Method for automatically coordinating, organizing and solving tasks by multiple Agents based on large model

Abstract

The invention relates to the technical field of large language models, and discloses a multi-Agent automatic coordination organization task solving method based on a large model, and the method comprises the steps: 1, receiving and collecting heterogeneous abnormal events from at least an order system, a warehousing system and a transportation system, uniformly converting the heterogeneous abnormal events into standardized abnormal event descriptions containing abnormal types, related orders, related resources and emergency degrees; and 2, calling an exception planning intelligent role, reasoning and planning the standardized exception event description by using a large model service, and dynamically generating an exception handling chain. According to the method, the technical scheme of unified collection and standardized description of heterogeneous abnormal events is adopted, the technical effect of automatically integrating cross-system information is achieved, clear abnormal task processing and automatic generation of a unified view are achieved, and the defects that manual data summarization consumes time and lacks an automatic integration mechanism are overcome.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, specifically to a method for automatic coordination and organization of multiple agents based on large models to solve tasks. Background Technology

[0002] The rapid development of e-commerce and modern logistics has dramatically increased the complexity of supply chains and order fulfillment. The entire process from receiving a customer order to final delivery involves numerous independent business systems, including order management, warehouse management, transportation management, customer service, and settlement systems. In this highly complex automated process, any stage can experience anomalies, posing a serious challenge to a company's fulfillment efficiency, operating costs, and customer satisfaction.

[0003] However, current exception handling and task coordination mechanisms face multiple bottlenecks:

[0004] In existing technologies, abnormal order fulfillment information is scattered across independent order systems, warehousing systems, transportation systems, customer service systems, and settlement systems. Operators need to spend a lot of time switching between systems to query and manually summarize data, resulting in a lack of technical mechanisms that can automatically integrate information from all systems and generate a clear and unified view of abnormal task handling.

[0005] Traditional exception handling systems rely on fixed programming rules or scripts for task assignment and early warning, which makes it difficult to quickly adapt to the differentiated requirements of different business lines, customers or carriers. Once business strategies change, rule adjustments lag behind actual needs, resulting in a lack of ability to make intelligent judgments and respond flexibly based on real-time changes in business semantics and complex exception types.

[0006] In complex exception handling processes, the lack of the ability to automatically break down exceptions into clear sub-tasks at the underlying level and automatically determine the responsible parties for these sub-tasks leads to exception handling being highly dependent on human experience for allocation. This can easily result in chaotic situations where multiple parties follow up repeatedly or no one follows up at all, seriously affecting overall processing efficiency and service quality.

[0007] Existing end-to-end automation solutions often require deep and complex integration of various heterogeneous systems within an enterprise, and require the configuration of costly access control and auditing structures, resulting in high development and maintenance costs for small and medium-sized enterprises with dispersed systems and limited IT resources.

[0008] To address the aforementioned issues, this invention proposes a method for automated coordination and organization of multi-agent tasks based on a large-scale model. This method achieves dynamic processing of heterogeneous anomalies and multi-agent collaboration through intelligent planning using a large-scale model, virtual simulation optimization, and task ledger tracking, overcoming the shortcomings of traditional processes such as lack of intelligent flexibility, information fragmentation, and high integration costs. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method for automatically coordinating and organizing multi-agent tasks based on a large model, thereby resolving the problems mentioned in the background section.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for automatic coordination and organization of multi-agent tasks based on a large model, comprising:

[0011] Step 1: Receive and collect heterogeneous abnormal events from at least the order system, warehousing system, and transportation system, and transform the heterogeneous abnormal events into standardized abnormal event descriptions that include the abnormality type, related orders, related resources, and urgency level.

[0012] Step 2: Invoke the intelligent role for anomaly planning, and use the large model service to reason and plan the standardized anomaly event description, and dynamically generate an anomaly handling chain;

[0013] Step 3: In the virtual warehouse environment, simulate the exception handling chain before execution, evaluate the potential impact of the exception handling chain on inventory, transportation resources and order status after execution, and iteratively optimize and correct the exception handling chain based on the simulation results;

[0014] Step 4: Write the optimized and corrected exception handling chain as a task snapshot into a unified task ledger. The task ledger records the initial state snapshot of the task, the current processing step, the corresponding intelligent role executing the task, and the state snapshot after execution.

[0015] Step 5: Based on the current processing step determined in the task ledger, the task is distributed to the corresponding intelligent role for execution. The intelligent role drives the external system to perform the corresponding operation through a standardized execution interface or status module, and at the same time updates the status snapshot in the task ledger.

[0016] Step 6: Monitor the execution feedback and current status snapshot of each intelligent role in the task ledger in real time. If the task execution result is found to be inconsistent with the expected result, call the anomaly planning intelligent role. Based on the historical snapshot and the latest feedback in the task ledger, re-utilize the large model to diagnose the anomaly handling chain, roll back to the historical state, and re-plan the anomaly handling chain.

[0017] Preferably, in step two, dynamically generating the exception handling chain specifically includes:

[0018] The standardized abnormal event description is used as input, and structured reasoning is performed using a large model service. Based on the reasoning results, several processing steps are generated. The processing steps are structured to define the operation content, required resources, expected results, and corresponding intelligent roles. The abnormal event handling chain is organized and stored using a flowchart or a directed acyclic graph data structure.

[0019] Preferably, in step three, assessing the potential impact of the exception handling chain on inventory, transportation resources, and order status specifically includes:

[0020] Calculate the performance risk index after the execution of the exception handling chain. The performance risk index includes at least one of the following: expected order delay time, inventory shortage, or transportation cost change rate. The iterative optimization and correction of the exception handling chain based on simulation results is based on the judgment and adjustment according to the preset threshold of the performance risk index.

[0021] Preferably, in step four, the task ledger records the initial state snapshot of the task, specifically including: recording the order status, associated inventory quantity and location, carrier information, estimated delivery time, and business rule version number at the start of the task.

[0022] Preferably, step two further includes:

[0023] Sub-step The standardized abnormal event description As input, through a pre-trained word embedding model Transform into high-dimensional event feature vectors ; Calculate the event feature vector With the set of vectors of known historical anomaly types Cosine similarity between ;

[0024] If cosine similarity Exceeding the preset similarity threshold Directly invoke the corresponding historical exception handling plan. The historical anomaly handling plan Includes a predefined set of processing steps ;

[0025] If cosine similarity The similarity threshold was not exceeded. The standardized abnormal event description Input to the largest model service Perform structured reasoning ;

[0026] ,

[0027] in, For the first known exception type vector in the set Types, For structured reasoning functionality in large models;

[0028] Sub-step The structured reasoning Through the mind chain prompting project, standardized abnormal event descriptions are made available. and business strategy knowledge base As contextual input, the output is a set of candidate processing steps. The set of processing steps Includes several optional processing steps The candidate processing steps Define the operation content, required resources, expected results, and corresponding intelligent roles in a structured manner. ;

[0029] Sub-step Processing steps for candidates Perform timing constraint verification Ensure that resource dependencies and time sequence between steps meet the requirements of the business strategy knowledge base. Constraints in ; Calculate the candidate processing steps Execution cost The execution cost It is caused by resource consumption and expected delay Obtained by weighted calculation;

[0030] ,

[0031] in, The execution cost of the alternative processing steps, This is a weighting coefficient for resource consumption. Resource consumption for the pending processing steps. The weighting coefficient is the expected delay. The expected delay for the pending processing steps;

[0032] Sub-step Based on the execution cost Using a heuristic search algorithm Alternatively, dynamic programming methods can be used to verify compliance with time constraints. Under the premise of the set of candidate processing steps Select the optimal processing step subset The optimal processing steps subset Total execution cost minimize;

[0033] Subset of the optimal processing steps The data structure, organized into a flowchart or directed acyclic graph according to the execution logic, serves as the exception handling chain. Perform dynamic output.

[0034] Preferably, step three further includes:

[0035] Sub-step Based on the initial state snapshot of the task ledger records Build a virtual warehouse environment Initial simulation state The initial simulation state Includes real-time inventory, transportation resources, and estimated delivery time for task-related orders;

[0036] The exception handling chain Input as a sequence to be executed into the virtual warehouse environment Simulation preparation is being carried out.

[0037] Sub-step The virtual warehouse environment Discrete event simulation model Simulate the exception handling chain The sequential execution of each operation step;

[0038] In each step After completion, calculate the performance risk indicators. The aforementioned performance risk indicators It is the expected delay time of the order. Inventory gap and rate of change in transportation costs The weighted sum is used to quantify the potential impact;

[0039] ,

[0040] in, For steps Performance risk indicators after the end of the contract period For the expected delay time of the order, This represents the inventory shortfall. For the rate of change in transportation costs,

[0041] , , The corresponding weight coefficients, and satisfying ;

[0042] Sub-step In the exception handling chain After all simulation steps are completed, the final total performance risk index is obtained. ;

[0043] Determine the total performance risk indicator Does it exceed the preset risk threshold? ;

[0044] like > The total performance risk indicator and a snapshot of the state after simulation As feedback, the information is re-entered into the anomaly planning intelligent role to iteratively optimize the anomaly handling chain. ;

[0045] like ≤ The exception handling chain This is confirmed as the final optimized and corrected exception handling chain;

[0046] ,

[0047] in, As the final overall performance risk indicator, This represents the total number of steps in the exception handling chain. As a preset risk threshold, This is a snapshot of the final state after simulation.

[0048] Preferably, step four further includes:

[0049] Sub-step The optimized and corrected exception handling chain Transform flowcharts or directed acyclic graph data structures into task snapshots of or Formatted data structure; the task snapshot Includes the exception handling chain An ordered list of all processing steps Expected results of each step and the initial state snapshot of the task A secure hash algorithm is used. Calculate the task snapshot Snapshot hash As a unique identifier;

[0050] ,

[0051] in, For task snapshots, For a secure hash algorithm, The snapshot hash value of the task snapshot;

[0052] Sub-step , take the task snapshot and the snapshot hash value Write to task ledger Initial task record The task ledger mentioned above; Employing a distributed data structure such as blockchain or hash chain, the snapshot hash value is verified. Ensure the initial record of the task Data integrity;

[0053] Sub-step In the task ledger Create a task status tracking table The task status tracking table By snapshot hash value Associate the task initial record ; Task status tracking table Current processing steps Initialize as the exception handling chain The first step; the task status tracking table The reserved field is used to record the corresponding smart role to be executed. and post-execution state snapshot .

[0054] Preferably, step five further includes:

[0055] Sub-step From the task status tracking table Get the current processing step All information All the information mentioned This includes the operation content of this step. and the corresponding intelligent role to perform. According to the corresponding intelligent role. Registry of predefined smart roles The corresponding intelligent role is located in the middle. execution address and standardized execution interface ;

[0056] Sub-step According to the execution address and standardized execution interface The operation content As instructions Distributed to the corresponding intelligent role The corresponding intelligent role Receive instructions Through the standardized execution interface or status module Calling external systems atomic operations The external system This includes one of the order system, warehousing system, or transportation system;

[0057] Sub-step The external system Perform the atomic operation Then, to the corresponding intelligent role Return execution status code and business results data The corresponding intelligent role Execution status code and business results data Encapsulate as a post-execution state snapshot ; Take a snapshot of the state after the execution Including the time to complete each step Write to the task status tracking table In addition, according to the logic of the exception handling chain, the task status tracking table is... Current processing steps Updated to the next pending step The next step to be executed The selection follows the topological sorting rules of directed acyclic graphs;

[0058] Sub-step Calculate the current task status tracking table The set of all completed steps In the exception handling chain Completion rate in the total set of steps The degree of completion Exceeding If the task is completed, it is determined that the task has been executed; otherwise, the task continues to the next pending step. ;

[0059] ,

[0060] in, For completion level, For the set of completed steps, This is an exception handling chain.

[0061] Preferably, step six further includes:

[0062] Sub-step Read the task status tracking table in real time. The latest post-execution state snapshot And obtain the expected result corresponding to this step. Define the set of key fields. The set of key fields It includes one of the following: order status code, inventory update value, and execution status code reported by external system;

[0063] Calculate the state snapshot after execution set of key fields The actual value and the expected result set of key fields Inconsistency measure between expected values ;

[0064] If inconsistency measurement Exceeding the preset inconsistency threshold The result of the task execution is determined to be inconsistent with the expected result;

[0065] ,

[0066] in, As a measure of inconsistency, The inconsistency threshold, Key fields The weighting coefficients, Actual value Compared with expected value The function for calculating the difference between them;

[0067] Sub-step If the determination results are inconsistent, a snapshot of the state after execution will be taken. and inconsistency measurement As the latest feedback Input to the abnormal planning intelligent role Perform a diagnosis; the diagnosis determines the root cause of the inconsistency and identifies the rollback point; the rollback point is selected from the task status tracking table. A snapshot of the state of the previous successful execution step. Or a snapshot of the initial state of the task. ;

[0068] Sub-step Based on the state snapshot corresponding to the rollback point Calling external systems State rollback interface The external system The state logic is rolled back to the corresponding state snapshot. Simultaneously, the task status tracking table Current processing steps Update to the rollback point; the abnormal planning intelligent role Based on the corresponding state snapshot And the latest feedback The large model is then used to re-plan the exception handling chain, generating a new exception handling chain.

[0069] Preferably, in step six, determining that the task execution result is inconsistent with the expected result specifically includes: comparing the differences in key fields between the post-execution status snapshot in the task ledger and the expected result in the exception handling chain; the key fields include at least one of the order status code, inventory update value, or execution status code fed back by the external system; the rollback to the historical state is a rollback operation based on the initial status snapshot recorded in the task ledger or the status snapshot of the previous successfully executed step.

[0070] This invention provides a method for automatically coordinating and organizing multi-agent teams to solve tasks based on a large model. It has the following beneficial effects:

[0071] 1. This invention adopts a unified collection and standardized description technology for heterogeneous abnormal events, which achieves the technical effect of automatically integrating cross-system information, realizes the automatic generation of a clear and unified view of abnormal task processing, and solves the shortcomings of manual data collection being time-consuming and lacking an automated integration mechanism.

[0072] 2. This invention adopts a large model-driven structured reasoning and dynamic generation of anomaly handling chain technology to achieve real-time understanding of business semantics and complex anomaly types, realize intelligent judgment of the processing flow and flexible adaptation to different business strategies, and solve the shortcomings of traditional fixed rule systems that have poor adaptability and are lagging behind the needs.

[0073] 3. The present invention adopts a multi-agent collaborative execution and task ledger step tracking mechanism technical solution to achieve the technical effect of automatically breaking down anomalies into clear sub-tasks and automatically determining the responsible parties, realizing efficient and orderly execution and follow-up of multi-party collaboration, and solving the shortcomings of anomaly handling that relies on manual allocation and is prone to duplication or lack of follow-up.

[0074] 4. This invention adopts a technical solution that uses intelligent roles to drive external systems through standardized interfaces or state modules, achieving low-coupling, low-cost system integration and step-by-step collaboration. It enables the construction of full-process automation without complex deep integration, solving the shortcomings of full-process automation, which has high requirements for system integration and high development and maintenance costs. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the overall architecture of a method for automatically coordinating and organizing multi-agent organizations to solve tasks based on a large model, according to the present invention.

[0076] Figure 2 This is a flowchart illustrating the dynamic generation and optimization of the exception handling chain in this invention.

[0077] Figure 3 This is a schematic diagram of the multi-agent collaborative execution and task ledger status tracking mechanism of the present invention. Detailed Implementation

[0078] To enable those skilled in the art to 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 some, but not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0079] The present invention will now be described in detail with reference to the accompanying drawings:

[0080] Example:

[0081] Please see the appendix Figure 1 -Appendix Figure 3 This invention provides a method for automatically coordinating and organizing multi-agent tasks based on a large model, comprising:

[0082] Step 1: Receive and collect heterogeneous abnormal events from at least the order system, warehousing system, and transportation system, and transform the heterogeneous abnormal events into standardized abnormal event descriptions that include the abnormal event type, related orders, related resources, and urgency level.

[0083] Step 2: Invoke the intelligent role for anomaly planning, and use the large model service to reason and plan the standardized anomaly event descriptions to dynamically generate an anomaly handling chain;

[0084] Step 3: In the virtual warehouse environment, simulate the exception handling chain before execution, evaluate the potential impact of the exception handling chain on inventory, transportation resources and order status after execution, and iteratively optimize and correct the exception handling chain based on the simulation results;

[0085] Step 4: Write the optimized and corrected exception handling chain as a task snapshot into a unified task ledger. The task ledger records the initial state snapshot of the task, the current processing step, the corresponding intelligent role to be executed, and the state snapshot after execution.

[0086] Step 5: Based on the current processing steps determined in the task ledger, the task is distributed to the corresponding intelligent role for execution. The intelligent role drives the external system to perform the corresponding operation through a standardized execution interface or status module, while updating the status snapshot in the task ledger.

[0087] Step 6: Monitor the execution feedback and current status snapshot of each intelligent role in the task ledger in real time. If the task execution result is found to be inconsistent with the expected result, call the anomaly planning intelligent role. Based on the historical snapshot and the latest feedback in the task ledger, re-utilize the large model to diagnose the anomaly handling chain, roll back to the historical state, and re-plan the anomaly handling chain.

[0088] Step two further includes:

[0089] Sub-step Standardize the description of abnormal events As input, through a pre-trained word embedding model Transform into high-dimensional event feature vectors ; Calculate the event feature vector With the set of vectors of known historical anomaly types Cosine similarity between ;

[0090] If cosine similarity Exceeding the preset similarity threshold Directly invoke the corresponding historical exception handling plan. Historical anomaly handling plan Includes a predefined set of processing steps ;

[0091] If cosine similarity The similarity threshold was not exceeded. Standardize the description of abnormal events Input to the largest model service Perform structured reasoning ;

[0092] ,

[0093] in, For the first known exception type vector in the set Types, For structured reasoning functionality in large models;

[0094] Sub-step Structured reasoning Through the mind chain prompting project, standardized abnormal event descriptions are made available. and business strategy knowledge base As contextual input, the output is a set of candidate processing steps. ; Set of processing steps Includes several optional processing steps Pending processing steps Define the operation content, required resources, expected results, and corresponding intelligent roles in a structured manner. ;

[0095] Sub-step Processing steps for candidates Perform timing constraint verification Ensure that resource dependencies and time sequence between steps meet the requirements of the business strategy knowledge base. Constraints in ; Calculate the candidate processing steps Execution cost Execution cost It is caused by resource consumption and expected delay Obtained by weighted calculation;

[0096] ,

[0097] in, The execution cost of the alternative processing steps, This is a weighting coefficient for resource consumption. Resource consumption for the pending processing steps. The weighting coefficient is the expected delay. The expected delay for the pending processing steps;

[0098] Sub-step Based on execution cost Using a heuristic search algorithm Alternatively, dynamic programming methods can be used to verify compliance with time constraints. Under the premise of the set of pending processing steps Select the optimal processing step subset Optimal processing steps subset Total execution cost minimize;

[0099] Subset of optimal processing steps Data structures organized according to execution logic, such as flowcharts or directed acyclic graphs, serve as exception handling chains. Perform dynamic output.

[0100] Step three further includes:

[0101] Sub-step An initial state snapshot based on the task ledger records. Build a virtual warehouse environment Initial simulation state Initial simulation state Includes real-time inventory, transportation resources, and estimated delivery time for task-related orders;

[0102] Exception handling chain Input as a sequence to be executed into the virtual warehouse environment Simulation preparation is being carried out.

[0103] Sub-step Virtual warehouse environment Discrete event simulation model Simulated exception handling chain The sequential execution of each operation step;

[0104] In each step After completion, calculate the performance risk indicators. Performance risk indicators It is the expected delay time of the order. Inventory gap and rate of change in transportation costs The weighted sum is used to quantify the potential impact;

[0105] ,

[0106] in, For steps Performance risk indicators after the end of the contract period For the expected delay time of the order, This represents the inventory shortfall. For the rate of change in transportation costs,

[0107] , , The corresponding weight coefficients, and satisfying ;

[0108] Sub-step In the exception handling chain After all simulation steps are completed, the final total performance risk index is obtained. ;

[0109] Judging the overall performance risk indicators Does it exceed the preset risk threshold? ;

[0110] like > The overall performance risk index and a snapshot of the state after simulation As feedback, the information is re-input into the anomaly planning intelligence role to iteratively optimize the anomaly handling chain. ;

[0111] like ≤ , the exception handling chain This is confirmed as the final optimized and corrected exception handling chain;

[0112] ,

[0113] in, As the final overall performance risk indicator, This represents the total number of steps in the exception handling chain. As a preset risk threshold, This is a snapshot of the final state after simulation.

[0114] Step four further includes:

[0115] Sub-step The optimized and corrected exception handling chain Transform flowcharts or directed acyclic graph data structures into task snapshots of or Formatted data structures; task snapshots Includes exception handling chain An ordered list of all processing steps Expected results of each step and the initial state snapshot of the task A secure hash algorithm is used. Computation task snapshot Snapshot hash As a unique identifier;

[0116] ,

[0117] in, For task snapshots, For a secure hash algorithm, The snapshot hash value of the task snapshot;

[0118] Sub-step , take a snapshot of the task and snapshot hash Write to task ledger Initial task record Middle; Task Ledger Employing a distributed data structure such as blockchain or hash chain, and verifying snapshot hash values. Ensure initial task record Data integrity;

[0119] Sub-step In the task ledger Create a task status tracking table Task Status Tracking Table By snapshot hash value Initial record of associated tasks ; Task status tracking table Current processing steps Initialize as an exception handling chain The first step; Task status tracking table The reserved field is used to record the corresponding smart role to be executed. and post-execution state snapshot .

[0120] Step five further includes:

[0121] Sub-step From the task status tracking table Get the current processing step All information All information This includes the operation content of this step. and the corresponding intelligent role to perform. According to the corresponding intelligent role. Registry of predefined smart roles The corresponding intelligent role is located in the middle. execution address and standardized execution interface ;

[0122] Sub-step According to the execution address and standardized execution interface , to operate content As instructions Distribute to the corresponding smart role ; Corresponding intelligent role Receive instructions Through standardized execution interfaces or status module Calling external systems atomic operations external systems This includes one of the following: an order system, a warehousing system, or a transportation system;

[0123] Sub-step external systems Perform atomic operations Then, to the corresponding intelligent role Return execution status code and business results data ; Corresponding intelligent role Execution status code and business results data Encapsulate as a post-execution state snapshot ; Snapshot the state after execution Including the time to complete each step Write to task status tracking table Furthermore, based on the logic of the exception handling chain, the task status tracking table is... Current processing steps Updated to the next pending step Next step to be executed The selection follows the topological sorting rules of directed acyclic graphs;

[0124] Sub-step Calculate the current task status tracking table The set of all completed steps In the exception handling chain Completion rate in the total set of steps ;Completeness Exceeding If the task is completed, it is determined to be finished; otherwise, the task continues to the next pending step. ;

[0125] ,

[0126] in, For completion level, For the set of completed steps, This is an exception handling chain.

[0127] Step six further includes:

[0128] Sub-step Real-time reading of task status tracking table The latest post-execution state snapshot And obtain the expected result corresponding to this step. Define the set of key fields. Key field set It includes one of the following: order status code, inventory update value, and execution status code reported by external system;

[0129] Calculate the state snapshot after execution set of key fields Actual value vs. expected result set of key fields Inconsistency measure between expected values ;

[0130] If inconsistency measurement Exceeding the preset inconsistency threshold The result of the task execution is determined to be inconsistent with the expected result;

[0131] ,

[0132] in, As a measure of inconsistency, The inconsistency threshold, Key fields The weighting coefficients, Actual value Compared with expected value The function for calculating the difference between them;

[0133] Sub-step If the judgment results are inconsistent, a snapshot of the state after execution will be taken. and inconsistency measurement As the latest feedback Input to the abnormal planning intelligent role Perform a diagnosis; the diagnosis identifies the root cause of the inconsistency and determines the rollback point; the rollback point is selected from the task status tracking table. A snapshot of the state of the previous successful execution step. Or a snapshot of the initial state of the task. ;

[0134] Sub-step Based on the state snapshot corresponding to the rollback point Calling external systems State rollback interface external systems The state logic is rolled back to the corresponding state snapshot. At the same time, the task status tracking table Current processing steps Update to rollback point; Exception planning intelligent role Based on the corresponding state snapshot And the latest feedback The large model is reused to re-plan the exception handling chain and generate a new exception handling chain.

[0135] By receiving and collecting abnormal events from multiple heterogeneous systems, and uniformly transforming them into standardized descriptions with clear structure and complete elements, the system achieves automated aggregation and unified modeling of scattered information such as orders, warehousing, and transportation. This fundamentally eliminates the tedious work of operators frequently switching between several systems and manually summarizing data, thereby improving the efficiency of abnormal event identification and the integrity of information.

[0136] Leveraging the powerful semantic understanding and logical reasoning capabilities of large models, and based on standardized abnormal event descriptions combined with pre-defined business strategy knowledge, the system dynamically and in real-time derives structured and optimal exception handling chains for specific anomalies. This breaks the limitations of traditional fixed rules and pre-defined scripts, enabling the system to make flexible and efficient decisions and task decompositions based on real-time changing business scenarios and complex anomaly types, ensuring the relevance and executability of the exception handling solutions.

[0137] In an independent virtual environment, dynamically generated exception handling chains are pre-executed and quantitatively risk-assessed. By calculating performance risk indicators, potential adverse impacts on inventory, transportation resources, and order status are predicted. Potential risks are identified and avoided before actual execution, and the handling chain can be rapidly iterated and corrected accordingly. This ensures the high feasibility and low risk of the final solution and significantly reduces actual business losses and resource waste caused by execution errors.

[0138] The optimized processing chain is written into a unified task ledger with anti-tampering features in the form of task snapshots, ensuring that the initial state of the task, execution logic and responsible entity information are reliable and traceable throughout the process. At the same time, a status tracking table is created as the core credential and reference basis for subsequent multi-intelligent role collaborative execution and status updates, providing a solid and reliable data guarantee for the orderly flow of tasks and accurate backtracking when problems occur.

[0139] This achieves a seamless transition from centralized task planning to distributed execution. By distributing tasks to corresponding intelligent roles, these roles drive the underlying heterogeneous business systems to perform atomic operations through standardized interfaces or state modules. This effectively isolates the coupling between the upper-layer intelligent planning logic and the complex integration of the underlying systems, eliminating the need for deep and complex system integration. It enables synchronous collaborative operations across systems and multiple entities, significantly reducing system deployment and maintenance costs.

[0140] By quantitatively comparing key fields with real-time monitoring of task execution feedback and preset expected results, the system immediately triggers the large model diagnostic function once the execution result is found to be inconsistent with the expectation. This function identifies the root cause and intelligently selects the optimal historical state rollback point. The system rollback interface restores the state logic and quickly replans based on the latest feedback, providing the ability to self-heal and self-repair faults, ensuring the robustness and reliability of the abnormal task handling process.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatically coordinating and organizing multi-agent tasks based on a large model, characterized in that, include: Step 1: Receive and collect heterogeneous abnormal events from at least the order system, warehousing system, and transportation system, and transform the heterogeneous abnormal events into standardized abnormal event descriptions that include the abnormality type, related orders, related resources, and urgency level. Step 2: Invoke the intelligent role for anomaly planning, and use the large model service to reason and plan the standardized anomaly event description, and dynamically generate an anomaly handling chain; Step 3: In the virtual warehouse environment, simulate the exception handling chain before execution, evaluate the potential impact of the exception handling chain on inventory, transportation resources and order status after execution, and iteratively optimize and correct the exception handling chain based on the simulation results; Step 4: Write the optimized and corrected exception handling chain as a task snapshot into a unified task ledger. The task ledger records the initial state snapshot of the task, the current processing step, the corresponding intelligent role executing the task, and the state snapshot after execution. Step 5: Based on the current processing step determined in the task ledger, the task is distributed to the corresponding intelligent role for execution. The intelligent role drives the external system to perform the corresponding operation through a standardized execution interface or status module, and at the same time updates the status snapshot in the task ledger. Step 6: Monitor the execution feedback and current status snapshot of each intelligent role in the task ledger in real time. If the task execution result is found to be inconsistent with the expected result, call the anomaly planning intelligent role. Based on the historical snapshot and the latest feedback in the task ledger, re-utilize the large model to diagnose the anomaly handling chain, roll back to the historical state, and re-plan the anomaly handling chain.

2. The method for automatically coordinating and organizing multi-agent task resolution based on a large model according to claim 1, characterized in that, Step two, specifically, includes the following: The standardized abnormal event description is used as input, and structured reasoning is performed using a large model service. Based on the reasoning results, several processing steps are generated. The processing steps are structured to define the operation content, required resources, expected results, and corresponding intelligent roles. The abnormal event handling chain is organized and stored using a flowchart or a directed acyclic graph data structure.

3. The method for automatically coordinating and organizing multi-agent task resolution based on a large model according to claim 1, characterized in that, Step three, specifically assessing the potential impact of the exception handling chain on inventory, transportation resources, and order status, includes: Calculate the performance risk index after the execution of the exception handling chain. The performance risk index includes at least one of the following: expected order delay time, inventory shortage, or transportation cost change rate. The iterative optimization and correction of the exception handling chain based on simulation results is based on the judgment and adjustment according to the preset threshold of the performance risk index.

4. The method for automatically coordinating and organizing multi-agent task resolution based on a large model according to claim 1, characterized in that, In step four, the task ledger records the initial state snapshot of the task, specifically including: recording the order status, associated inventory quantity and location, carrier information, estimated delivery time, and business rule version number at the start of the task.

5. The method for automatically coordinating and organizing multi-agent task resolution based on a large model according to claim 1, characterized in that, Step two further includes: Sub-step The standardized abnormal event description As input, through a pre-trained word embedding model Transform into high-dimensional event feature vectors ; Calculate the event feature vector With the set of vectors of known historical anomaly types Cosine similarity between ; If cosine similarity Exceeding the preset similarity threshold Directly invoke the corresponding historical exception handling plan. The historical anomaly handling plan Includes a predefined set of processing steps ; If cosine similarity The similarity threshold was not exceeded. The standardized abnormal event description Input to the largest model service Perform structured reasoning ; ， in, For the first known exception type vector in the set Types, For structured reasoning functionality in large models; Sub-step The structured reasoning Through the mind chain prompting project, standardized abnormal event descriptions are made available. and business strategy knowledge base As contextual input, the output is a set of candidate processing steps. The set of processing steps Includes several optional processing steps The candidate processing steps Define the operation content, required resources, expected results, and corresponding intelligent roles in a structured manner. ; Sub-step Processing steps for candidates Perform timing constraint verification Ensure that resource dependencies and time sequence between steps meet the requirements of the business strategy knowledge base. Constraints in ; Calculate the candidate processing steps Execution cost The execution cost It is caused by resource consumption and expected delay Obtained by weighted calculation; ， in, The execution cost of the alternative processing steps, This is a weighting coefficient for resource consumption. Resource consumption for the pending processing steps. The weighting coefficient is the expected delay. The expected delay for the pending processing steps; Sub-step Based on the execution cost Using a heuristic search algorithm Alternatively, dynamic programming methods can be used to verify compliance with time constraints. Under the premise of the set of candidate processing steps Select the optimal processing step subset The optimal processing steps subset Total execution cost minimize; Subset of the optimal processing steps The data structure, organized into a flowchart or directed acyclic graph according to the execution logic, serves as the exception handling chain. Perform dynamic output.

6. The method for automatically coordinating and organizing multi-agent task resolution based on a large model according to claim 1, characterized in that, Step three further includes: Sub-step Based on the initial state snapshot of the task ledger records Build a virtual warehouse environment Initial simulation state The initial simulation state Includes real-time inventory, transportation resources, and estimated delivery time for task-related orders; The exception handling chain Input as a sequence to be executed into the virtual warehouse environment Simulation preparation is being carried out. Sub-step The virtual warehouse environment Discrete event simulation model Simulate the exception handling chain The sequential execution of each operation step; In each step After completion, calculate the performance risk indicators. The aforementioned performance risk indicators It is the expected delay time of the order. Inventory gap and rate of change in transportation costs The weighted sum is used to quantify the potential impact; ， in, For steps Performance risk indicators after the end of the contract period For the expected delay time of the order, This represents the inventory shortfall. For the rate of change in transportation costs, , , The corresponding weight coefficients, and satisfying ; Sub-step In the exception handling chain After all simulation steps are completed, the final total performance risk index is obtained. ; Determine the total performance risk indicator Does it exceed the preset risk threshold? ; like > The total performance risk indicator and a snapshot of the state after simulation As feedback, the information is re-entered into the anomaly planning intelligent role to iteratively optimize the anomaly handling chain. ; like ≤ The exception handling chain This is confirmed as the final optimized and corrected exception handling chain; ， in, As the final overall performance risk indicator, This represents the total number of steps in the exception handling chain. As a preset risk threshold, This is a snapshot of the final state after simulation.

7. The method for automatically coordinating and organizing multi-agent task resolution based on a large model according to claim 1, characterized in that, Step four further includes: Sub-step The optimized and corrected exception handling chain Transform flowcharts or directed acyclic graph data structures into task snapshots of or Formatted data structure; the task snapshot Includes the exception handling chain An ordered list of all processing steps Expected results of each step and the initial state snapshot of the task A secure hash algorithm is used. Calculate the task snapshot Snapshot hash As a unique identifier; ， in, For task snapshots, For a secure hash algorithm, The snapshot hash value of the task snapshot; Sub-step , take the task snapshot and the snapshot hash value Write to task ledger Initial task record The task ledger mentioned above; Employing a distributed data structure such as blockchain or hash chain, the snapshot hash value is verified. Ensure the initial record of the task Data integrity; Sub-step In the task ledger Create a task status tracking table The task status tracking table By snapshot hash value Associate the task initial record ; Task status tracking table Current processing steps Initialize as the exception handling chain The first step; the task status tracking table The reserved field is used to record the corresponding smart role to be executed. and post-execution state snapshot .

8. The method for automatically coordinating and organizing multi-agent task resolution based on a large model according to claim 1, characterized in that, Step five further includes: Sub-step From the task status tracking table Get the current processing step All information All the information mentioned This includes the operation content of this step. and the corresponding intelligent role to perform. According to the corresponding intelligent role. Registry of predefined smart roles The corresponding intelligent role is located in the middle. execution address and standardized execution interface ; Sub-step According to the execution address and standardized execution interface The operation content As instructions Distributed to the corresponding intelligent role The corresponding intelligent role Receive instructions Through the standardized execution interface or status module Calling external systems atomic operations The external system This includes one of the order system, warehousing system, or transportation system; Sub-step The external system Perform the atomic operation Then, to the corresponding intelligent role Return execution status code and business results data The corresponding intelligent role Execution status code and business results data Encapsulate as a post-execution state snapshot ; Take a snapshot of the state after the execution Including the time to complete each step Write to the task status tracking table In addition, according to the logic of the exception handling chain, the task status tracking table is... Current processing steps Updated to the next pending step The next step to be executed The selection follows the topological sorting rules of directed acyclic graphs; Sub-step Calculate the current task status tracking table The set of all completed steps In the exception handling chain Completion rate in the total set of steps The degree of completion Exceeding If the task is completed, it is determined that the task has been executed; otherwise, the task continues to the next pending step. ; ， in, For completion level, For the set of completed steps, This is an exception handling chain.

9. The method for automatically coordinating and organizing multi-agent task resolution based on a large model according to claim 1, characterized in that, Step six further includes: Sub-step Read the task status tracking table in real time. The latest post-execution state snapshot And obtain the expected result corresponding to this step. Define the set of key fields. The set of key fields It includes one of the following: order status code, inventory update value, and execution status code reported by external system; Calculate the state snapshot after execution set of key fields The actual value and the expected result set of key fields Inconsistency measure between expected values ; If inconsistency measurement Exceeding the preset inconsistency threshold The result of the task execution is determined to be inconsistent with the expected result; ， in, As a measure of inconsistency, The inconsistency threshold, Key fields The weighting coefficients, Actual value Compared with expected value The function for calculating the difference between them; Sub-step If the determination results are inconsistent, a snapshot of the state after execution will be taken. and inconsistency measurement As the latest feedback Input to the abnormal planning intelligent role Perform a diagnosis; the diagnosis determines the root cause of the inconsistency and identifies the rollback point; the rollback point is selected from the task status tracking table. A snapshot of the state of the previous successful execution step. Or a snapshot of the initial state of the task. ; Sub-step Based on the state snapshot corresponding to the rollback point Calling external systems State rollback interface The external system The state logic is rolled back to the corresponding state snapshot. Simultaneously, the task status tracking table Current processing steps Update to the rollback point; the abnormal planning intelligent role Based on the corresponding state snapshot And the latest feedback The large model is then used to re-plan the exception handling chain, generating a new exception handling chain.

10. The method for automatically coordinating and organizing multi-agent task resolution based on a large model according to claim 9, characterized in that, In step six, determining that the task execution result is inconsistent with the expected result specifically includes: comparing the differences in key fields between the post-execution status snapshot in the task ledger and the expected result in the exception handling chain; the key fields include at least one of the order status code, inventory update value, or execution status code fed back by the external system; the rollback to the historical state is a rollback operation based on the initial status snapshot recorded in the task ledger or the status snapshot of the previous successfully executed step.

Images