A multi-agent task scheduling method based on four types of cooperation protocols

Abstract

The application discloses a multi-agent task arrangement system and method based on four types of cooperation protocols, which comprises a task distribution module, a result checking module, an abnormality reporting module, a debate optimization module and a protocol scheduling module. The task distribution module generates a standardized task description object containing a task target, an execution standard and an output format; the result checking module performs compliance checking on the execution result and controls a retry mechanism; the abnormality reporting module generates a structured report and reports when an abnormality outside the capability boundary is detected; the debate optimization module collects arguments of each party when a decision difference is detected, and outputs an optimal scheme based on a weighted scoring model; and the protocol scheduling module dynamically schedules the above modules to form a closed loop in response to a task life cycle event. The application eliminates understanding ambiguity between agents, improves task execution quality, enhances system fault tolerance, solves decision conflicts by using a data-driven debate mechanism, and significantly reduces system expansion and maintenance costs.

Description

Technical Field

[0001] This application relates to artificial intelligence and multi-agent system technologies, specifically a multi-agent task orchestration system and method for complex business scenarios such as fresh food industry internet platforms and supply chain management systems. This system achieves fully automated collaboration across the entire process of task distribution, result verification, anomaly reporting, and debate optimization through four standardized collaboration protocols, effectively improving the execution efficiency and decision-making reliability of multi-agent clusters. Background Technology

[0002] With the rapid development of artificial intelligence technology, multi-agent systems are widely used in fields such as automated scheduling and intelligent decision-making. However, existing multi-agent collaboration mechanisms still have significant technical bottlenecks. First, in master-slave architectures, the lack of a standardized result verification mechanism after the master agent issues tasks leads to the inability to automatically intercept erroneous results, resulting in a high task rework rate. Second, when agents encounter problems beyond their capabilities (such as fuzzy intent recognition or business rule conflicts), the lack of a unified reporting process often causes tasks to be suspended indefinitely or errors to accumulate. Third, when faced with disagreements among multiple agents on the same issue (such as pricing conflicts or classification differences), existing systems lack a structured negotiation mechanism, usually relying on manual arbitration, which results in delayed responses and strong subjectivity. Finally, the inconsistent interface specifications between different agents lead to poor system scalability and high costs for integrating new agents. Therefore, building a standardized collaboration protocol system that covers the entire task lifecycle and possesses self-correction and decision-making game capabilities has become an urgent technical challenge. Summary of the Invention

[0003] The technical problem to be solved by this invention is that existing multi-agent systems generally suffer from the following problems when collaboratively executing complex tasks: unclear communication of task intent leading to execution deviations, lack of standardized verification mechanisms leading to uncontrollable output quality, lack of abnormal handling procedures leading to easy task interruption, and lack of structured resolution mechanisms for multi-agent decision-making disagreements leading to low decision-making efficiency and high subjectivity.

[0004] To address the aforementioned technical problems, the present invention provides a multi-agent task orchestration system based on four types of cooperative protocols. This system coordinates four standardized protocol modules—task distribution, result verification, anomaly reporting, and debate optimization—through a protocol scheduling module, thus constructing a closed-loop cooperative system covering the entire task lifecycle. The system includes: The module includes a task distribution module, a result verification module, an exception reporting module, a debate optimization module, and a protocol scheduling module.

[0005] The task distribution module generates standardized task description objects based on the task distribution protocol. These objects include fields for task objective, execution criteria, deadline, and output format. The execution criteria field defines quantifiable metrics such as accuracy and timeliness, while the output format field standardizes the data structure. This design eliminates ambiguities in understanding between humans and machines, and between machines themselves, through structured descriptions, thereby improving the accuracy of task execution.

[0006] The result verification module interfaces with the task distribution module to verify the compliance and accuracy of the results returned by the sub-agents. Internally, it includes a compliance verification unit and a retry counting unit. When verification fails, the instruction is re-executed; when the number of consecutive failures exceeds a threshold, the retry counting unit automatically triggers an exception reporting process. This closed-loop verification mechanism effectively ensures the minimum quality of the task output.

[0007] The anomaly reporting module is responsible for handling abnormal states during system operation. The anomaly detection unit identifies anomalies such as unrecognizable input and business rule conflicts. The report generation unit generates an anomaly report containing the anomaly type, problem description, and anonymized context data, and reports it to the directly responsible human administrator through the interaction unit. This establishes an efficient communication channel between the intelligent agent and human experts, preventing system crashes due to unknown anomalies.

[0008] The debate optimization module resolves decision-making disagreements among multiple agents. The opinion management unit collects the judgments and supporting data from each agent (such as historical cases, model confidence, and feature importance); the iterative convergence unit controls the number of debate rounds to prevent resource waste; and the optimal solution output unit outputs the final decision based on weighted voting or confidence ranking algorithms. This module endows the system with self-consistent conflict resolution capabilities, improving the robustness of decision-making.

[0009] The protocol scheduling module is the core hub of the system, communicating with all modules. Responding to events triggered during the task lifecycle, it dynamically schedules and launches the corresponding protocol modules. These events include, but are not limited to, changes in task status to completion, failure, or timeout, and API callback signals sent by external systems. This module establishes four types of logical call relationships between protocols: verification is triggered after task distribution; reporting is triggered when verification fails and exceeds limits; debate is triggered when disagreements exist before distribution; and the debate result is distributed as a new task. This scheduling mechanism achieves an automated closed-loop for task processing.

[0010] In addition, the system also includes a knowledge base storage module, configured to store the input data, intermediate debate logs and the final converged optimal solution generated by the debate optimization module; the debate optimization module is also configured to call the historical data in the knowledge base storage module for auxiliary decision-making when a debate is triggered.

[0011] Based on the above system, this embodiment also provides a multi-agent task orchestration method, which coordinates the execution order of task distribution, verification, reporting, and debate optimization through a protocol scheduling module to achieve automated business process flow. Simultaneously, this application also provides a storage medium storing the corresponding computer program to physically support the above method.

[0012] This technical solution has significant advantages over existing technologies: 1. By setting up standardized task description objects that include task objectives, execution standards, and output formats, unstructured business requirements are transformed into machine-readable structured contracts, eliminating the ambiguity and interpretation of natural language at the source. Combined with a result verification module, this performs mandatory closed-loop verification of the compliance and accuracy of execution results, effectively intercepting unexpected outputs and significantly improving the consistency and reliability of task delivery.

[0013] 2. A retry counting mechanism and an anomaly reporting mechanism were introduced, constructing a threshold-based negative feedback control loop. The system can automatically identify sporadic errors and attempt self-correction. Simultaneously, when encountering unknown anomalies exceeding its own capabilities, it automatically generates a structured diagnostic report containing a context snapshot and reports it to the human monitoring end. This "machine diagnosis, human decision-making" model avoids tasks being suspended indefinitely due to unknown errors, significantly enhancing the system's robustness in complex environments.

[0014] 3. To address decision-making disagreements among multiple agents, a simple majority voting mechanism is abandoned in favor of a data-driven structured debate and weighted scoring model. This model comprehensively considers the confidence level and historical weight of each agent's output, effectively filtering out low-quality noise signals and reducing the interference of extreme individual opinions on the final decision. This results in a more robust and scientifically sound optimal solution in scenarios with conflicting multi-source information.

[0015] 4. The four types of collaboration protocols define unified interface specifications and data interaction standards, enabling decoupling between agents within the system. New agents can quickly join the cluster simply by following the established protocol format, without requiring large-scale modifications to the existing system. This "plug-and-play" characteristic significantly reduces the system's development difficulty and subsequent maintenance costs, while improving its scalability.

[0016] 5. By setting up a knowledge base storage module, the system can associate and store input data, evidence exchange records, and the final convergence solution during the debate process. When facing similar dispute scenarios in the future, the system can automatically call upon historical data to assist in decision-making, avoiding redundant calculations and debates. This allows the system's decision-making efficiency and accuracy to continuously improve with the increase of running time, achieving the system's self-evolution. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the overall relationship between the four types of collaboration protocols in this embodiment.

[0018] Figure 2 This is a flowchart illustrating the task distribution protocol in this embodiment.

[0019] Figure 3 This is a flowchart illustrating the result verification protocol in this embodiment.

[0020] Figure 4 This is a timing diagram of the anomaly reporting protocol in this embodiment.

[0021] Figure 5 This is a flowchart illustrating the debate optimization protocol in this embodiment.

[0022] Figure 6 This is a sequence diagram showing the collaboration of the four types of protocols in the AI-powered e-commerce competition in this embodiment. Detailed Implementation

[0023] The embodiments of this technical solution will now be described in detail with reference to the accompanying drawings.

[0024] like Figures 1 to 6 As shown, this technical solution provides a multi-agent task orchestration system based on four types of cooperative protocols, including a task distribution module, a result verification module, an anomaly reporting module, a debate optimization module, and a protocol scheduling module. The protocol scheduling module is communicatively connected to the task distribution module, result verification module, anomaly reporting module, and debate optimization module, and is configured to schedule and start the corresponding protocol module in response to event triggering in the task lifecycle, so that the modules are sequentially connected to form a closed loop for task processing.

[0025] The task distribution module is configured to generate standardized task description objects according to the task distribution protocol and distribute these objects to at least one sub-agent. The task description object includes a task objective field, an execution criterion field, a deadline field, and an output format field. The execution criterion field defines quantifiable quality indicators, including at least one of accuracy and timeliness; the output format field defines the structure, field names, and data type of the result data. The deadline field specifies the time limit for the sub-agent to complete and return the execution result. Through this structured description, the system eliminates ambiguities in understanding between humans and machines, and between machines themselves, ensuring the controllability of the execution result from the outset.

[0026] The result verification module is configured to verify the execution results returned by the sub-agent according to the result verification protocol, and trigger a re-execution instruction when the verification fails. The result verification module includes a compliance verification unit and a retry counting unit. The compliance verification unit is configured to compare the execution results item by item based on the execution standard fields to determine whether the execution results meet the execution standards defined in the task description object. The retry counting unit is configured to record the number of consecutive failures of the same task, and trigger the exception reporting module when the number of consecutive failures exceeds a preset threshold. This closed-loop verification mechanism effectively ensures the lower limit of the quality of task output and avoids resource waste caused by invalid retries.

[0027] The anomaly reporting module is configured to generate an anomaly report and report it to the supervisory interface when preset anomaly conditions are detected, based on the anomaly reporting protocol, while suspending related tasks. The anomaly reporting module includes an anomaly detection unit, a report generation unit, and an interaction unit. The anomaly detection unit is configured to identify anomaly states, including unrecognizable input, business rule conflicts, and unavailable dependent services. The report generation unit is configured to generate an anomaly report containing the anomaly type, problem description, attempted resolution steps, and anonymized context data. The interaction unit is configured to receive intervention commands from the Directly Responsible Person (DRI) to lift the task suspension or terminate the task. This module establishes an efficient communication channel between the agent and human experts, preventing system crashes due to unknown anomalies.

[0028] The debate optimization module is configured to collect structured viewpoints from multiple agents and organize cross-debates to output the optimal solution when the output differences of multiple agents on the same decision objective exceed a preset threshold, based on the debate optimization protocol. The debate optimization module includes a viewpoint management unit, an iterative convergence unit, and an optimal solution output unit. The viewpoint management unit is configured to collect the judgment conclusions and supporting data submitted by each agent. The supporting data includes at least one of historical cases, model confidence, and feature importance. The iterative convergence unit is configured to control the number of rounds of cross-debates to not exceed a preset maximum value and update the judgments of each agent based on the acquired arguments after each round of debate. The optimal solution output unit is configured to output the optimal solution based on weighted voting or a confidence ranking algorithm after the debate ends.

[0029] In this embodiment, specifically: the optimal solution output unit determines the optimal solution through the following weighted scoring model:

[0030] in, This represents the optimal solution output. Represents the set of all candidate solutions. This represents the total number of agents participating in the debate. Indicates the first Preset weight coefficients for each agent Indicates the first Each agent considers candidate solutions The given quantitative score is positively correlated with the agent's model confidence. In actual computation, By the The internal reasoning model of each agent is calculated based on input features and normalized to ensure that the scores output by different agents are on the same scale. The derivation of this formula is based on multi-attribute decision theory. By maximizing the sum of weighted scores, the system not only considers the judgment of an individual agent but also the credibility of that judgment, thereby outputting a globally optimal solution.

[0031] In this embodiment, the specific call relationships established by the protocol scheduling module include: triggering the result verification module to start after the task distribution module completes execution; triggering the exception reporting module to start when the number of consecutive failures in the result verification module exceeds a preset threshold; triggering the debate optimization module to start if a decision disagreement is detected before the task distribution module executes; and using the optimal solution output by the debate optimization module as a new task description object, feeding it back to the task distribution module for iterative execution. This scheduling mechanism achieves an automated closed loop for task processing, enabling the system to dynamically orchestrate four types of protocols based on the context information of task execution.

[0032] In this embodiment, specifically: the system also includes a knowledge base storage module, configured to store the input data generated by the debate optimization module, intermediate debate logs, and the final converged optimal solution. The debate optimization module is also configured to call historical data in the knowledge base storage module for decision support when a debate is triggered. By accumulating the debate results into system knowledge, subsequent similar disputes can directly call historical data, thereby reducing redundant calculations and achieving continuous self-enhancement of system performance.

[0033] The operation of the above system will be described in detail below through three specific embodiments.

[0034] Example 1: Generation and Verification of Marketing Copy for Fresh Food E-commerce Platforms

[0035] In response to the "100-person group launch" event, the protocol scheduling module calls the task distribution module to generate a task description object. The task objective field is set to "generate tomato promotion copy targeting stay-at-home moms in the community"; the execution standard field is set to an accuracy rate of "must include zero pesticide residue selling points and a 30% discount"; the deadline field is set to 18:00 on the same day; and the output format field is limited to a JSON structure, including title and content fields. After receiving the task, the sub-agent "Wen Xiaochang" calls the corpus to generate the copy and returns the execution result to the result verification module. The compliance verification unit finds that the copy omits the "mentorship benefits" description and determines that the verification fails. The retry counting unit records the number of failures as 1, triggering a re-execution instruction. The sub-agent supplements the "mentorship benefits" content based on feedback and resubmits. After the compliance verification unit verifies that everything is correct, the task is completed. Through the rigid constraints of the execution standard field, the system automatically identifies and corrects errors involving missing key marketing information, ensuring the completeness of the marketing information.

[0036] Example 2: AI-powered shopping guide anomaly reporting and knowledge accumulation

[0037] When a user asks, "I want to buy the safe vegetable advertised on TV," the AI ​​shopping guide (sub-agent)'s intent recognition model has a confidence level of only 55%, and the anomaly detection unit determines it as an "input unrecognizable" anomaly. The report generation unit generates an anomaly report, including the user's original words, the confidence level value, and the keyword matching steps already attempted. The interaction unit reports this report to the DRI ticket system and suspends the current session. After reviewing the report, DRI identifies the user's intended meaning as "TA-grade Qilin watermelon" and enters the mapping relationship in the management terminal. The interaction unit receives the DRI instruction, updates the thesaurus, and releases the task suspension. The sub-agent re-identifies the user's intent and successfully recommends the product. This mechanism not only solves the blocking problem of a single session but also accumulates DRI's domain knowledge into system capabilities, achieving a knowledge accumulation effect.

[0038] Example 3: Multi-agent pricing debate and optimal solution output

[0039] For the task of "adjusting the premium for TA-grade tomatoes," three pricing agents output different suggestions: Price Advisor suggests 180% (based on tight supply and demand), Matching Agent suggests 150% (based on sufficient inventory), and Price Activation Agent suggests 170% (based on an elasticity model). If the difference exceeds a preset threshold of 5%, a debate optimization protocol is triggered. The opinion management unit collects supporting data from all parties: Price Advisor provides data on "30% increase in inquiries," Matching Agent provides data on "15 days of inventory," and Price Activation Agent provides data on "price elasticity coefficient -0.6." The iterative convergence unit controls the number of debate rounds to not exceed a preset maximum, and each agent updates its judgment based on the acquired arguments. The optimal solution output unit uses the aforementioned weighted scoring model for decision-making, inputting data and calculating to output 175% as a compromise premium solution. Simultaneously, the knowledge base storage module records the entire debate process. By introducing a mathematical model to replace manual voting, the system comprehensively considers the strength of data support and historical authority, resulting in a 175% premium solution that balances market opportunities and inventory risks, enhancing the persuasiveness and scientific rigor of the decision.

[0040] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A multi-agent task orchestration system, characterized in that, include: The task distribution module is configured to generate a standardized task description object according to the task distribution protocol, and distribute the task description object to at least one sub-agent. The task description object includes a task target field, an execution standard field, and an output format field. The result verification module is configured to verify the execution result returned by the sub-agent according to the result verification protocol, and trigger a re-execution instruction when the verification fails. The anomaly reporting module is configured to generate an anomaly report and report it to the regulatory interface when a preset anomaly condition is detected, in accordance with the anomaly reporting protocol, and simultaneously suspend related tasks. The debate optimization module is configured to collect the structured viewpoints of each agent and organize cross-debate when the output differences of multiple agents on the same decision goal exceed a preset threshold, in accordance with the debate optimization protocol, so as to output the optimal solution. The protocol scheduling module is communicatively connected to the task distribution module, result verification module, exception reporting module, and debate optimization module, and is configured to respond to event triggering in the task lifecycle, schedule and start the corresponding protocol module so that the modules are sequentially connected to form a closed loop of task processing.

2. The system according to claim 1, characterized in that, The task description object also includes a deadline field, which specifies the time limit for the sub-agent to complete and return the execution result; The execution standard field defines quantifiable quality indicators, which include at least one of accuracy and timeliness. The output format field defines the structure, field name, and data type of the result data.

3. The system according to claim 1, characterized in that, The result verification module includes: The compliance verification unit is configured to compare the execution results item by item based on the execution standard field to determine whether the execution results conform to the execution standards defined in the task description object; The retry counting unit is configured to record the number of consecutive failures of the same task, and trigger the exception reporting module when the number of consecutive failures exceeds a preset threshold.

4. The system according to claim 1, characterized in that, The anomaly reporting module includes: The anomaly detection unit is configured to identify abnormal states, including unrecognizable input, conflicting business rules, and unavailable dependent services. The report generation unit is configured to generate the exception report, which includes the exception type, problem description, attempted resolution steps, and de-identified context data. The interaction unit is configured to receive intervention instructions from the person directly responsible for the task to lift the task suspension or terminate the task.

5. The system according to claim 1, characterized in that, The debate optimization module includes: The opinion management unit is configured to collect the judgment conclusions and supporting data submitted by each intelligent agent. The supporting data includes at least one of historical cases, model confidence, and feature importance. The iterative convergence unit is configured to control the number of rounds of cross-debate to not exceed a preset maximum value, and to update the judgments of each agent based on the acquired arguments after each round of debate. The optimal solution output unit is configured to output the optimal solution based on a weighted voting or confidence ranking algorithm after the debate.

6. The system according to claim 5, characterized in that, The optimal solution output unit determines the optimal solution using the following weighted scoring model: in, This represents the optimal solution output. Represents the set of all candidate solutions. This represents the total number of agents participating in the debate. Indicates the first Preset weight coefficients for each agent Indicates the first Each agent considers candidate solutions The given quantitative score value is positively correlated with the model confidence of the agent.

7. The system according to claim 1, characterized in that, The call relationships established by the protocol scheduling module specifically include: After the task distribution module completes its execution, the result verification module is triggered to start. When the number of consecutive failures in the result verification module exceeds a preset threshold, the exception reporting module is triggered. If a decision disagreement is detected before the task distribution module is executed, the debate optimization module is triggered to start. The optimal solution output by the debate optimization module is used as a new task description object and fed back to the task distribution module for iterative execution.

8. The system according to claim 1, characterized in that, It also includes a knowledge base storage module, configured to store the input data, intermediate debate logs and the final converged optimal solution generated by the debate optimization module; The debate optimization module is also configured to call historical data in the knowledge base storage module to assist in decision-making when a debate is triggered.

9. A multi-agent task orchestration method based on four types of cooperative protocols, characterized in that, Includes the following steps: In response to a task triggering event, the protocol scheduling module calls the task distribution module to generate a task description object containing task definition information according to the task distribution protocol, and sends the task description object to at least one sub-agent; After the sub-agent returns the execution result, the protocol scheduling module calls the result verification module to verify the execution result according to the result verification protocol. If the verification fails, the re-execution instruction is triggered until the verification passes or the number of consecutive failures exceeds a preset threshold. When the number of consecutive failures exceeds a preset threshold or an anomaly exceeding the preset capability boundary is detected, the protocol scheduling module calls the anomaly reporting module to generate an anomaly report according to the anomaly reporting protocol and reports it to the regulatory interface, while suspending the associated task. When a decision disagreement is detected among multiple agents, the protocol scheduling module calls the debate optimization module to collect arguments from all parties and organize cross-debates according to the debate optimization protocol, and outputs the converged optimal solution. The protocol scheduling module encapsulates the optimal solution into a new task description object and feeds it back to the task distribution module to trigger a new round of task processing closed loop.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the multi-agent task orchestration method based on four types of cooperative protocols as described in claim 9.

Images