning, planning, and execution capabilities. Combined with the task decomposition table, extract the task decomposition start and end times, calculate the difference with the planned time points, and classify the trends to generate task progress trend label groups.

[0052] S3: Call the task progress trend tag group, extract the fluctuation task partition number, identify the regional tool call frequency data, compare the tool call period with the fluctuation task time period, record the number of overlapping time periods, and generate a list of task node segments affected by tool calls;

[0053] S4: Based on the list of segments affected by tool calls on task nodes, filter out unaffected partition node tasks, extract agent segment task records and operation log data, map the daily task allocation volume with the agent operation period, determine whether there is an inefficient distribution of tasks, and obtain the agent node task completion fluctuation group;

[0054] S5: Call the agent node task completion fluctuation group, extract the abnormal fluctuation task group, calculate the ratio of resource input to task volume in the knowledge base, record the difference distribution between resource input cycle and task execution cycle, and generate task progress monitoring structure indicators.

[0055] The task node synchronization offset list includes the task offset number, execution status label, time offset, and task node classification. The task progress trend label group includes the progress offset level, trend change type, plan comparison result, and task association number. The task node affected by the tool call segment list includes the tool call type, affected time segment, number of overlapping time periods, and affected task number. The agent node task completion fluctuation group includes the task uneven distribution number, running time record, task completion deviation, and agent operation matching degree. The task progress monitoring structure indicators include resource allocation ratio, task intensity level, execution cycle deviation, and task efficiency indicator.

[0056] See also Figure 2 , the specific steps for obtaining the task node synchronization offset list are:

[0057] S111: Obtain the knowledge base documents, tool interface configuration information, and external API call records in the task scenario, extract the agent task node plan time and number, match the knowledge base document update time and coordinates, compare the matching document time range with the node plan time, and generate the partition node document record period;

[0058] Obtain the knowledge base document data set in the task scenario from the agent work platform background, such as project requirement documents, interface instruction manuals, original task reports, etc.; read the tool interface configuration information list from the Agent system configuration, including the description, parameter definition, calling method, expected response format, etc. of each tool API; query the external API call record details within a specific time range from the log storage, such as capturing the request, response log, call timestamp, return status code, etc. of the Agent interacting with the third-party service when executing the task; extract the planned start and end time of the agent task node and the corresponding task node number from the task decomposition plan table, for example, the node Node_Alpha planned execution time is from 10-27-10:00 to 10-27-10:15, the node number is N001, and the node Node The Beta plan execution time is from 10-27-10:20 to 10-27-10:35, the node number is N002, and the last update timestamp of the knowledge base document is matched. For example, the update time of document DocA (project requirement document) is 10-27-10:05, the update time of document DocB (interface instruction manual) is 10-27-10:22, and the update time of document Doc_C (another version of the interface instruction manual) is 10-27-10:30. The associated coordinates of the knowledge base documents in the task scenario are matched. The coordinates can be document version numbers, specific section identifiers, or associated labels with task content. For example, the associated coordinates of Doc_A are "Requirementv1.1, Section3", the associated coordinates of Doc_B are "APIManualv2.0, EndpointA", and the associated coordinates of Doc_C are "APIManualv2.1, Endpoint". Compare the time ranges of the matched documents, and the effective impact of the documents is taken here. Time is defined as its update time point. If the update time point is earlier than the node's planned start time, it will not be used as a recording period. If the update time point is within the planned interval, the update time point will be recorded. For nodes with multiple matching documents or multiple update time points, all update time points within the planned interval are recorded. For example, if the node N001 has a planned execution time of [10:00, 10:15] and the associated document Doc_A is updated at 10:05, the recording period is 10:05. If the node N002 has a planned execution time of [10:20, 10:35] and the associated documents Doc_B and Doc_C are updated at 10:22 and 10:30, the recording periods are 10:22 and 10:30. A partitioned node document recording period is generated, which contains each task node number and all corresponding related knowledge base document update time points that fall within its planned execution time range. For example, the set is {N001:[10:05], N002:[10:22, 10:30]}.

[0059] S112: Based on the partition node document record period, the overlap with the task node planned time interval is determined, the ratio of the overlap time to the total duration is extracted, the node numbers with a ratio lower than the benchmark value are filtered, and the node document status label is updated to obtain the partition node progress coverage deviation rate;

[0060] Based on the partition node document record time period set, for example, the set is {N001: [10:05] N002: [10:22, 10:30]}, and the coincidence with the planned time interval of each task node is judged. For node N001, its planned time interval is [10:00, 10:15], and the record period is [10:05]. It is judged whether the record period 10:05 falls within the planned interval [10:00, 10:15], indicating that the document update and the node execution time overlap. The ratio of the overlap time to the total duration is extracted. Here, the impact of the document update time point is set as a For example, a fixed-length impact window is set to 5 minutes. This means that the task may be affected within 5 minutes after the document is updated. For the recording period 10:05, the impact window is [10:05, 10:10]. The total planned duration of node N001 is 15 minutes. Calculate the overlap between the impact window [10:00, 10:10] and the planned interval [10:00, 10:15]. The overlap interval is [10:05, 10:10], and the overlap duration is 10:10-10:05=5 minutes. Calculate the ratio of overlap time to total duration. Ratio = overlap time / planned total duration = 5 / 15 ≈ 0.333, for node N002, the planned interval is [10:20, 10:35], the total duration is 15 minutes, and the recording period is [10:22, 10:30]. Consider two recording periods, each affecting a window of 5 minutes: [10:22, 10:27] and [10:30, 10:35]. The overlapping time with the planned interval |10:20, 10:35] is 5 minutes and 5 minutes respectively, and the total overlapping time is 5+5=10 minutes (assuming that the two impact windows do not overlap or the overlapping part is only calculated once, here we simplify the calculation of each overlapping time and then sum it up). Calculate the ratio of overlapping time to total duration, ratio = 10 / 15≈0.667, and filter the node number whose ratio is lower than the benchmark value. The benchmark value is used to define the situation where the document update has little impact on the task. Its setting refers to the task's impact on the knowledge base. The degree of document dependency and document update frequency. For example, data processing tasks rely on data specification documents, and their updates directly impact processing logic, resulting in a high degree of dependency. Tasks that only reference interface parameters have a relatively low degree of dependency. Benchmark values can be determined through raw data analysis, such as statistically analyzing the correlation between document updates and task failure and delay rates over a period of time to determine an empirical value. Alternatively, they can be manually set based on task criticality. For example, a benchmark value of 0.5 indicates that task nodes where document updates affect less than half of the planned window duration need to be further marked. Comparing node N001's ratio of 0.333 with the benchmark value of 0.5 reveals that 0.333 is less than 0.5, so node N001 is screened out. Comparing node N002's ratio of 0.667 with the benchmark value of 0.5 reveals that 0.667 is greater than 0.5. Node N002 is not filtered. Update the node document status label, marking the filtered node N001 as "Document Coverage is Low" and the unfiltered node N002 as "Document Coverage Meets Standard." Obtain the partition node progress coverage deviation rate. For the filtered node, the deviation rate is defined as 1-ratio. That is, the coverage deviation rate of node N001 is 1-0.333≈0.667. For the unfiltered node, the deviation rate can be considered 0 or not high.

[0061] S113: Determine the deviation status of the task node number based on the partition node progress coverage deviation rate, identify the task node number whose deviation rate exceeds the node synchronization threshold, integrate the node number, document coverage information and deviation rate value, and generate a task node synchronization offset list;

[0062] According to the coverage deviation rate of the partition node progress, for example, the coverage deviation rate of node N001 is about 0.667, and the coverage deviation rate of node N002 is not high (for example, 0), the deviation status of the task node number is determined. The basis for the determination is whether the coverage deviation rate is higher than the set node synchronization threshold. The node synchronization threshold is used to define the minimum requirement for the synchronization of document updates and task execution. Its setting refers to the real-time requirements of the task and the sensitivity to the timeliness of the document. For example, for tasks with high real-time requirements and document updates must be reflected in the execution in a timely manner (such as financial trading strategy execution), the threshold can be set lower (such as 0.1), and any deviation will be marked. On the contrary, for tasks that can tolerate a certain delay (such as offline data analysis), a higher threshold (such as 0.5) can be set. The setting of the threshold It can also be determined by analyzing the correlation between synchronization deviation and task anomaly in the original data. For example, the node synchronization threshold is set to 0.4, indicating that the task node with a coverage deviation rate higher than 0.4 is determined to be synchronization deviation. The task node numbers with deviation rates higher than the node synchronization threshold are identified. The deviation rate of node N001, 0.667, is compared with the threshold of 0.4. Since 0.667>0.4, node N001 is identified as a synchronization deviation node. The deviation rate of node N002, 0, is compared with the threshold of 0.4. Since 0<0.4, node N002 is not identified as a synchronization deviation node. The identified synchronization deviation node numbers, related document coverage information, and calculated deviation rate values are integrated, and the information is organized into structured data, such as a list or table, to generate a task node synchronization deviation list.

[0063] Table 1: Task node synchronization offset list;

[0064] .

[0065] As shown in Table 1, the task node synchronization offset list lists the task nodes that are judged to have synchronization problems due to the low overlap between the knowledge base document update and the task plan time, including the node number, document coverage, coverage deviation rate and its judgment status.

[0066] See also Figure 3 ,The specific steps for obtaining the task progress trend tag group are:

[0067] S211: Based on the task node synchronization offset list, identify an agent-type intelligent assistant with reasoning-planning-execution capabilities and a task decomposition table, extract the real-time task decomposition start and completion time, calculate the start and end time difference of the task segment, and compare it with the planned task node time difference to obtain the task time offset value;

[0068] Based on the task node synchronization offset list, for example, the list contains nodes N001, N004, and N005 to identify agent-type intelligent assistants with reasoning-planning-execution capabilities. The agent has a specific role or capability label in the configuration, such as "Agent:RPE", and has the ability to adjust plans according to environmental changes and call tools to execute tasks. Identify the task decomposition table related to the agent execution and obtain the decomposition details of the currently executing or recently executed agent tasks (corresponding to task nodes) from the agent's task schedule or task executor log. The decomposition includes a series of finer-grained For example, node N001 is decomposed into steps A.1 (data loading) and A.2 (data preprocessing). The start and completion times of the real-time task decomposition are extracted. The actual start and end times of each sub-step or segment after the task decomposition are obtained from the agent's running log or task execution record. For example, when Agent_A executes the task corresponding to node N001, the real-time execution time of step A.1 is 10-27-10:02 and completed 10-27-10:08, and the real-time execution time of step A.2 is 10-27-10:09 and completed 10-27-10:17. The start and end time difference of the task segment is the real-time completion time minus the real-time start time. For example, the real-time time difference of step A.1 is 10:08-10:02=6 minutes, and the real-time time difference of step A.2 is 10:17-10:09=8 minutes. It is compared with the time difference of the planned task node. Here, it means comparing the duration of the real-time execution segment with the duration of the segment in the original plan. The original plan decomposes the total duration of the node into each segment. For example, the total duration of node N001 is 15 minutes, and the decomposition plan is 5 minutes for step A.1 (for example, from 10:00 to 10:05), and the planned duration of step A.2 is 15 minutes. For example, the time offset of step A.1 is 6 minutes (10:05 to 10:15) and the planned time is 5 minutes. The time offset of step A.2 is 8 minutes (10:15) and the planned time is 10 minutes. The time offset of step A.2 is 8 minutes (10:15) and the planned time is 10 minutes. The time offset of the task is obtained. The offset value is defined as the real-time duration minus the planned duration. For example, the time offset of step A.1 is 6-5=+1 minute. A positive value indicates a delay and a negative value indicates an advance. The time offset of step A.2 is 8-10=-2 minutes. Repeat this process to obtain the time offset value of each task segment executed by the agent. For node N001, the time offset value of its task segment is [+1, -2] minutes.

[0069] S212: Call the task time offset value, combine the segment distribution, offset trend and adjustment frequency, collect the task node offset data, identify and calculate the trend offset degree by segment number, determine the trend direction based on the adjustment frequency, and obtain the task progress trend label group;

[0070] Call the task time offset value, for example, the task segment time offset value of node N001 is [+1, -2] minutes. Combined with segment distribution, offset trend and adjustment frequency, segment distribution refers to the position of the task segment in the entire task process, the relationship between the predecessor and successor, the task node to which it belongs, and other related information. The offset trend refers to whether the offset value of the same type of segment (for example, all "data preprocessing" steps) or the segment executed by the same Agent within a period of time is continuously positive (continuous delay), continuously negative (continuous advance) or fluctuating (positive and negative alternation). The adjustment frequency refers to how many times the user retries, adjusts parameters, changes resource allocation or modifies the plan for the segment, related tasks or execution agents. The service node offset data is used to associate the offset values (such as +1, -2) of all task segments under the same task node (such as N001), the segment numbers (such as A.1, A.2), the executing agents (such as Agent_A), and the related adjustment frequencies. The trend offset is identified and calculated by segment number. For the same segment type or number, the average offset, cumulative offset, or maximum offset when it is executed multiple times (in different tasks or repeated executions of the same task) or on different agents is calculated as the trend offset. For example, if step A.1 is executed multiple times in different tasks and its offset values are +1, +0.5, +2, and +1.5 respectively, the average trend offset is =(1+0.5+2+1.5) / 4=+1.25 minutes. The trend direction is determined based on the adjustment frequency. If the deviation of a certain segment is high (for example, the average delay is large) and its adjustment frequency is also high, it indicates that there is a persistent problem that causes repeated adjustments but still delays. The trend direction is determined to be "persistent delay fluctuations". If the deviation is positive but the adjustment frequency is low, it is a one-time problem or an unreasonable plan. The trend direction is determined to be "occasional delays". If the deviation is close to zero and the adjustment frequency is low, it is determined to be "stable progress". Here, the "high" deviation threshold is set to an average deviation absolute value greater than 1.0 minute, and the "high" adjustment frequency threshold is set to 3 times / day. For example, the average value of step A.1 The average deviation is +1.25 minutes (the absolute value is higher than 1.0 minute), and the adjustment frequency is 4 times / day (higher than 3 times / day). Therefore, the trend direction is determined to be "continuous delay fluctuation". The average deviation in step A.2 is -2 minutes (the absolute value is higher than 1.0 minute), and the adjustment frequency is 1 time / day (less than 3 times / day). Therefore, the trend direction is determined to be "occasional advance". The determination result is associated with the task node to obtain the task progress trend label group, for example, {N001: contains continuous delay fluctuation segment, N002: stable progress}. Specific to the segment level, the label group information is more detailed, such as (N001, A.1): continuous delay fluctuation, (N001, A.2): occasional advance, ...}.

[0071] See also Figure 4 The steps for obtaining the list of segments affected by tool calls on task nodes are as follows:

[0072] S311: Call the task progress trend tag group, filter the fluctuation offset task partition number, extract the task time period according to the node association table, process the segment task time period according to the time dimension, identify the task time period index table, and obtain the fluctuation task activity time period set;

[0073] Call the task progress trend label group. For example, the label group shows that node N001 contains "continuous delay fluctuation segment". Filter the fluctuation offset task partition number and identify the task node number or segment number containing the word "fluctuation" or "offset" in the label. For example, node N001 and its subordinate fluctuation segments (such as A.1) extract the task time period according to the node association table, and find the planned or actual execution time period corresponding to the fluctuation node or segment from the task plan or execution record. For example, the planned time period of node N001 is [10:00, 10:15], the planned time period of fluctuation segment A.1 is [10:00, 10:05], and the actual execution time period is [10:02, 10:08]. Process the segment task time period according to the time dimension and extract it. The time periods are sorted, merged, or split to facilitate subsequent time window analysis. For example, the actual execution periods of all fluctuation segments are collected and sorted by time to obtain a time period list, identify the task time period index table, create an index, and associate the processed time periods with the corresponding fluctuation task nodes or segments. For example, the index item {[10:02, 10:08]:(N001, A1), [10:09, 10:17]:(N001, A.2)} (assuming that A.2 is accidentally advanced but is still included in the fluctuation analysis), and obtain the fluctuation task activity period set, that is, the set containing the actual execution time of all fluctuation task nodes or segments, for example, the set is {[10:02, 10:08], [10:09, 10:17]}.

[0074] S312: Based on the fluctuating task activity period set, collect the tool call frequency of the same period section, identify the impact table, determine the interference based on the interference threshold, match it with the fluctuating task construction period, analyze the tool intervention intensity, determine the progress monitoring abnormality correlation, and generate a list of task node sections affected by the tool call;

[0075] According to the set of fluctuating task activity periods, for example, the set is {[10:02, 10:08], [10:09, 10:17]}, the tool call frequency of the same period is collected. During the activity period, the number of times the agent calls the tool interface or the number of calls per minute is counted from the external API call records of the agent system. For example, in the time period [10:02, 10:08] (a total of 6 minutes), the agent calls a tool APIX15 times and APIY5 times. In the time period [10:09, 10:17] (a total of 8 minutes), it calls APIX8 times and APIY20 times. The APIX in [10:02, 10:08] is calculated as The average call frequency is 15 / 6=2.5 times / minute. APIY is based on the fluctuating task activity period set, for example, the set is {[10:02, 10:08], [10:09, 10:17]}, and the tool call frequency of the same period segment is collected. During the activity period, the number of times the Agent calls the tool interface or the number of calls per minute is counted from the external API call records of the Agent system. For example, in the time period of 10:02 and 10:08 (a total of 6 minutes), the Agent called a tool APIX 15 times and APIY 5 times. In the time period of 10:09 and 10:17 (a total of 8 minutes), it called APIX 8 times and APIY 20 times, totaling The average call frequency of APIX between 10:02 and 10:08 is 15 / 6 = 2.5 times / minute, and that of APIY is 5 / 6 ≈ 0.83 times / minute. The impact table is identified. An impact table is a predefined rule or data table that describes the type and intensity of the impact of a specific tool's call behavior (such as high-frequency calls, call failures, and abnormal parameters) on the task progress. For example, the impact table entry {tool APIX call frequency > a certain threshold, potential impact: progress delay, impact intensity coefficient: 1.2} is used to determine interference based on the interference value. The interference value is used to define the degree to which the tool call affects the task progress. Its setting refers to the characteristics of the tool, the agent's dependence on the tool, and the normal Consider the tool call pattern in this situation. For example, for tool APIX, the average call frequency is 1.0 times / minute. The interference threshold can be set to 1.5 times its normal average, that is, 1.0 x 1.5 = 1.5 times / minute. This means that if the average call frequency exceeds 1.5 times per minute, it is considered interference. The threshold setting can be determined by analyzing the correlation between tool call frequency and task anomalies or delays in the original run log. Here, the calculated APIX frequency is 2.5 times / minute, and the value is 1.5 times / minute. Because 2.5 > 1.5, APIX is considered to be interfering. The APIY frequency is 0.83 times / minute. Assuming its threshold is 1.0 times / minute, 0.83 < 1.0, APIY is judged to have no interference and is matched with the fluctuating task construction period. Check whether the tool call period judged to be interfering (i.e., the period [10:02, 10:08] where the frequency is calculated) overlaps with the period in the previously identified fluctuating task activity period set {[10:02, 10:08], [10:09, 10:17]}. [10:02, 10:08] overlaps with itself. Analyze the tool intervention intensity. The intervention intensity can be measured by the degree or duration of the exceeded value and calculated in combination with the intensity coefficient in the impact table. For example, for the interference of APIX at [10:02, 10:08], the frequency of exceeding the threshold is 2.5 -1.5 = 1.0 times / minute, with a duration of 6 minutes. Assuming an impact intensity coefficient of 1.2, the intervention intensity is quantified as: excess frequency x duration x impact coefficient = 1.0 x 6 x 1.2 = 7.2 units of intensity. Determine the association between progress monitoring anomalies. If tool interference exists and the interference period matches the fluctuating task period, then the progress fluctuation anomaly (the fluctuation identified in S212, such as the continuous delay fluctuation in N001) is determined to be associated with the tool call interference. Generate a list of task node segments affected by the tool call, and record the task node or segment number determined to be affected by the tool call, along with the affected period, tool, and intervention intensity.

[0076] Table 2: List of task node sections affected by tool calls;

[0077] .

[0078] See Table 2. The list of task nodes affected by tool calls shows the task progress fluctuations related to tool call interference, including the affected nodes or segments, the time period when the interference occurs, the specific interference tool, the quantified intervention intensity, and the associated task fluctuation label.

[0079] See also Figure 5 ,The specific steps for obtaining the agent node task completion fluctuation group are:

[0080] S411: Based on the list of sections of the task node affected by the tool call, filter the unmarked partition nodes, extract and record the job task list of the node, including the task number, planned process, start time and end time, and obtain the task set of the node not affected by the tool call;

[0081] Based on the list of segments of the task node affected by the tool call, for example, the list shows that segment A.1 of node N001 is affected by the tool call, filter the partition nodes that have not been marked, select those nodes that are not in the list affected by the tool call generated by S312 from all task nodes, and the nodes are considered to be nodes that are not directly affected by the tool call. For example, node N003 is not in the list affected by the tool, and node N002 is not in it either. Extract and record the job task list of the node, and obtain the specific task list of the undisturbed node from the task decomposition plan table, including the number of each task, the description of the planned execution process, the planned start time and the planned end time. For example, the job list of node N003: task number T003a, planned process "data preparation" planned start time 10:30, planned end time 10:40: task number T003b, planned process "model call", planned start time 10:40, planned end time 10:55, the job list of node N002 : Task number T002a, planned process "interface parameter configuration", planned start time 10:20, planned end time 10:25; Task number T002b, planned process "interface call", planned start time 10:25, planned end time 10:35, obtain the task set of nodes not interfered by tool calls, this set contains the detailed planned job information of all task nodes that are determined to be not directly interfered by tool calls, for example, the set contains {node: N003, task clearing Order: {Number: T003a, Process: Data preparation, Start: 10:30, End: 10:40}, {Number: T003b, Process: Model call, Start: 10:40, End: 10:55}, Node: N002, Task list: {Number: T002a, Process: Interface parameter configuration, Start: 10:20, End: 10:25}, {Number: T002b, Process: Interface call, Start: 10:25, End: 10:35}}.

[0082] S412: Calling the task set of nodes that are not interfered with by the tool call, matching the agent's daily segmented task records with the operation log data, extracting the planned task volume by task number, and calculating the running time and real-time running time to generate the agent task execution matching data set;

[0083] Call the node task set that is not interfered by the tool call. For example, the set contains tasks T003a and T003b (belonging to node N003), and tasks T002a and T002b (belonging to node N002). Match the daily segmented task records of the agent with the operation log data. From the task execution log and the more fine-grained operation log of the agent system, find the actual execution record corresponding to the task number in the set. The record contains information such as the actual start and end time of the task, the amount of task processed, etc. Extract the planned task amount by task number, and obtain the expected workload of each task from the task plan or task decomposition table. For example, the task T003a (data preparation) is calculated. It is planned to process 100 data records, task T003b (model call) is planned to process 50 requests, task T002a (interface parameter configuration) is planned to configure 10 parameters, and task T002b (interface call) is planned to initiate 20 calls. The running time and real-time running period are counted, and the actual start time, end time and total duration of the task are recorded from the Agent running log. For example, the actual execution time of task T003a is 10:32 and ends at 10:45, with a running time of 13 minutes and a real-time running period of [10:32, 10:45]. The actual execution time of task T003b is 10:46 and ends at 10:50. The running time is 4 minutes, and the real-time running time period is [10:46, 10:50]. The actual execution time of task T002a is 10:20 and 10:24, with a running time of 4 minutes and a real-time running time period of [10:20, 10:24]. The actual execution time of task T002b is 10:25 and 10:32, with a running time of 7 minutes and a real-time running time period of [10:25, 10:32]. Generate an agent task execution matching dataset, which associates planning information with actual execution information. For example, the dataset contains the following entries: {Task number: T003a, Planned task amount: 100, Planned duration: 10 minutes , real-time running time: 13 minutes, real-time running time period: [10:32, 10:45]}, {Task number: T003b, planned task amount: 50, planned duration: 15 minutes, real-time running time: 4 minutes, real-time running time: [10:46, 10:50]} {Task number: T002a, planned task amount: 10, planned duration: 5 minutes, real-time running time: 4 minutes real-time running time period: [10:20, 10:24]}, {Task number: T002b, planned task amount: 20, planned duration: 10 minutes, real-time running time: 7 minutes, real-time running time: [10:25, 10:32]}.

[0084] S413: Based on the agent task execution matching data set, determine the matching degree between task execution efficiency and operation distribution, identify efficiency fluctuation nodes and evaluate the task deviation degree, analyze the task efficiency fluctuation degree of task nodes, mark tasks with fluctuation degrees exceeding the set benchmark value as abnormal task nodes, and obtain the agent node task completion fluctuation group;

[0085] According to the agent task execution matching dataset, for example, the dataset contains information about task T003a, the degree of match between task execution efficiency and running distribution is judged. Task execution efficiency is measured by calculating the ratio of planned task volume to planned duration (planned efficiency) and the ratio of actually completed task volume (assuming it is consistent with the planned task volume) to real-time running duration (actual efficiency). For task T003a, the planned efficiency is 10010=10 requests / minute, and the actual efficiency is 100 / 13-7.69 requests / minute. For task T003b, the planned efficiency is 50 / 15≈3.33 requests / minute, and the actual efficiency is 50 / 4=12.5 requests / minute. Running distribution refers to the position of the actual execution period of the task relative to the planned period, whether it is continuous, whether it exceeds the range, and other characteristics. For example, the planned period of task T003a is [10:30, 10:40], and the actual period is [ The actual start time for task T003a is [10:32, 10:45], and the actual start time is 2 minutes late and the end time is 5 minutes late, which is a delay and extension of the whole process. The planned time for task T003b is [10:40, 10:55], and the actual time is [10:46, 10:50]. The actual start time is 6 minutes late and the end time is 5 minutes early, which is a delay and shortened process. Identify efficiency fluctuation nodes and evaluate the degree of task deviation. If the actual efficiency is significantly lower or higher than the planned efficiency, or the operation distribution is significantly different from the plan, then the task or its nodes are considered to have efficiency fluctuations. The degree of deviation can be evaluated by the ratio of actual efficiency to planned efficiency (efficiency ratio) or the absolute value of the difference. For example, the efficiency ratio of task T003a is 7.69 / 10 = 0.769, and the degree of deviation can be defined as 1-efficiency ratio| = |1-0.769| = 0.231 (i.e., 23.1% deviation). The efficiency ratio of task T003b is 12.5 / 3.33 -3.75, and the degree of deviation is |1-3.75. =2.75 (i.e., 275% deviation). Analyze the task efficiency fluctuation of the task node. The fluctuation degree is the quantitative value of the deviation degree calculated above. Tasks with fluctuations exceeding the set baseline value are marked as abnormal task nodes. The baseline value is used to define the tolerance for efficiency fluctuations. Its setting refers to the criticality of the task, the performance expectations of the agent, the average fluctuation level of the raw data, and the business requirements for efficiency stability. For example, for non-critical tasks, efficiency fluctuations within 20% are acceptable, and the baseline value is set to 0.20. For tasks with high performance requirements, fluctuations exceeding 10% require attention, and the baseline value is set to 0.10. Here, the baseline value is set to 0.20, indicating that tasks with efficiency deviations exceeding 20% are marked as abnormal. Comparing the fluctuation degree of 0.231 of task T003a with the baseline value of 0.20, because 0.231 is greater than 0.20, task T003a is marked as abnormal. Comparing the fluctuation degree of 2.75 of task T003b with the baseline value of 0.20, because 2.75 is greater than 0.20, task T003b is also marked as abnormal, resulting in an agent node task completion fluctuation group. This group contains the task node number marked as abnormal, the task number, and its related efficiency fluctuation information, such as {node number: N003, task number: T003a, efficiency deviation: 23.1%}, {node number: N003, task number: T003b, efficiency deviation: 275%}. Nodes in this fluctuation group are those task nodes whose execution efficiency or rhythm is abnormal without direct tool interference.

[0086] See also Figure 6 ,The steps for obtaining the task progress monitoring structure indicators are as follows:

[0087] S511: Call the node number and the corresponding operation time interval in the task completion fluctuation group of the intelligent node, identify the task volume difference between adjacent tasks, filter out the nodes that exceed the threshold, record the time interval and task volume change range, and obtain the progress fluctuation abnormality identification set;

[0088] Call the agent node task to complete the node number and corresponding operation time interval in the fluctuation group. For example, the fluctuation group contains node N003, and the actual execution time interval of its abnormal task T003a is [10:32, 10:45], and the actual execution time interval of task T003b is [10:46, 10:50]. Identify the task volume difference between adjacent task days. For the node in the fluctuation group (such as N003), find the similar tasks it executed on adjacent task days (for example, yesterday 10-26 and today 10-27). Or associate the planned or actual task volume of the task and calculate the absolute value of the task volume difference. For example, node N003 contains another task T003c (planned to perform data cleaning every afternoon). The actual processing task volume from 10-26 yesterday was 500, and the actual processing task volume from 10-27 today was 450. The task volume difference is 1450-500=50. Select the node that exceeds the value. The task volume difference threshold is used to define the significance of the task volume change. Its setting refers to the original task volume fluctuation range, business demand stability or experience For example, for tasks with an average daily fluctuation of no more than 20 tasks, the threshold can be set to 30, or to 10% of the average task volume (if the average task volume is 500, the threshold is 500x10%=50). Here, the threshold is set to 40, which means that a task volume difference of more than 40 is considered a significant change. Compare the task volume difference of 50 for task T003c with the threshold of 40, because 50>40, node N003 is screened out due to a significant change in task volume, and record the time interval and task The quantity change range records the date range in which the task quantity changes (for example, 10-26 to 10-27) and the specific task quantity change value or percentage (for example, change of -50 items or change of -10%), and obtains the progress fluctuation abnormality identification set, which contains the task nodes that are marked because the task quantity changes significantly on adjacent dates and their change details. For example, the set contains {node: N003, associated task: T003c, change area: [10-26, 10-27], task quantity change: -50 items}.

[0089] S512: Based on the knowledge base resources corresponding to the nodes in the progress fluctuation abnormality identification center, identify the node resource input ratio sequence, extract the abnormal distribution interval and compare the critical coefficient, record the ratio deviation direction and node number, and form a task resource matching deviation indicator group;

[0090] Based on the knowledge base resources corresponding to the nodes in the progress fluctuation abnormal identification set, for example, the abnormal identification set includes node N003 and its associated task T003c. The knowledge base resources associated with this task (data cleaning) are Data Cleansing Rule Document v2.0 and Abnormal Data Processing Guide. The node resource input ratio sequence is identified. The resource input ratio can be defined as the frequency (such as the number of visits per minute) or duration of the Agent accessing or using a specific knowledge base resource during the execution of the node task, or the ratio sequence of the resource update frequency to the task execution frequency, such as node When executing task T003c at point N003, the access frequency for "Data Cleansing Rules Document v2.0" was 0.5 times / minute from October 26th yesterday. When executing the same task today from October 27th, the access frequency was 0.2 times / minute, forming a resource input ratio sequence..., 0.5 (yesterday), 0.2 (today), .... Anomaly distribution intervals are extracted and compared with critical coefficients. Anomaly distribution intervals refer to periods of significant change in the resource input ratio sequence. For example, the ratio dropped from 0.5 to 0.2 from yesterday to today, with a change of 0.5 - 0.2 = 0. 3. The critical coefficient is used to define the significance of the change in the resource input ratio. Its setting refers to the importance of the resource and the stability of the original ratio. For example, for core rule documents, the access frequency should not fluctuate greatly, and the critical coefficient is set lower (such as 0.1). For reference documents, the coefficient with high tolerance for fluctuation is set higher (such as 0.3). Here, the critical coefficient is set to 0.25, which means that the resource input is considered abnormal if the absolute value of the ratio change is greater than 0.25. Here, the ratio change range is 0.3>0.25, and the abnormal distribution range is from yesterday to today. Comparing the critical coefficient, we can confirm the abnormality. Identify the significance of the change, record the ratio offset direction and node number, the ratio drops from 0.5 to 0.2, the offset direction is "downward" or "negative deviation", record the node number N003 and its associated task T003c, and form a task resource matching offset indicator group. This group includes the task nodes marked due to abnormal changes in resource input ratios and their resource matching deviations. For example, the indicator group includes {node: N003, associated task: T003c, associated resource: data cleaning rule document v2.0, ratio offset direction: downward, ratio change: -0.3}.

[0091] S513: Based on the task resource matching offset indicator group, extract the knowledge base resource delivery and task execution time periods, analyze the difference between the resource delivery cycle and the operation cycle, identify the task time synchronization offset distribution table, sort and mark the differences according to the progress benchmark, and generate the task progress monitoring structure indicator;

[0092] According to the task resource matching offset indicator group, for example, the indicator group shows that the T003c task of node N003 has a resource matching decline offset, extract the knowledge base resource delivery and task execution time period, obtain the planned update, release or available time period of the knowledge base resource from the resource management, and obtain the planned or actual execution time period of the relevant task from the task plan or execution record. For example, the "Data Cleansing Rule Document v2.0" is scheduled for routine updates every Monday morning at 10:00 am, and an emergency temporary update was carried out at 15:00 pm today, 10-27 (Tuesday), while the T003c task of node N003 is scheduled to be executed from 14:00 to 16:00 pm every day, and the actual execution period today is 14:05-16:10. Analyze the difference between the resource delivery cycle and the operation cycle, and compare the matching degree and timing relationship between the resource update or available cycle and the task execution cycle. For example, today's resources were updated at 15:00 in the middle of the task execution, and the task started at 14:05, which means that the old version of the rule was used at the beginning of the task. , or the Agent needs to interrupt the current execution to adapt to the new rules, resulting in a decrease in efficiency (related to the efficiency fluctuation in T003c in S413), or combined with the change in task volume identified in S511 (for example, the number of tasks decreased by 50), analyze whether the delayed resource update or the change in resource content caused some data to be marked as abnormal and not processed. Identify the task time synchronization offset distribution table, build a table to record task nodes, associated resources, planned resource delivery time, actual resource delivery time, planned task execution time period, actual task execution time period, as well as timing differences and potential synchronization issues, and sort and mark the differences according to the progress baseline. According to the planned progress baseline of the entire project or phase, such as the total task completion time and the key milestone achievement rate, sort the entries in the task time synchronization offset distribution table, for example, by the degree of impact on the overall progress, the amount of tasks involved, the deviation duration, or the node criticality, and mark the deviation degree or type (for example, "resource lag synchronization offset" or "resource content change impact");

[0093] Table 3: Task progress synchronization offset distribution table;

[0094] .

[0095] As shown in Table 3, the task progress synchronization deviation distribution table summarizes various types of progress deviations related to time synchronization, covering aspects such as document updates, task execution, and resource allocation. By analyzing structured deviation data, we can generate structural indicators for task progress monitoring. These indicators quantify and summarize structural issues in the synchronization of task progress with knowledge base resources, tools, and the agent's own execution efficiency. For example, indicators may include "number of nodes affected by resource synchronization" (e.g., one node, N003, in Table 3), "average resource lag duration" (e.g., N003's resource lag is nearly 2 hours), "expected rework rate due to resource mismatch" (assessed based on the impact of resource changes on task outcomes), "proportion of delayed nodes due to agent efficiency fluctuations" (e.g., N003 due to efficiency fluctuations in T003a and T003b), and "accumulated delay duration due to tool call interference" (calculated based on the intervention intensity and duration in Table 2). These indicators are used for high-level progress monitoring, problem diagnosis, and optimization suggestions.

[0096] The multi-model agent collaboration system is used to execute the multi-model agent collaboration method described above, and the system includes:

[0097] The task status extraction module obtains the knowledge base documents, tool interface configuration information and external API call records in the task scenario, extracts the agent task node number and time interval, compares the real-time execution status with the task node identifier, marks the node tasks with inconsistent time, and generates a node status offset label group;

[0098] The node trend classification module is based on the node state offset label group, and is an agent-type intelligent assistant with positioning reasoning, planning, and execution capabilities. It extracts the task breakdown table and planned time, identifies the difference between the real-time task time and the planned time, and classifies and labels the difference type to generate a task node progress trend identification group.

[0099] The tool call identification module, based on the task node progress trend identification group, screens the fluctuating and lagging task nodes, locates the corresponding partitions, extracts the tool call frequency interference period, determines the overlap with the task time period, screens the frequent interference sections, and generates the task progress tool call interference mapping set;

[0100] The task efficiency diagnosis module calls the interference mapping set based on the task progress tool, eliminates interfering segment tasks, extracts agent task records and operation logs, matches task assignments with operation periods, calculates the operation volume to task ratio, identifies dense but inefficient node tasks, and obtains the node task execution deviation set;

[0101] The resource allocation analysis module locates the agent resource input records based on the node task execution deviation set, extracts the ratio of task volume to resource allocation, compares the difference between execution cycle and input cycle, maps resource usage and task progress status, and forms task progress monitoring structure indicators.

[0102] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A multi-model agent collaboration method, characterized in that: The following steps are involved: S1: Obtain the knowledge base documents, tool interface configuration information, and external API call records in the task scenario, extract the execution plan number and time interval of the agent task node, compare the agent's real-time execution status with the plan node identifier, identify the task node number that deviates from the plan, and generate a task node synchronization offset list; S2: Based on the task node synchronization offset list, select an agent-type intelligent assistant, extract the task decomposition start and end time in combination with the task decomposition table, calculate the difference with the planned time point and classify the trend, and generate a task progress trend label group; The specific steps for obtaining the task progress trend tag group are as follows: S211: Based on the task node synchronization offset list, identify an agent-type intelligent assistant with reasoning-planning-execution capabilities and a task decomposition table, extract the real-time task decomposition start and completion time, calculate the start and end time difference of the task segment, and compare it with the planned task node time difference to obtain the task time offset value; S212: Calling the task time offset value, combining the segment distribution, offset trend and adjustment frequency, collecting task node offset data, identifying and calculating the trend offset degree by segment number, determining the trend direction based on the adjustment frequency, and obtaining a task progress trend label group; S3: Call the task progress trend tag group, extract the fluctuation task partition number, identify the regional tool call frequency data, compare the tool call period with the fluctuation task time period, record the number of overlapping time periods, and generate a list of task node segments affected by the tool call; The steps for obtaining the list of sections of the task node affected by the tool call are as follows: S311: calling the task progress trend tag group, filtering the fluctuation offset task partition number, extracting the task time period according to the node association table, processing the segment task time period according to the time dimension, identifying the task time period index table, and obtaining the fluctuation task activity time period set; S312: Based on the fluctuating task activity period set, collect tool call frequencies for the same period, identify the impact table, determine interference based on the interference threshold, match it with the fluctuating task construction period, analyze tool intervention intensity, determine progress monitoring anomaly correlation, and generate a list of task node sections affected by tool calls; S4: Based on the list of sections of the task node affected by the tool call, filter the unaffected partition node tasks, extract the agent segment task records and operation log data, map the daily task assignment volume and the agent operation period, determine whether there is an inefficient distribution of tasks, and obtain the agent node task completion fluctuation group.

2. The multi-model agent collaboration method according to claim 1, characterized in that: The task node synchronization offset list includes the task offset number, execution status label, time offset, and task node classification; the task progress trend label group includes the progress offset level, trend change type, plan comparison result, and task association number; the task node affected by tool call segment list includes tool call type, affected time segment, number of overlapping time periods, and affected task number; the intelligent node task completion fluctuation group includes task uneven distribution number, running time record, task completion deviation, and intelligent agent operation matching degree.

3. The multi-model agent collaboration method according to claim 1, characterized in that: The steps for obtaining the task node synchronization offset list are as follows: S111: Obtain the knowledge base documents, tool interface configuration information, and external API call records in the task scenario, extract the agent task node plan time and number, match the knowledge base document update time and coordinates, compare the matching document time range with the node plan time, and generate the partition node document record period; S112: Based on the partition node document record period, the overlap with the task node planned time interval is judged, the ratio of the overlap time to the total duration is extracted, the node numbers with a ratio lower than the benchmark value are screened, and the node document status label is updated to obtain the partition node progress coverage deviation rate; S113: According to the partition node progress coverage deviation rate, the deviation status of the task node number is determined, the task node number with a deviation rate higher than the node synchronization threshold is identified, the node number, document coverage information and deviation rate value are integrated, and a task node synchronization offset list is generated.

4. The multi-model agent collaboration method according to claim 1, characterized in that: The steps for obtaining the agent node task completion fluctuation group are specifically as follows: S411: Based on the list of sections of the task node affected by the tool call, filter the unmarked partition nodes, extract and record the job task list of the node, including the task number, planned process, start time and end time, and obtain the task set of the node not affected by the tool call; S412: Calling the node task set that is not interfered with by the tool call, matching the agent's daily segmented task records with the operation log data, extracting the planned task amount by task number, counting the running time and the real-time running time period, and generating an agent task execution matching data set; S413: Based on the intelligent agent task execution matching data set, determine the matching degree between the task execution efficiency and the operation distribution, identify the efficiency fluctuation nodes and evaluate the task deviation degree, analyze the task efficiency fluctuation degree of the task nodes, mark the tasks whose fluctuation degree exceeds the set benchmark value as abnormal task nodes, and obtain the intelligent agent node task completion fluctuation group.

5. The multi-model agent collaboration method according to claim 1, characterized in that: The method further comprises step S5: S5: Call the task completion fluctuation group of the agent node, extract the abnormal fluctuation task group, calculate the ratio of resource input to task amount in the knowledge base, record the difference distribution between resource input cycle and task execution cycle, and generate task progress monitoring structure indicators; The task progress monitoring structure indicators include resource allocation ratio, task intensity level, execution cycle deviation, and task efficiency index.

6. The multi-model agent collaboration method according to claim 5, characterized in that: The steps for obtaining the task progress monitoring structure indicator are specifically as follows: S511: Call the node number and the corresponding operation time interval in the task completion fluctuation group of the agent node, identify the task volume difference between adjacent tasks, filter out the nodes that exceed the threshold, record the time interval and task volume change range, and obtain the progress fluctuation abnormality identification set; S512: Based on the knowledge base resources corresponding to the nodes in the progress fluctuation abnormality identification set, identify the node resource input ratio sequence, extract the abnormal distribution interval and compare the critical coefficient, record the ratio deviation direction and node number, and form a task resource matching deviation indicator group; S513: According to the task resource matching offset indicator group, extract the knowledge base resource delivery and task execution time period, analyze the difference between the resource delivery cycle and the operation cycle, identify the task time synchronization offset distribution table, and sort and mark the difference according to the progress benchmark to generate the task progress monitoring structure indicator.

7. A multi-model intelligent agent collaborative system, characterized in that: The system is used to implement the multi-model agent collaboration method according to any one of claims 1 to 6, and the system includes: The task status extraction module obtains the knowledge base documents, tool interface configuration information and external API call records in the task scenario, extracts the agent task node number and time interval, compares the real-time execution status with the task node identifier, marks the node tasks with inconsistent time, and generates a node status offset label group; The node trend classification module locates the agent-type intelligent assistant with reasoning-planning-execution capabilities based on the node state offset label group, extracts the task decomposition table and planned time, identifies the difference between the real-time task time and the planned time, classifies and labels the difference type, and generates a task node progress trend identification group; The tool call identification module selects the fluctuating and lagging task nodes based on the task node progress trend identification group, locates the corresponding partitions, extracts the tool call frequency interference period, determines the overlap with the task time period, selects the frequent interference sections, and generates a task progress tool call interference mapping set; The task efficiency diagnosis module calls the interference mapping set based on the task progress tool, eliminates the interfering segment tasks, extracts the agent task records and operation logs, matches the task assignments with the operation time periods, calculates the operation volume and task ratio, identifies the dense but inefficient node tasks, and obtains the node task execution deviation set; The resource allocation analysis module locates the agent resource input record based on the node task execution deviation set, extracts the task volume and resource allocation ratio, compares the execution cycle and input cycle difference, maps resource usage and task progress status, and forms a task progress monitoring structure indicator.

Images