Multi-agent cooperative task processing method

Abstract

The application discloses a kind of multi-agent collaborative task processing methods, wherein method includes receiving structured task by preconfigured native direct hair channel;Receive pending mail text by preconfigured mail gateway channel, and based on information extraction agent, task element extraction is carried out to the pending mail text, to output structured task data;Task routing agent matches the responsibility subject corresponding to structured task data;Cooperative scheduling agent identifies the subtask dependency relationship in task and completes disassembly, generates parallel or serial task execution scheduling scheme, and after task scheduling is completed, it is transferred to execution monitoring link;In execution monitoring link, execution monitoring agent real-time tracks the execution state of structured task, and based on the characteristics of the execution state of task, dynamically determines the time and mode of urging, and pushes task reminder to corresponding person in charge. Each agent data intercommunication forms task full-process automation closed-loop management and control.

Description

Technical Field

[0001] This invention relates to the field of information processing technology for multi-agent collaboration, and specifically to a task processing method for multi-agent collaboration. Background Technology

[0002] Since its inception, email systems have long been used by various organizations as a core tool for internal management and communication due to their advantages such as openness, stability, and decentralization. However, email was originally designed for asynchronous communication, and its underlying data structure does not include any fields required for task management—such as responsible person, deadline, task status, priority, and feedback loop. This design creates a fundamental mismatch between email and its actual application in management scenarios. Eppler et al. pointed out that information overload not only reduces individual decision-making efficiency but also leads to the omission of key information and delayed responses at the organizational level. This judgment is fully confirmed in email-driven task management scenarios.

[0003] Taking an organization as an example, a statistical analysis of over k business emails from a quarter in 2025 revealed that emails containing explicit task instructions accounted for over p%. However, the task information in these emails was entirely in unstructured text format. For instance, some emails wrote "Please submit by May 20th," others wrote "Deadline: May 20th," some wrote "Deadline: Next Thursday," and still others simply omitted the deadline, using vague instructions lacking a clear deadline. This high heterogeneity in expression makes traditional rule-based keyword matching methods ineffective in handling such situations.

[0004] Therefore, from an execution perspective, managers spend a significant amount of time each week manually processing emails, such as extracting tasks, categorizing them, setting reminders, expediting, and reviewing them; the overdue rate for tasks is generally high. From a management perspective, the organization lacks a holistic view of tasks, management effectiveness cannot be quantitatively assessed, and decisions rely on experience rather than data. From a knowledge management perspective, questions, experiences, and lessons learned during the task process are scattered in individual inboxes, failing to form organizational memory (organizational memory is defined as the retrieval and maintenance of information stored within the organization that can be used for decision-making). The implicit loss of task knowledge in the email system constitutes precisely the biggest break in organizational memory. Summary of the Invention

[0005] The main objective of this invention is to provide a multi-agent collaborative task processing method to address the shortcomings of related technologies.

[0006] To achieve the above objectives, according to a first aspect of the present invention, a multi-agent collaborative task processing method is provided, comprising: receiving a structured task through a pre-configured native direct sending channel; receiving a text message to be processed through a pre-configured email gateway channel, and extracting task elements from the text message based on an information extraction agent to output structured task data containing task subject, source department, responsible person, deadline, priority, and task type; a task routing agent matching the responsible entity corresponding to the structured task data; a collaborative scheduling agent identifying and decomposing sub-task dependencies within the task, generating a parallel or serial task execution scheduling scheme, and transferring the task to the execution monitoring stage after scheduling is completed; in the execution monitoring stage, an execution monitoring agent tracking the execution status of the structured task in real time, and dynamically determining the timing and method of reminders based on the characteristics of the task's execution status, and pushing task reminders to the corresponding responsible person.

[0007] Optionally, the method also includes the execution monitoring agent counting the trigger frequency of similar tasks during task tracking, and notifying the knowledge management agent when a preset trigger condition is met. The knowledge management agent then extracts reusable experience from the task execution process data and stores it in the knowledge base to form organizational assets.

[0008] Optionally, the task element extraction of the email text to be processed includes two parsing mechanisms: a rule engine and an LLM parser. The rule engine is used to extract elements for well-formatted notification emails by matching keywords. For emails with free formatting, the LLM parser is used to extract elements by semantic understanding. The rule engine and the LLM parser serve as backups for each other, and the extraction results are cross-validated.

[0009] Optionally, the method also includes human-machine collaborative correction: the structured task data parsed by the email gateway channel is silently stored, and a parsing result verification notification is pushed to the corresponding responsible person's terminal; if a correction instruction submitted by the responsible person's terminal for the parsing result is received, the original email text, parsing result and user correction result are recorded as annotation samples; when the cumulative number of annotation samples reaches a preset threshold, the adaptation and fine-tuning of the LLM parser is triggered to update the LLM parser parameters.

[0010] Before the matching of responsible parties is completed and the task execution scheduling plan is generated, the sub-step of permission pre-configuration is also included: through the permission management intelligent agent based on the role-based and hierarchical permission system, the visibility range of corresponding task information is configured for each matched responsible party; the isolation rules for feedback content across responsible parties are set; and the hierarchical matching rules for notification objects corresponding to task overdue warnings are pre-configured.

[0011] Optionally, the execution monitoring agent dynamically determines the timing and method of reminders based on the characteristics of the task's execution state, including: modeling the task reminder decision as a Markov decision process, using a feature vector consisting of the remaining time before the deadline, task priority, task type, and the responsible person's historical response rate as the state space; using an action set consisting of immediate reminder, delayed reminder, escalation notification, and no reminder as the action space; predicting the task overdue probability based on the current state space using a preset classification model; calculating a reward value using the task overdue probability, the number of reminders sent, and the user satisfaction score as input; and iteratively learning the optimal reminder strategy using a Q-learning algorithm, outputting the optimal reminder action for the corresponding state.

[0012] Optionally, during the task element extraction process, it is identified whether the email text contains preset high-risk clauses; if high-risk clauses are identified, the reward weight corresponding to the task overdue probability in the Markov decision process is increased to a preset multiple, and the task is subject to an enhanced reminder strategy by moving the reminder trigger threshold forward.

[0013] Optionally, extracting reusable experience from task execution data through a knowledge management agent includes: statistically analyzing business consultation questions generated during the execution of similar tasks; automatically generating standardized answers and storing them in the knowledge base when the number of times the same business consultation question is repeatedly asked reaches a preset threshold; automatically extracting the common process structure and standard execution rhythm of the same type of task when the number of times the same type of task is repeated reaches a preset threshold to generate standardized task templates and store them in the knowledge base; and extracting temporal dependencies between tasks based on historical task data using association rule mining algorithms to generate task association recommendation rules and store them in the knowledge base.

[0014] Optionally, identifying subtask dependencies within a structured task through a collaborative scheduling agent includes: automatically breaking down multi-step composite tasks to generate a subtask dependency graph; and automatically solidifying recurring task combinations into standardized task execution templates for use as scheduling references for subsequent similar tasks.

[0015] The method also includes: when the execution monitoring agent tracks the execution status of structured tasks in real time, it generates an early warning escalation signal and sends it to the permission management agent when it detects the risk of task overdue; after the permission management agent determines the scope of the early warning notification object based on the preset permission rules, it triggers the user interaction agent to match the corresponding notification channel and completes the targeted delivery of the early warning information.

[0016] This embodiment presents a multi-agent collaborative task processing method, which includes receiving structured tasks through a pre-configured native direct sending channel; receiving email text to be processed through a pre-configured email gateway channel, and extracting task elements from the email text based on an information extraction agent to output structured task data containing task subject, source department, responsible person, deadline, priority, and task type; a task routing agent matching the responsible entity corresponding to the structured task data; a collaborative scheduling agent identifying and decomposing sub-task dependencies within the task, generating a parallel or serial task execution scheduling scheme, and transferring the task to the execution monitoring stage after scheduling; in the execution monitoring stage, the execution monitoring agent tracks the execution status of the structured tasks in real time, and dynamically determines the timing and method of reminders based on the characteristics of the task's execution status, pushing task reminders to the corresponding responsible person. A dual-channel parsing agent completes the structured conversion of heterogeneous tasks; routing and scheduling agents achieve responsibility matching and process decomposition; a monitoring agent dynamically reminds users based on reinforcement learning; a knowledge agent accumulates business experience at three levels; and data is shared among the agents to form a fully automated closed-loop management system for the entire task process. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a multi-agent collaborative task processing method according to an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention 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 for the embodiments of the invention 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.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] The system architecture applicable to this embodiment includes a first data layer for storing structured task data, user and organizational permission data, a knowledge base (FAQ, experience summaries, work templates), and operation logs. A relational database manages structured information, and a vector database manages unstructured knowledge. The second layer is the intelligent parsing layer—the system's perception module. When an email enters the system, this layer calls a large language model to extract six key elements of the task: task subject, source department, responsible person, deadline, priority, and task type. For emails with incomplete information, the system does not force guessing but automatically generates structured follow-up questions. The third layer is the collaborative decision-making layer, the core of multi-agent collaboration. Seven types of agents in this layer complete task distribution, scheduling, monitoring, knowledge accumulation, and permission control according to predefined collaboration protocols. The fourth layer, the execution service layer, provides basic service capabilities such as structured form generation, multi-template notification sending, data statistics and summarization, version control, and permission verification, which can be called by upper-layer agents at any time. The fifth layer is the user interaction layer, providing three interaction channels: web, mobile, and API interfaces, supporting users of different roles to access the system in an appropriate manner. The five-layer architecture described above is a reference framework. Different organizations have different organizational structures, approval processes, and permission systems, and can be tailored and adjusted according to actual needs. For example, if a mature OA system already exists, the user interaction layer can be reused from the existing OA system, requiring only proper data integration.

[0023] According to embodiments of the present invention, a multi-agent cooperative task processing method is provided, such as... Figure 1 As shown, steps 101 to 103 are included below: Step 101: Receive structured task data and unstructured email text to be processed directly from two input channels: the pre-configured native direct sending channel and the email gateway channel. Extract task elements from the email text to be processed using an information extraction agent to output structured task data containing task subject, source department, responsible person, deadline, priority, and task type.

[0024] In this step, task input may take two forms: first, standardized tasks actively created by users in the business system, which possess complete structured fields and are deterministic structured input; second, informal tasks carried in emails, which exist entirely in natural language text and are unstructured, uncertain input. Traditional task management systems often only support one of these input forms, either forcing users to manually enter tasks, disrupting the existing email workflow, or only being able to process email tasks and unable to cover standardized creation scenarios.

[0025] This solution utilizes two parallel input channels—a native direct mail channel and an email gateway—to precisely adapt to the two input formats mentioned above. The native direct mail channel handles structured tasks, directly reusing their structured attributes; the email gateway channel handles unstructured email text, performing structured transformation through element extraction. Both channels simultaneously cover all task entry points within the organization, preserving the efficiency of standardized tasks while remaining compatible with email workflow usage habits, thus resolving the issue of heterogeneous task input at its source.

[0026] As an optional implementation of this embodiment, the task element extraction of the email text to be processed includes two parsing mechanisms: a rule engine and an LLM parser. The rule engine is used to extract elements for well-formatted notification emails by keyword matching, while the LLM parser is used to extract elements for freely formatted emails by semantic understanding. The rule engine and the LLM parser serve as parsing backups for each other, and the extraction results are cross-validated.

[0027] In this optional implementation, the rule engine performs deterministic matching based on keywords and preset rules. The processing speed is in the millisecond range and the computing power consumption is extremely low. However, it can only adapt to templated emails with fixed formats and standardized expressions. Once the email expression deviates from the preset rules, the parsing accuracy will drop significantly. In addition, it is not reusable across business scenarios and cannot cover all email types.

[0028] The rules engine channel is enabled for scheduling well-formatted notification emails. Based on a pre-configured keyword library and regular expression matching rules, it performs deterministic pattern matching on the email text to quickly extract six core elements: task subject, source department, responsible person, deadline, priority, and task type. This channel offers fast processing speed and low computational consumption, making it suitable for the most prevalent standardized notification scenarios and ensuring efficient batch email processing.

[0029] LLM parsers extract elements based on semantic understanding, making them adaptable to non-standard emails with flexible wording and higher parsing accuracy. However, their processing speed is in the sub-second range, and their computational cost is far higher than that of rule engines. If all emails were to use LLM, it would result in unnecessary waste of computational resources and response delays. Therefore, an architecture that combines format determination and dual-channel scheduling is adopted to achieve the optimal balance between parsing speed, accuracy, and operating cost while ensuring parsing coverage across all scenarios.

[0030] The LLM parser channel is enabled for emails with flexible formats. Leveraging the semantic understanding capabilities of a large language model, it infers and identifies task information expressed in natural language, extracting accurate six elements of the task from non-standard expressions. This channel does not rely on fixed rules and can cover various personalized business collaboration emails, ensuring parsing accuracy in complex scenarios.

[0031] If a single channel's parsing results in missing fields, the system automatically calls another channel for semantic completion, ensuring the integrity of the output structured data fields and avoiding incomplete parsing results due to the limitations of a single channel's capabilities. For core elements such as deadlines and responsible parties, a two-way comparison is performed using the dual-channel results. If the outputs from both mechanisms are consistent, the final structured data is directly output. If discrepancies exist, corrections are made based on the LLM's semantic understanding result, or the data is marked as requiring verification and pushed for manual confirmation, reducing the probability of misjudgment using a single parsing method.

[0032] For example, let the input text be T, and call LLM to generate structured output O = (title, dept, assignee, deadline, priority, category). Employ a few-shot prompting strategy, embedding 3-5 parsing examples in the prompt words.

[0033] Dual-channel parsing: Connects to two input channels. "Native direct sending channel"—users directly issue tasks within the system, resulting in naturally structured data. "Email gateway channel"—received emails automatically undergo LLM (Low-Level Modeling) for six-factor extraction. For well-formatted notification emails, the rules engine takes precedence (faster and less computationally efficient); for freely formatted emails, LLM is used.

[0034] The information extraction agent extracts structured task information from unstructured text. Its core mechanism is a dual-channel approach combining LLM parsing and a rule engine, which automatically generates follow-up questions when information is incomplete. It can convert unstructured text into structured tasks by extracting six core elements: task subject, source department, responsible person, deadline, priority, and task type, mapping fragmented email text into manageable task entities. The dual input channels cover both standardized task creation and email-based unstructured tasks, preserving the accuracy and efficiency of in-system task entry while remaining fully compatible with long-standing enterprise email workflows. Users can achieve structured task management without changing their existing operating habits, significantly reducing system deployment and user adaptation costs.

[0035] It should be noted that the understanding of LLM is probabilistic rather than deterministic. Asking the same question twice in an email might yield two slightly different answers. For information that cannot be ambiguous, such as "deadlines," a probabilistic understanding carries risks. Therefore, a "dual-channel parsing + human-machine collaborative closed-loop" mechanism was designed.

[0036] As an optional implementation of this embodiment, the method further includes human-machine collaborative correction: the structured task data parsed by the email gateway channel is silently stored, and a parsing result verification notification is pushed to the corresponding responsible person's terminal; if a correction instruction submitted by the responsible person's terminal for the parsing result is received, the original email text, parsing result and user correction result are recorded synchronously as annotation samples; when the cumulative number of annotation samples reaches a preset threshold, the adaptation and fine-tuning of the LLM parser is triggered to update the LLM parser parameters.

[0037] In this optional implementation, silent storage means that the system first stores the structured task data generated by parsing the email gateway channel into the background database as temporary data to be confirmed. This process is completed automatically in the background without requiring the user to actively trigger the operation, and without forcibly interrupting the user's existing workflow with pop-ups. After the data is stored in the database, the system only pushes the parsing results to the corresponding responsible person in the form of a lightweight notification, prompting the user to verify as needed. If the result is correct, it will take effect by default. If there is a deviation, the user can click to correct it.

[0038] When the responsible party submits a correction instruction for the parsing results, the system simultaneously records three sets of corresponding data, which together constitute a complete labeled sample: the original email text (corresponding to the model's input source data); the automatic parsing result (corresponding to the original output of the LLM parser); and the user-corrected result (corresponding to manually labeled ground truth data). The core logic of this step is data labeling, which automatically transforms the process data into training samples for supervised learning. This eliminates the need for dedicated labeling personnel and avoids additional work time, naturally accumulating high-quality domain data within daily business processes. This solves the pain point of high data acquisition costs for fine-tuning large models.

[0039] When the cumulative number of labeled samples reaches a preset threshold, the domain adaptation fine-tuning of the LLM parser is automatically triggered, optimizing model parameters using the accumulated business corpus of the organization. The logic of this step is data-driven autonomous iteration. General-purpose models have limited understanding of the business terminology, expression habits, and management styles of specific organizations or enterprises. By continuously injecting labeled data that fits the actual context of the organization, the parser can gradually adapt to the corresponding business scenarios, and the parsing accuracy continues to improve with the system's runtime.

[0040] The structured data extracted by this implementation is first "silently stored in the database," and then a notification is pushed to the user: "Your new task has been automatically parsed: [Task Description], Responsible Person: [Name], Deadline: [Date]. If there are any errors, please click to correct." Each time the user clicks "correct," three pieces of information are recorded: (1) the original email text (input); (2) the automatic parsing result (AI output); and (3) the result after user correction (human annotation). The combination of these three pieces of information constitutes a high-quality training data, which is naturally generated in normal daily use and does not require anyone to spend time "annotating." Every time a certain number of corrected data entries are accumulated, the model is automatically retrained to adapt to the unit's language habits and management style.

[0041] The complete process of human-machine collaboration closed loop is as follows: flowchart LR EMAIL [email input] --> LLM [Six elements of LLM parsing] LLM --> SILENT [Silent Ingestion] SILENT --> NOTIFY [Push notification of parsing results] NOTIFY --> USER {User Verification} USER -->|Confirmed Correct| DONE [Data Accuracy] USER --> | Click to correct | CORRECT [Record the three elements] CORRECT --> |Original Email| REC1 [(Original Text)] CORRECT --> |AI Output| REC2[(AI parsing result)] CORRECT -->|Manual annotation| REC3[(Revised result)] REC1 --> TRAIN [Cumulative Training Data] REC2 -->TRAIN REC3 -->TRAIN TRAIN --> |≥[Preset training trigger threshold]| RETRAIN [Trigger model retraining] RETRAIN --> LLM DONE -->KB[(Knowledge Base)].

[0042] Step 102: The task routing agent matches the responsible entity corresponding to the structured task data; the collaborative scheduling agent identifies the sub-task dependencies within the task and completes the decomposition, generating a parallel or serial task execution scheduling scheme, and after the task scheduling is completed, it flows to the execution monitoring stage.

[0043] In this step, the task routing agent automatically distributes tasks to the correct executors. It implements a task-role-person mapping, using the source department, responsible person, and task type fields from structured task data as input. Based on pre-configured mapping rules, such as "task type-role-person subject" mapping rules, it automatically matches the multi-level responsible entities corresponding to the task, including three levels: the primary executor, collaborating executors, and the management level (for copying). Simultaneously, it binds corresponding permission scopes, including setting visibility and operation permissions for task information based on role level. The primary executor can see all task information and has editing permissions, collaborating executors can only see the task nodes they are responsible for, and the management level (for copying) can only see the global progress without operation permissions. For cross-departmental transfer scenarios, it supports setting the visibility level of information after transfer, achieving information boundary control in the task flow.

[0044] The core of this logic is "synchronous matching of rights and responsibilities," which avoids the problems of "responsibility assigned to individuals but authority mismatch" and "information forwarding out of control" in traditional task allocation, and ensures that the responsibility boundaries and information boundaries of task flow are determined synchronously.

[0045] Task dependencies are analyzed through a collaborative scheduling agent to coordinate parallel / serial execution. Identifying subtask dependencies within structured tasks includes: automatically decomposing multi-step composite tasks to generate a subtask dependency graph; automatically solidifying recurring task combinations into standardized task execution templates for future scheduling reference of similar tasks. The scheduling scheme is a task execution scheduling scheme that includes a list of subtasks, dependencies, and execution order. Multi-step composite tasks, i.e., a single structured task containing multiple interconnected and parallel sub-tasks that cannot be completed in one step, can be decomposed into the smallest independent executable subtask unit by relying on the task theme, task type, and task description fields within the structured task data, combined with historical samples of similar task decompositions. Based on the task text and historical execution order, two types of constraints between subtasks are identified: ① Sequential dependency: the output of the previous subtask is a prerequisite for the execution of the next subtask, and they must be executed sequentially; ② Parallel dependency: there are no prerequisite constraints between subtasks, and they can be carried out synchronously. Output a subtask dependency graph, recording a complete list of subtasks, their start and end times, and the dependency constraints between each pair of subtasks in a visual / structured data format. This provides the execution monitoring agent with a basis for full-node tracking, allowing for progress tracking and triggering reminders for each node in the dependency graph.

[0046] Furthermore, a task combination is defined as a set of fixed subtasks and fixed dependencies generated after multiple decompositions of the same type of composite task, which constitutes a stable task combination. When the same type of composite task is repeatedly initiated and decomposed to generate highly similar subtask dependency graphs, and a preset repetition threshold is reached, temporary and personalized subtasks in each task are removed, while the common subtask structure, serial / parallel dependencies, and standard execution interval rhythm of multiple rounds of tasks are retained and encapsulated into a standardized task execution template. When the same type of structured task is initiated again, the collaborative scheduling agent directly calls this template without repeated semantic decomposition and dependency identification, and generates a complete subtask dependency graph with one click, unifying the execution process of similar tasks.

[0047] For example, if a composite task of the "quarterly inventory count" type is initiated multiple times consecutively, and the sub-task lists and sequential dependencies generated each time are highly consistent, reaching the preset repetition threshold: 1) The collaborative scheduling agent compares the subtask dependency graphs of multiple rounds of historical tasks and removes temporary subtasks added in a single instance (such as temporary special spot checks). 2) Extract the general fixed process: A. Inventory → B. Statistics → C. Reporting → D. Rectification, and the standard execution time of each sub-task; 3) Solidify the entire set of sub-task structures, dependencies, and standard execution rhythm into the "Standardized Execution Template for Quarterly Inventory Counting"; 4) Reusability: When the operations staff initiates the "Quarterly Inventory Check" task again, the collaborative scheduling agent can directly load the template and automatically output the complete subtask dependency graph. There is no need to re-decompose the semantics, and a complete execution plan can be generated with one click.

[0048] As an optional implementation method in this embodiment, before the matching of responsible entities is completed and the task execution scheduling scheme is generated, a permission pre-configuration sub-step is also included: the permission management intelligent agent configures the visibility range of corresponding task information for each matched responsible entity based on the role-based and hierarchical permission system; sets the feedback content isolation rules across responsible entities; and pre-configures the notification object hierarchical matching rules corresponding to the task overdue warning.

[0049] In this optional implementation, the access control agent dynamically controls the visibility of information and approval permissions. Four core roles are defined: task initiator, management layer, department liaison, and task executor. This role system is designed based on Mintzberg's framework for analyzing organizational roles and information rights—information visibility within an organization should match role responsibilities, rather than simply being "all or nothing." Specifically, department liaisons can set visibility levels when forwarding tasks, addressing the pain point of uncontrollable information visibility in traditional email "forwarding" models. In document review scenarios, the access control agent also implements "feedback content isolation"—opinions submitted by one department are not visible to other departments, avoiding the convergence effect of opinions due to herd mentality. This design choice is not technically necessary, but rather an explicit manifestation of implicit constraints in management scenarios—in policy revisions, genuine opinions are more valuable than "polite words."

[0050] As an optional implementation in this embodiment, when the monitoring agent tracks the execution status of structured tasks in real time, it generates an early warning escalation signal and sends it to the permission management agent when it detects a risk of task overdue. After the permission management agent determines the scope of the early warning notification object based on preset permission rules, it triggers the user interaction agent to match the corresponding notification channel and completes the targeted delivery of the early warning information.

[0051] In this optional implementation, the following chain is implemented: execution monitoring agent (detects overdue risk) → upgrade signal is sent to permission management agent (determines who needs to be notified) → user interaction agent (delivers through appropriate channel).

[0052] The execution monitoring agent generates an overdue risk signal as a pre-input: It receives the task execution scheduling scheme output by the collaborative scheduling agent in the task flow protocol, using the node time limits, responsible parties, and sub-task dependencies within the scheme as the full-cycle tracking benchmark. The execution monitoring agent collects real-time data on the current task progress, node completion status, and historical response data of responsible parties, dynamically calculating the task overdue probability. When the overdue probability exceeds a preset escalation threshold, the task is formally determined to have an overdue risk. A standardized overdue risk signal is generated (containing core fields such as unique task ID, risk level, remaining time limit, affiliated department, and directly responsible person), and sent to the permission management agent through a standardized inter-agent collaboration interface, triggering an early warning escalation process.

[0053] The access control agent determines the scope of notification recipients, receiving overdue risk signals from the monitoring agent as pre-input. It invokes pre-configured hierarchical access rules from the task routing phase, combining the task's department and risk level as core dimensions to match the compliant scope of early warning notification recipients, strictly limiting the personnel authorized to receive the task's early warning information to prevent unauthorized access to sensitive task information. Output: The output includes a list of confirmed notification recipients (including personnel identification, hierarchical level, and access priority), along with the core early warning content, sent as an execution command to the user interaction agent.

[0054] The user interaction intelligent agent completes multi-channel targeted delivery, receiving a list of notification recipients and alert content from the access management intelligent agent. Based on the alert risk level and the preset reach preferences of each notification recipient, it automatically matches the optimal combination of notification channels to achieve differentiated targeted delivery. The alert information is accurately delivered to the corresponding users through the matched channels. After the push is completed, the delivery status and read receipt are sent back to the execution monitoring intelligent agent, forming a complete risk prevention and control closed loop.

[0055] Step 103: In the execution monitoring stage, the execution monitoring agent tracks the execution status of structured tasks in real time, and dynamically determines the timing and method of reminders based on the characteristics of the task's execution status, and pushes task reminders to the corresponding responsible persons.

[0056] In this step, the monitoring agent tracks task status in real time and proactively warns of risks. An independent status ledger is established for each task, collecting and updating its execution status (e.g., not started, in progress, completed, overdue) in real time. Unlike traditional methods that only track the final deadline, this tracking covers all sub-task nodes, including all stages of sequential tasks and branches of parallel tasks, achieving full-process status visualization of the task. Status data also serves as the foundational input for reminder decisions, ensuring that reminder actions accurately match the actual task progress and avoiding invalid reminders that are out of sync with the task's progress.

[0057] This embodiment implements the following intelligent agent for extracting link information → intelligent agent for task routing → intelligent agent for collaborative scheduling (which needs to be decomposed if there are dependencies) → intelligent agent for execution monitoring (starting tracking). This is the main line from the "birth" of a task to its "being tracked".

[0058] As an optional implementation of this embodiment, the execution monitoring agent dynamically determines the timing and method of reminders based on the characteristics of the task's execution state, including: modeling the task reminder decision as a Markov decision process, using a feature vector consisting of the remaining time before the deadline, task priority, task type, and the historical response rate of the responsible person as the state space; using an action set consisting of immediate reminder, delayed reminder, escalation notification, and no reminder as the action space; predicting the task overdue probability based on the current state space using a preset classification model; calculating the reward value using the task overdue probability, the number of reminders sent, and the user satisfaction score as inputs; iteratively learning the optimal reminder strategy using the Q-learning algorithm; and outputting the optimal reminder action in the corresponding state.

[0059] In this optional implementation, the key design considerations for the monitoring agent are real-time tracking of task status, proactive risk warnings, and intelligent reminders. Its core innovation lies in dynamically determining the timing and method of reminders based on reinforcement learning strategies—different responsible parties have varying tolerances for reminders, and fixed-frequency reminders are either too disruptive or too late. A "reminder strategy model" is maintained, learning the optimal reminder timing for each responsible party based on historical data, and implementing risk-weighted strategies for high-risk clauses.

[0060] Expediting decisions are a typical sequential decision problem. The risk of a task being overdue changes dynamically over time, and the current expediting action directly affects the subsequent execution status and decisions of the task. Moreover, the task state at the next moment is only related to the current state and the current action. Therefore, modeling expediting decisions as a Markov Decision Process (MDP) can transform the experience-based management problem of when and how to expedite tasks into a computable and iteratively optimizeable mathematical problem by defining the three elements of state, action, and reward. This provides a framework for the algorithm's autonomous learning.

[0061] The state space selects four features: remaining time before deadline, task priority, task type, and the responsible person's historical response rate. All features are derived from structured task data and historical execution data already output from previous stages, requiring no additional data collection. The remaining time before deadline directly determines the urgency of the task and is a core time-dimensional indicator affecting overdue risk. Task priority corresponds to the importance of the task; higher priority means greater business impact from overdue deadlines. Regarding task type, different types of tasks have different execution complexities and process lengths, resulting in different evolutionary patterns of overdue risk. The responsible person's historical response rate reflects their execution habits and response speed; a lower historical response rate indicates a higher probability of task overdue. These four features together constitute a complete quantitative characterization of the current task's overdue risk, providing an objective basis for expediting decisions.

[0062] The action space is configured with four action categories: immediate reminder, delayed reminder, escalation notification, and no reminder. This covers a complete control gradient from no intervention to strong intervention, matching the actual management needs of different risk levels. For example, no reminder is suitable for scenarios with extremely low overdue risk and responsible persons with proactive execution habits, avoiding unnecessary disturbances; delayed reminder is suitable for scenarios with ample time remaining before the deadline and low risk, delaying the reminder trigger to balance the timeliness of control and the degree of disturbance; immediate reminder is suitable for scenarios with significantly increased overdue risk, instantly pushing task reminders to urge responsible persons to advance the task; and escalation notification is suitable for scenarios with extremely high overdue risk and unresponsive responsible persons, synchronizing the risk with superiors and escalating the control intensity.

[0063] The pre-defined classification model is an intermediate computational module connecting the state space and the reward function, and it is pre-trained based on historical task data. After inputting the current state feature vector, the model outputs a task overdue probability value in the range of 0-1, transforming multi-dimensional discrete state features into a unified and quantifiable risk value, providing a standardized risk metric for the subsequent calculation of the reward function.

[0064] The reward function aims to reduce delinquency rates, control the frequency of reminders, and ensure user experience. It corresponds to three input parameters, and the reward formula is as follows: , The probability of a task being overdue corresponds to the overdue risk management target. The higher the probability, the greater the penalty, guiding the algorithm to prioritize reducing the possibility of task overdue. The number of reminders sent corresponds to the target for controlling reminder disturbances. The more reminders sent, the greater the penalty. The algorithm is guided to reduce invalid and redundant reminders. User satisfaction ratings correspond to user experience assurance goals; higher satisfaction results in greater rewards, guiding the algorithm to balance control intensity and collaborative experience. The weighting of these three factors (α, β, and γ) can be adjusted to suit different organizational management styles. The final reward value comprehensively reflects the overall effectiveness of follow-up actions.

[0065] Q-learning is a value-iterative reinforcement learning algorithm. Its core principle is to continuously refine the strategy through task feedback, gradually learning the most rewarding prompting action under different states. The specific execution logic consists of three steps: The basic definitions of Q-value and Q-table are as follows: A Q-value table is maintained to store the Q-values ​​corresponding to all combinations of states and actions. The Q-value is the estimated long-term cumulative reward that can be obtained after performing a certain reminder action in a specific task state. The higher the Q-value, the better the overall management effect of the action in the corresponding state. After each complete reminder action and the end of the task cycle, the actual reward value corresponding to this action is calculated based on the actual overdue result, the cumulative number of reminders, and user satisfaction feedback. Then, the Q-value of the corresponding state-action is adjusted in reverse using the time-series difference update formula. Simply put, if the actual effect of this reminder action is good (on-time completion, few reminders, high satisfaction), the Q-value of that action in the corresponding state is increased, and this action is more likely to be chosen when encountering similar states in the future; if the effect is poor, the corresponding Q-value is decreased, and the selection of this action is reduced in the future. When a reminder decision needs to be made in real time, the state vector of the current task is read, the Q-values ​​of the four actions in that state are queried from the Q-table, and the action with the highest Q-value is selected as the optimal reminder action output for the current state, which is the reminder timing and method with the highest expected overall benefit.

[0066] As the sample size of tasks running in the system continues to increase, the Q table will gradually adapt to the organization's work rhythm and the execution habits of different responsible persons after a large number of real-world scenarios. The accuracy of reminders will improve autonomously and continuously without the need for manual adjustment of reminder rules.

[0067] Specifically, when the information extraction agent identifies high-risk clauses such as "overdue is considered no objection," the risk weight is increased by 1.5 times, causing the overall reminder threshold to be moved forward, forming a "strengthened reminder strategy." This design stems from a pain point in actual management: such clauses are common in notices soliciting opinions on system revisions, and the "hidden loss" of losing the right to speak due to departments being too busy to provide feedback in manual management is almost impossible to quantify.

[0068] The following is an example of a decision-making process for Q-learning follow-up strategies: flowchart TD START [Enter New Task] --> EXTRACT [Information Extraction Agent Parsing Task] EXTRACT -->CHECK {High-risk terms identified?} CHECK --> |Yes| RISK [Risk weight α × 1.5] CHECK --> | No | NORMAL [Standard Weight α] RISK --> STATE [Construct state vector S] NORMAL --> STATE STATE --> QTABLE [Optimal action for Q table queries] QTABLE --> EPSILON {ε-greedy selection} EPSILON -->|Explore| RANDOM [Randomly select an action] EPSILON --> | Use | BEST [Select the action with the highest Q value] RANDOM --> ACTION BEST -->ACTION ACTION [Execute Action A: Remind / Delay / Escalate / Do Not Remind] ACTION --> REWARD [Calculate Reward R] REWARD --> UPDATE [Update Q value Q←Q+αδ] UPDATE --> NEXT {Task complete?} NEXT -->|NO| STATE NEXT --> | Yes | END [Task Archive].

[0069] As an optional implementation of this embodiment, during the task element extraction process, it is identified whether the email text contains preset high-risk clauses; if high-risk clauses are identified, the reward weight corresponding to the task overdue probability in the Markov decision process is increased to a preset multiple, and the task is subject to an enhanced reminder strategy by moving the reminder trigger threshold forward.

[0070] In this optional implementation method, some tasks in daily collaboration contain high-risk clauses such as "failure to provide feedback by the deadline is considered as no opinion" and "failure to submit by the deadline is considered as abandonment of qualification". Once such tasks are overdue, they will cause hidden losses such as loss of speaking rights and failure to apply for qualification. Moreover, such losses are difficult to quantify in manual management and are easily overlooked due to being busy.

[0071] Traditional fixed-cycle follow-up applies a uniform pace to all tasks, making it impossible to escalate control measures for high-risk tasks. This mechanism embeds risk identification into the analysis process, automatically matching differentiated follow-up strategies to ensure that the control measures precisely correspond to the task's risk level, proactively preventing hidden losses.

[0072] The identification process is embedded in the task element extraction stage before the action is performed, and is completed simultaneously with the six-factor extraction. No additional steps are required. While parsing the email text and extracting elements such as the task subject and deadline, the LLM parser identifies pre-defined high-risk clause types through semantic matching and marks the corresponding tasks as high-risk, simultaneously transmitting this information to the subsequent execution monitoring agent. This identification relies on the semantic understanding capabilities of LLM, is not limited to fixed keywords, and can adapt to similar risk clauses with different expressions, achieving a higher accuracy rate than pure keyword matching.

[0073] When a task is marked as high-risk, the weight coefficient α corresponding to the task's overdue probability in the reward function of the Markov decision process is increased to a preset multiple, such as 1.5 times. The transmission logic is that in the reward function, the overdue probability is a negative penalty term. After the weight α is increased, the penalty corresponding to the same overdue probability will be increased accordingly. In ordinary tasks, the state of "no reminder / delayed reminder" which originally had higher benefits will have a significantly increased proportion of overdue penalty after weighting, and the overall benefit of "immediate reminder / early reminder" will surpass that of other actions.

[0074] This change will be directly reflected in the Q-value calculation, making the algorithm more inclined to choose earlier and stronger reminder actions when there is the same remaining time and the same priority. Ultimately, this will result in the reminder trigger threshold being moved forward as a whole, without the need for manual setting of reminder rules.

[0075] The weighting adjustment ultimately translates into three levels of enhanced control: first, the initial reminder time is moved forward, triggering the first reminder earlier than for ordinary tasks of the same level; second, the follow-up interval is shortened, with a higher frequency of follow-ups within the same period to continuously strengthen risk warnings; and third, the upgrade notification threshold is moved forward, triggering the action of synchronizing risks with superiors earlier, thus upgrading the control level. Through these three layers of differentiated control, the probability of overdue high-risk tasks is proactively reduced, avoiding hidden business losses that are easily overlooked in manual management.

[0076] As an optional implementation method in this embodiment, the execution monitoring agent counts the trigger frequency of similar tasks during the task tracking process. When the preset trigger conditions are met, the knowledge management agent is notified, and the knowledge management agent extracts reusable experience from the task execution process data and stores it in the knowledge base to form organizational assets.

[0077] In this optional implementation, the execution monitoring agent uses the generated structured task data as a basis to identify similar tasks and count the trigger frequency by accurately matching the core structured fields and combining semantic similarity verification. When repeated triggering of similar tasks is detected, the knowledge management agent is notified. After the knowledge management agent verifies that the frequency reaches the corresponding threshold, a three-level accumulation mechanism is used to complete the experience archiving.

[0078] Task execution data, including basic task information, reminder records, process consultations and Q&A, and feedback, all originate from the flow data of preceding intelligent agents, requiring no additional collection. The data accumulation process is divided into three levels: FAQ accumulation, task templating, and association rule accumulation. The trigger thresholds, processing methods, and reuse value of each level are progressively upgraded, ultimately being stored together in a knowledge base to form a task management asset that can be accessed by the entire organization. The accumulated knowledge base content then supports the preceding task scheduling and execution monitoring processes, forming an organizational learning closed loop of "execution—accumulation—reuse—optimization."

[0079] The key design principle of the knowledge management agent is to transform "experienced knowledge" gained during task execution into reusable organizational assets. The specific mechanism includes three levels of accumulation: automatic FAQ accumulation (automatically generating standard answers when the same question is asked more than three times), task templating (automatically extracting common structures after repeated occurrences of similar tasks), and association recommendation (analyzing related matters from historical data based on association rule algorithms). These three levels of accumulation, progressing from "answering questions" to "providing templates" to "predicting associations," create an "experience accumulation" mechanism. This is a key design feature that distinguishes TMS-Agent from traditional task management systems. In the traditional model, once a task is completed, the process ends, and the execution experience dissipates. This predicament is consistent with the problem of "tacit knowledge being difficult to transfer" revealed by Polanyi—a large amount of experience gained during task execution exists at the individual level and is difficult to make explicit and accumulate into organizational assets. Argyris's two-loop learning theory points out that the key to organizational learning lies not only in correcting behavior (single loop) but also in reflecting on and correcting the assumptions behind the behavior (double loop). The experience accumulation mechanism proposed in this embodiment unifies single-loop learning (task execution feedback) and double-loop learning (experience rule accumulation) at the computational level.

[0080] For example, the experience accumulation mechanism includes three levels: Level 1: FAQ Accumulation. If the same question is asked more than 3 times, a standard answer is automatically generated and added to the knowledge base. This is the lightest level of accumulation—when a new task is released, the system recommends relevant historical FAQs based on vector retrieval, directly answering common questions and reducing redundant communication.

[0081] The second level: Task templateization. When similar tasks are repeated, common structures are automatically extracted to generate templates for subsequent reuse. For example, after a "project application" task is executed multiple times, a standard "four-stage sequential" process template and a standard "20-15-5 day" follow-up schedule are automatically extracted.

[0082] The third level: Association rule accumulation. High-frequency association patterns between tasks discovered through the Apriori algorithm form empirical rules such as "doing task A usually requires doing task B as well." When a user creates a new task, related items are proactively recommended.

[0083] The three levels progress from simple to complex, from "answering questions" to "providing templates" to "predicting connections," gradually transforming personal experience into organizational assets.

[0084] The above implementation achieves the following chain: execution monitoring agent (detecting that similar tasks are repeatedly asked) → notification to knowledge management agent (whether the FAQ threshold has been reached) → knowledge management agent generates a template and stores it in the knowledge base. This is the technical implementation path of "experience accumulation".

[0085] As an optional implementation of this embodiment, extracting reusable experience from task execution process data through a knowledge management intelligent agent includes: statistically analyzing business consultation questions generated during the execution of similar tasks; automatically generating standardized answers and storing them in the knowledge base when the number of times the same business consultation question is repeatedly asked reaches a preset threshold; automatically extracting the common process structure and standard execution rhythm of the same type of task when the number of times the same type of task is repeatedly asked reaches a preset threshold, generating standardized task templates and storing them in the knowledge base; and extracting the temporal dependencies between tasks based on historical task data using an association rule mining algorithm, generating task association recommendation rules and storing them in the knowledge base.

[0086] In this optional implementation, the three levels of FAQ accumulation can be implemented in various ways, and no limitation is made here. For example, the data source for the automatic FAQ accumulation logic is the consultation and Q&A records submitted by the responsible persons and collaborators during the execution of each task; the judgment rule is that the agent performs semantic clustering on the consultation text, identifies similar questions with the same meaning, and counts the frequency of repeated inquiries; the trigger condition is that the number of inquiries for the same semantic question is greater than or equal to a preset threshold (e.g., the threshold is 3 times); the processing action is to integrate the official replies from previous times, extract unified standard Q&A items, and store them in the knowledge base; the reuse value refers to the fact that when a new similar task is created, the system automatically pushes FAQs, reducing the cost of repeated Q&A communication.

[0087] The standardized task template accumulation logic includes extracting comprehensive execution data from the execution process ledger, including the subtask breakdown sequence, subtask dependencies, start and end times of each node, reminder trigger nodes, and overdue records for all historical tasks of this type. A multi-sequence alignment algorithm is used to globally align the subtask execution sequences of all historical tasks, extracting common subtask sequences and fixed dependencies to form a standard process framework. For example, for project application tasks, the aligned process follows a four-stage sequential standard process: notification issuance → material writing → departmental preliminary review → institute-level review → external submission. The historical median execution time and optimal reminder trigger time for each standard node are statistically analyzed to fit the standard execution rhythm and reminder strategy for this type of task. For example, the standard reminder rhythm for project application tasks is: first reminder 20 days before the deadline, second reminder 15 days later, and escalation warning 5 days later. The standard process framework, dependencies, execution rhythm, and reminder strategy are integrated into a structured task execution template, including template ID, applicable task type, subtask configuration, scheduling rules, and reminder rule fields, and stored in the knowledge base template library.

[0088] When the collaborative scheduling agent receives a new task, it matches the standard template corresponding to the task type and directly calls the template to complete the automatic decomposition of subtasks, configuration of dependencies, and scheduling of time nodes; the execution monitoring agent synchronously applies the reminder rules built into the template to perform task tracking without having to configure the scheduling strategy from scratch.

[0089] The task sequence association rule accumulation logic is executed by the monitoring agent. At a preset period (such as monthly) or when the total number of historical tasks reaches a preset mining threshold, the association rule mining trigger signal is sent to the knowledge management agent, and the read permission of the full historical task ledger is opened simultaneously.

[0090] The Apriori algorithm is used for iterative mining of the transaction dataset: First, the entire dataset is scanned to generate one-item frequent sets that meet the minimum support threshold. Second, two-item frequent sets are generated based on the one-item frequent sets, and these two-item frequent sets that meet the support threshold are selected. This process is repeated iteratively until all k-item frequent sets that meet the preset support threshold are mined. Based on the mined frequent itemsets, all possible association rules are generated, and the confidence of each rule is calculated. Rules that simultaneously meet the preset minimum support and minimum confidence thresholds are selected; for example, the association rule "Project Application → Funding Budget Approval" with 22% support and 87% confidence is selected. The selected rules undergo business rationality verification. Invalid rules with accidental associations are removed, and structured association rules containing fields such as "prerequisite task, associated task, support, confidence, and applicable scenario" are generated and stored in the knowledge base rule library.

[0091] When a user creates a new task, the system reads the type and attributes of the current task, matches it against association rules in the knowledge base that use previous tasks as matching items, and proactively recommends corresponding related tasks to the user. The user can initiate the corresponding task with one click, avoiding omissions in the business process.

[0092] This embodiment features parallel input via both native direct sending channels and email gateway channels, covering both standardized data entry and traditional email collaboration scenarios. Users do not need to change their existing work habits, significantly reducing the cost of system deployment and promotion. The email gateway employs a rule engine + LLM hierarchical scheduling. Formatted notifications use the low-computing-power rule engine, while free-form emails use LLM. The two serve as backups and cross-validations for each other, controlling computing power consumption and reducing response latency while ensuring the completeness of parsing all types of emails. It uniformly outputs standardized task data with six elements, and all backend modules share a single data standard, simplifying the system architecture and facilitating global task statistics and performance analysis.

[0093] The task routing agent automatically matches the responsible parties and synchronously binds hierarchical permissions, clarifying the boundaries of task operation and information visibility, and avoiding shirking of responsibility and cross-departmental information leakage; the collaborative scheduling agent automatically breaks down composite tasks, identifies serial / parallel dependencies of subtasks and generates dependency graphs, replacing manual scheduling; repetitive tasks automatically solidify standard templates, which can be reused with one click for subsequent similar tasks, unifying the organization and execution standards, and eliminating omissions and process differences caused by manual breakdown.

[0094] The reminder decision-making process is modeled as a Markov decision process, autonomously learning the optimal reminder timing and method based on task characteristics and the responsible person's historical behavior. This replaces the "one-size-fits-all" fixed reminders, reducing unnecessary disturbances and lowering the overall task overdue rate. It automatically identifies high-risk clauses in emails, dynamically increases the overdue risk weight, and moves the reminder trigger threshold forward to strengthen control and proactively avoid hidden business losses caused by overdue payments. The Q-learning algorithm continuously iterates the reminder strategy based on real business feedback, eliminating the need for repeated manual adjustments to reminder rules and resulting in lower long-term maintenance costs.

[0095] FAQ auto-conversion enables the automatic generation of standard Q&As for repetitive business inquiries, reducing redundant communication across different roles; task template consolidation enables the standardization of high-frequency similar tasks, unifying execution rhythm and improving task scheduling efficiency; association rule mining automatically extracts cross-task temporal relationships based on the full historical tasks, and proactively pushes corresponding work reminders for new tasks to avoid omissions in business processes; the three layers of knowledge are uniformly stored in the knowledge base, forming long-term organizational assets that are unaffected by personnel changes.

[0096] The sub-process data within a single task output by the collaborative scheduling serves as a standardized data source for the association rule mining algorithm; the timing sequence of the inner single task sub-steps and the timing sequence of the outer cross-task global matching work together to fully cover the timing constraints of the entire task lifecycle; the cross-task rules of the knowledge base provide reverse support for task scheduling, forming a complete business closed loop of "parsing-allocation-monitoring-accumulation-reuse".

[0097] It silently stores email parsing results and automatically generates labeled samples based on users' daily correction operations. When the samples accumulate to a threshold, it automatically fine-tunes the LLM parsing model without the need for additional dedicated labeling personnel. It continuously improves the accuracy of email element extraction as business operations progress and adapts to the enterprise's specific business context.

[0098] This embodiment overcomes the following four shortcomings: First, there is a lack of systematic design in task management. There is no complete management ecosystem from task release, tracking, and feedback to knowledge accumulation; and no closed-loop management extending to task execution.

[0099] Second, there is insufficient research on role collaboration mechanisms in management scenarios.

[0100] Third, there is insufficient attention paid to knowledge accumulation during task execution. Existing OA systems and project management software (such as Jira and Asana) all take "task completion" as the end point, failing to transform the questions, experiences, and lessons learned during the execution process into reusable organizational assets.

[0101] Fourth, there is a lack of empirical verification based on real email data. Most studies remain at the theoretical framework or simulation level.

[0102] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0103] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A multi-agent collaborative task processing method, characterized in that, include: Receive structured tasks through a pre-configured native direct transmission channel; The system receives email text to be processed through a pre-configured email gateway channel, and extracts task elements from the email text based on the information extraction agent to output structured task data containing task subject, source department, responsible person, deadline, priority, and task type. The task routing agent matches the responsible parties corresponding to the structured task data; the collaborative scheduling agent identifies the sub-task dependencies within the task and breaks them down, generating parallel or serial task execution scheduling schemes, and after the task scheduling is completed, it flows to the execution monitoring stage; During the execution monitoring phase, the execution monitoring agent tracks the execution status of structured tasks in real time and dynamically determines the timing and method of reminders based on the characteristics of the task's execution status, pushing task reminders to the corresponding responsible persons.

2. The multi-agent collaborative task processing method according to claim 1, characterized in that, The method also includes: The execution monitoring agent counts the trigger frequency of similar tasks during task tracking. When the preset trigger conditions are met, it notifies the knowledge management agent, which then extracts reusable experience from the task execution process data and stores it in the knowledge base to form organizational assets.

3. The multi-agent collaborative task processing method according to claim 1, characterized in that, The task element extraction process for the email text to be processed includes two parsing mechanisms: a rule engine and an LLM parser. The scheduling rule engine identifies the format regularity of the email text to be processed. For well-formatted notification emails, the engine extracts elements through keyword matching. For emails with flexible formats, the scheduling LLM parser extracts elements through semantic understanding; The rule engine and the LLM parser serve as backups for each other, and the extraction results are cross-validated.

4. The multi-agent cooperative task processing method according to claim 1, characterized in that, The method also includes human-machine collaborative correction: The structured task data parsed by the email gateway channel is silently stored, and a parsing result verification notification is pushed to the terminal of the corresponding responsible person. If a correction instruction is received from the responsible party's terminal regarding the parsing results, the original email text, parsing results, and user correction results are recorded simultaneously as annotation samples. When the cumulative number of labeled samples reaches a preset threshold, the LLM parser is triggered to adapt and fine-tune to update the LLM parser parameters.

5. The multi-agent cooperative task processing method according to claim 1, characterized in that, Before the matching of responsible parties is completed and the task execution scheduling plan is generated, the process also includes a permission pre-configuration sub-step: By using an intelligent agent with access control based on a role-based and hierarchical access system, the visibility range of corresponding task information is configured for each matched responsible entity; Set up feedback content isolation rules across responsible entities; and pre-configure the notification object hierarchy matching rules corresponding to task overdue warnings.

6. The multi-agent cooperative task processing method according to claim 1, characterized in that, The dynamic determination of the timing and method of reminders based on the characteristics of the execution status of the monitoring agent's tasks includes: The task expediting decision is modeled as a Markov decision process, with the feature vector consisting of the remaining time before the deadline, task priority, task type, and the historical response rate of the responsible person as the state space. The action space is defined by a set of actions consisting of immediate reminders, delayed reminders, escalation notifications, and no reminders. The probability of task overdue is predicted based on the current state space using a pre-defined classification model. The reward value is calculated using the task overdue probability, the number of reminders sent, and the user satisfaction rating as inputs; the optimal reminder strategy is learned iteratively through the Q-learning algorithm, and the optimal reminder action is output for the corresponding state.

7. The multi-agent cooperative task processing method according to claim 6, characterized in that, During the task element extraction process, it is identified whether the email text contains preset high-risk clauses; If a high-risk clause is identified, the reward weight corresponding to the probability of task overdue during the Markov decision-making process will be increased to a preset multiple, and the reminder trigger threshold will be moved forward to implement an enhanced reminder strategy for the task.

8. The multi-agent cooperative task processing method according to claim 2, characterized in that, Extracting reusable experience from task execution data through knowledge management agents includes: The system analyzes business inquiries generated during the execution of similar tasks. When the number of times the same business inquiry is asked repeatedly reaches a preset threshold, a standardized answer is automatically generated and stored in the knowledge base. When the number of times the same type of task is repeated reaches a preset threshold, the common process structure and standard execution rhythm of the task are automatically extracted to generate a standardized task template and store it in the knowledge base. Based on historical task data, the temporal dependencies between tasks are extracted using association rule mining algorithms to generate task association recommendation rules and store them in a knowledge base.

9. The multi-agent cooperative task processing method according to claim 1, characterized in that, The subtask dependencies within a structured task can be identified through a collaborative scheduling agent, including: Automatically break down multi-step complex tasks and generate subtask dependency graphs; For recurring task combinations, they are automatically fixed into standardized task execution templates for use as a reference for scheduling subsequent similar tasks.

10. The multi-agent cooperative task processing method according to claim 5, characterized in that, The method also includes: When the monitoring agent tracks the execution status of structured tasks in real time, it generates an escalation warning signal and sends it to the permission management agent when it detects a risk of task overdue. After the permission management agent determines the scope of the warning notification object based on the preset permission rules, it triggers the user interaction agent to match the corresponding notification channel and completes the targeted delivery of the warning information.

Images