A multi-model scheduling and reasoning optimization method supporting collaborative work of intelligent agents

Abstract

The application discloses a kind of multi-model scheduling and reasoning optimization methods of supporting agent collaborative work, and the application relates to computer technology field, including: the method is executed by scheduling hub, including the following steps: S1: task analysis step: receiving complex task, and it is parsed as multiple by different agent processing subtask.This multi-model scheduling and reasoning optimization method of supporting agent collaborative work, by introducing dynamic credibility flow control mechanism, effectively overcome the inherent uncertainty of existing static workflow scheduling method in coping with agent collaborative task.The method can be adjusted according to the credibility of runtime intermediate result real-time task execution path and resource allocation, so as to avoid the system resource deadlock and task flow blockage problem caused by task dependency relationship dynamic change or model reasoning time fluctuation, improve the overall throughput and resource utilization efficiency of multi-agent system.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a multi-model scheduling and inference optimization method that supports collaborative work of intelligent agents. Background Technology

[0002] Current technologies face a significant challenge in the program control layer of multi-model scheduling. Most systems currently rely on pre-defined static workflows for scheduling. This approach treats task dependencies between agents as a fixed, ordered sequence, allocating resources and planning the order before execution. However, the actual operating environment of agent collaboration is fraught with uncertainty. On one hand, an agent's intermediate output can fundamentally alter the execution path and dependency logic of subsequent tasks; for example, a preliminary classification result might render several subsequent analysis steps unnecessary or require priority processing. On the other hand, different models exhibit significant fluctuations in inference time due to variations in input data complexity, often causing pre-defined execution time windows to deviate from actual conditions.

[0003] This rigid program control model creates a profound contradiction with the dynamic operating environment. The direct consequence is that the entire task flow becomes exceptionally fragile. When a critical agent is blocked waiting for the delayed output of a preceding agent, it not only becomes idle but may also permanently occupy valuable computing resources, causing other ready tasks to be unable to acquire those resources and become stuck in a waiting state. This deadlock state caused by resource looping or non-critical tasks blocking the critical path severely slows down the overall system response speed, resulting in a huge waste of computing resources and preventing the parallel potential of multi-agent systems from being realized. Therefore, the urgent problem to be solved is: how to overcome the limitations of static program control and achieve a flexible collaborative control mechanism that can dynamically adapt to real-time changes in task dependencies and uncertainties in model inference. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-model scheduling and inference optimization method that supports collaborative work of intelligent agents, so as to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a multi-model scheduling and inference optimization method supporting agent collaborative work, comprising: the method being executed by a scheduling center, including the following steps: Task parsing steps: Receive a complex task and parse it into multiple subtasks to be handled by different agents; Dynamic dependency graph construction steps: Based on the logical relationship between the subtasks, construct a dynamic dependency graph. The dynamic dependency graph contains nodes and edges, where nodes represent the agent and the corresponding subtasks, and edges represent the dependency relationship between subtasks and are associated with logical strength coefficients. Scheduling and execution steps: Based on the real-time state of the dynamic dependency graph, schedule the corresponding agents to execute subtasks; The scheduling execution steps include a dynamic credibility flow control process, which dynamically adjusts the execution order and resource allocation of nodes in the dynamic dependency graph by introducing and circulating credibility parameters to adapt to the uncertainties in the task execution process.

[0006] Furthermore, the dynamic credibility flow control process includes: When the leading agent completes a subtask and produces intermediate results, the first verification mechanism is initiated. The first verification mechanism includes: calling a credibility evaluator to quickly analyze the intermediate results and output a real-time credibility score; The scheduling center compares the real-time credibility score with the expected credibility threshold preset by the subsequent dependency nodes in the dynamic dependency graph to obtain the comparison result. Based on the comparison results, the readiness status of subsequent dependent nodes is determined and resource pre-allocation is performed.

[0007] Furthermore, the method for determining the ready state of subsequent dependent nodes based on the comparison results is as follows: If the real-time credibility score is higher than the expected credibility threshold of the subsequent node, the path to the node is determined to be ready, and computing resources are immediately pre-allocated to the node, triggering it to enter the execution queue. If the real-time credibility score is lower than the expected credibility threshold of the subsequent node, the path to that node is temporarily suspended, and the second verification mechanism is initiated.

[0008] Furthermore, the second verification mechanism includes: The verification agent is initiated to perform a secondary verification of the output of the predecessor agent that generated the intermediate result; Based on the results of the secondary verification, the real-time credibility score is updated; Based on the updated real-time reliability score, the scheduling center re-executes the step of determining the readiness status of subsequent dependent nodes according to the comparison results.

[0009] Furthermore, before any agent completes its final subtask reasoning and prepares the output result, a final screening mechanism is also included: The scheduling center performs a consistency fusion judgment by integrating the agent's own output confidence level and the credibility flow history information of its input data. If the consistency fusion judgment finds a potential contradiction, a targeted secondary verification process is triggered. The disputed results are arbitrated before the final output is allowed.

[0010] Furthermore, the targeted secondary verification process includes calling a preset authoritative model to make a final ruling on the disputed result, obtaining the ruling result, and using the ruling result as the final output.

[0011] Furthermore, the credibility evaluator's rapid analysis of intermediate results is based on the analysis of the data structure integrity, logical consistency, and multi-dimensional features in the original confidence scores provided by the precursor agent, rather than reproducing the complete computational logic of the precursor agent.

[0012] Furthermore, the expected credibility threshold of each node in the dynamic dependency graph is dynamically set based on the historical execution success rate and task complexity information of the model corresponding to each agent.

[0013] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements steps such as a multi-model scheduling and inference optimization method supporting agent cooperative work.

[0014] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements steps such as a multi-model scheduling and inference optimization method supporting agent cooperative work.

[0015] This invention provides a multi-model scheduling and inference optimization method that supports collaborative work among intelligent agents. It has the following beneficial effects: This multi-model scheduling and inference optimization method, which supports agent collaboration, effectively overcomes the shortcomings of existing static workflow scheduling methods in handling inherent uncertainties in agent collaboration tasks by introducing a dynamic credibility flow control mechanism. This method can adjust task execution paths and resource allocation in real time based on the credibility of intermediate results during runtime, thereby avoiding system resource deadlock and task flow blocking problems caused by dynamic changes in task dependencies or fluctuations in model inference time, and improving the overall throughput and resource utilization efficiency of the multi-agent system.

[0016] This multi-model scheduling and inference optimization method, which supports collaborative work among intelligent agents, enhances the reliability and consistency of agent outputs by constructing a multi-level verification process that includes initial evaluation, secondary verification, and final screening. This quality control mechanism, which runs throughout task execution, not only reduces the risk of erroneous results propagating and amplifying within the system but also makes the entire collaborative system exhibit higher robustness and reliability in complex application scenarios. Attached Figure Description

[0017] Figure 1This is a flowchart illustrating a multi-model scheduling and inference optimization method for supporting agent collaboration according to the present invention. Figure 2 This is a schematic diagram of the system module interaction of a multi-model scheduling and inference optimization method that supports collaborative work of intelligent agents according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0019] Please see Figure 1 and Figure 2 This invention provides a technical solution: a multi-model scheduling and inference optimization method supporting agent collaborative work, comprising: the method is executed by a scheduling center, including the following steps: Task parsing steps: Receive a complex task and parse it into multiple subtasks to be handled by different agents; Dynamic dependency graph construction steps: Based on the logical relationships between subtasks, construct a dynamic dependency graph, where nodes represent agents and their corresponding subtasks, and edges represent the dependency relationships between subtasks and are associated with logical strength coefficients; Scheduling and execution steps: Based on the real-time state of the dynamic dependency graph, schedule the corresponding agents to execute subtasks; The scheduling and execution steps include a dynamic credibility flow control process, which dynamically adjusts the execution order and resource allocation of nodes in the dynamic dependency graph by introducing and circulating credibility parameters to adapt to the uncertainties in the task execution process.

[0020] It should be further explained that, in the specific implementation process, a multi-model scheduling and inference optimization method supporting agent collaborative work begins with the scheduling center receiving a complex task to be processed. The scheduling center first parses the task, reasonably decomposing it into a series of logically related subtasks, each handled by a specialized agent. Subsequently, based on the inherent logical relationships between these subtasks, the system constructs an initial dynamic dependency graph. In this graph, each node corresponds to an agent and its required subtasks, the directed edges between nodes represent the dependencies between subtasks, and each edge is associated with a logical strength coefficient, which initially reflects the strength of the dependency.

[0021] The core of the method lies in the dynamic credibility flow control process employed in the scheduling and execution steps. This process is implemented as follows: When the agent corresponding to a predecessor node in the dynamic dependency graph completes its computation and outputs an intermediate result, this result is not directly and unconditionally passed to subsequent dependent nodes. Instead, the system immediately initiates a dynamic evaluation process. A dedicated credibility evaluator rapidly analyzes this intermediate result. This analysis does not repeat the complete computation of the predecessor agent but focuses on meta-features such as the data structure completeness, internal logical consistency, and the confidence level provided by the agent itself, and generates a real-time credibility score based on these features.

[0022] Next, the scheduling center compares this real-time reliability score with the predefined expected reliability thresholds of each of the direct successor nodes in the dynamic dependency graph. This comparison is the basis for dynamic scheduling decisions. If the score is higher than the threshold of a certain successor node, the scheduling center determines that the execution path to that node is ready, immediately pre-allocates the necessary computing resources to this ready node, and places it in the execution queue, thereby achieving advance preparation of computing resources and efficient task connection. Conversely, if the score is lower than the threshold of a certain successor node, it is determined that the path has uncertainty and needs to be postponed.

[0023] At this point, the system initiates a secondary verification mechanism. For example, it assigns a backup verification agent to specifically verify the aforementioned intermediate results and updates the real-time credibility score based on the verification results. Subsequently, the scheduling center re-determines the path based on the new score. Through this closed-loop flow with credibility parameters as the information carrier and control link, the entire system can perceive changes in the task execution environment in real time and dynamically adjust the direction of task flow and resource allocation strategies. This effectively overcomes the blocking and deadlock problems caused by static workflows' inability to adapt to runtime uncertainties, achieving flexible and efficient collaboration between agents.

[0024] The dynamic credibility flow control process includes: When the leading agent completes a subtask and produces intermediate results, the first verification mechanism is initiated. The first verification mechanism includes: calling the credibility evaluator to quickly analyze intermediate results and output a real-time credibility score; The scheduling center compares the real-time credibility score with the expected credibility threshold preset by the subsequent dependent nodes in the dynamic dependency graph; Based on the comparison results, the readiness status of subsequent dependent nodes is determined and resource pre-allocation is performed.

[0025] It should be further explained that, in the specific implementation process, when the leading agent completes its assigned sub-task according to the instructions of the dynamic dependency graph and generates an intermediate result data packet, this data packet is not immediately pushed to subsequent waiting agents. The system first activates a dedicated reliability evaluator to analyze the intermediate result. The evaluator is designed to quickly assess the reliability of the result. Its analysis process does not repeat the complete computational logic of the leading agent, but examines the meta-features presented by the result data packet itself. These meta-features include, but are not limited to, whether the integrity of the data structure conforms to predefined specifications, whether there are logical contradictions within the data, and the self-confidence score added by the leading agent when outputting. The evaluator integrates these meta-feature information and outputs a quantitative real-time reliability score through built-in evaluation rules. This score represents the scheduling center's preliminary judgment on the current reliability of this intermediate result.

[0026] Subsequently, this real-time reliability score is immediately sent to the decision module of the scheduling center. The decision module accesses the dynamic dependency graph and reads the preset expected reliability thresholds for all subsequent nodes that directly depend on the predecessor node. The scheduling center compares the real-time reliability score with the threshold of each subsequent node one by one. This comparison operation is the direct basis for subsequent resource scheduling and task triggering. If the score meets the threshold requirement of a subsequent node, the center determines that the path is valid and generates resource pre-allocation instructions and task ready signals; if the score does not reach the threshold, the path is marked as risky, which may trigger a more in-depth verification process. Through this series of continuous and automated steps, key nodes for data quality control in the task flow are controlled, providing primary and fundamental protection for the dynamic adaptability of the entire system.

[0027] The method for determining the ready state of subsequent dependent nodes based on the comparison results is as follows: If the real-time credibility score is higher than the expected credibility threshold of the subsequent node, the path to the node is determined to be ready, and computing resources are immediately pre-allocated to the node, triggering it to enter the execution queue. If the real-time credibility score is lower than the expected credibility threshold of subsequent nodes, the path to that node is temporarily suspended, and a second verification mechanism is initiated.

[0028] It should be further explained that, in the specific implementation process, after the scheduling center compares the real-time reliability score with the expected reliability threshold of subsequent nodes, it will generate branching logic based on the comparison result. If the currently obtained real-time reliability score is higher than the expected reliability threshold set by a specific subsequent node, the scheduling center will immediately make a judgment to confirm that the execution path to that node is in a valid ready state.

[0029] Subsequently, the central control system issues instructions to the resource management unit to pre-allocate the necessary computing resources, such as graphics processing unit memory or computing cores, to the ready node. Simultaneously, it marks the node's status as executable and adds it to a high-priority task queue, ensuring that the task can begin inference immediately once it obtains all input data. This achieves proactive scheduling of computing resources and seamless task chaining. Conversely, if the real-time reliability score falls below the threshold of a subsequent node, the scheduling central control system determines that the path currently poses a reliability risk, and its status is set to postpone.

[0030] In this scenario, the system will not allocate resources but will instead trigger a more cautious secondary verification mechanism. This mechanism aims to reconfirm and evaluate the reliability of the current intermediate result, thereby avoiding resource waste and error propagation caused by continuing to execute subsequent tasks based on potentially unreliable data. This binary decision-making mechanism based on a clearly defined threshold enables the system to handle intermediate results of varying reliability with different response strategies, making it a key element in achieving efficient and robust collaborative scheduling.

[0031] The second verification mechanism includes: The verification agent is initiated to perform secondary verification on the output of the predecessor agent that generated the intermediate results; The real-time credibility score is updated based on the results of the second verification. Based on the updated real-time reliability score, the scheduling center re-executes the steps to determine the readiness status of subsequent dependent nodes according to the comparison results.

[0032] It should be further explained that, specifically, in the implementation process, when the scheduling center determines that a path is temporarily suspended due to insufficient real-time reliability score, a second verification mechanism is initiated. This mechanism first requires calling a dedicated verification agent. This verification agent can be a more simplified, faster inference model of the same type, or a rule engine dedicated to consistency checks. This verification agent will perform a secondary verification of the output of the predecessor agent that caused the dispute. This verification process is not a simple repetition of the original calculation, but may adopt different perspectives or focus on key data fragments for targeted analysis, in order to obtain an independent and comparable judgment conclusion.

[0033] Subsequently, the system updates the original real-time reliability score based on the conclusions of this secondary verification. For example, if the verification result supports the original intermediate result, the score may be increased; if potential problems are found, it will be decreased. This update process ensures that the system's judgment on data reliability is dynamic and correctable. Next, the scheduling center will, based on the updated, theoretically more accurate real-time reliability score, re-execute the process of determining the readiness status of subsequent dependent nodes based on the comparison results. That is, it will again compare the score with the expected reliability threshold of the relevant subsequent nodes and make a new determination of path readiness or postponement accordingly.

[0034] Through this closed-loop process of "evaluation-verification-update-re-decision", the system can not only respond to the uncertainty of the initial evaluation, but also self-correct based on the additional information acquired later, thereby improving the accuracy of scheduling decisions and the robustness of the system, and effectively avoiding erroneous scheduling or resource deadlock caused by a single evaluation deviation.

[0035] Before any agent completes its final subtask reasoning and prepares the output result, a final screening mechanism is also included: The scheduling center performs a consistency fusion judgment by comprehensively considering the agent's own output confidence level, the credibility of its input data, and historical information. If a potential contradiction is found in the consistency fusion judgment, a targeted secondary verification process is triggered. The disputed results are arbitrated before the final output is allowed.

[0036] It should be further explained that, in the specific implementation process, when any agent completes its final sub-task reasoning and generates the final result to be output, the system does not immediately publish the result. The scheduling center will initiate a final screening mechanism at this moment. This mechanism first performs a consistency fusion judgment: the scheduling center will retrieve the confidence level of the agent's own output, trace the confidence flow history of all input data relied upon in the process of generating the result, that is, each preceding intermediate result relied upon by the result has undergone confidence assessment and verification, and these historical records of scores constitute a confidence chain, and this information is comprehensively analyzed.

[0037] This analysis aims to detect logical inconsistencies or potential contradictions between the final result and the data chain that generated it. For example, if the final result exhibits extremely high confidence, but the historical confidence score of a key intermediate result on which it is based has consistently been low and has not been adequately verified, it will be flagged as having potential risk. When the consistency fusion judgment identifies such noteworthy risks or contradictions, the final screening mechanism will trigger a targeted secondary verification process.

[0038] This process differs from the previous second verification mechanism for intermediate results; it is more targeted. It might involve designating a higher-authority arbitration model to make a final ruling on disputed points, or convening relevant agents for a simplified round of cross-validation. Only after confirmation or correction through this final verification process will the result be allowed to be output. By adding this final screening step, the system achieves multi-level, end-to-end quality control of task execution results, minimizing the disordered output of erroneous or unreliable results, thereby improving the reliability and robustness of the entire agent-cooperative system's output.

[0039] The targeted secondary verification process involves invoking a pre-defined authoritative model to make a final ruling on the disputed results, obtaining the ruling as the final output. It should be further explained that, in the specific implementation process, when the consistency fusion judgment in the final screening mechanism identifies potential contradictions or risks, thereby triggering the targeted secondary verification process, the system will invoke an agent designated as the authoritative model according to a pre-defined strategy. This authoritative model typically has a recognized high accuracy or stronger generalization ability within a specific task domain, and its role is similar to that of an arbitrator. The scheduling center will submit the disputed final result, along with related key input data or contextual information that triggered the dispute, to this authoritative model.

[0040] The authoritative model, independent of the original processing flow, re-analyzes and evaluates the submitted data, outputting a ruling. This ruling has final effect and directly replaces the original questionable final result, being published as the final output of the agent's task. By introducing this arbitration mechanism based on the authoritative model, a final and reliable guarantee is provided for the output quality of the entire system, ensuring the authority and correctness of the final conclusion in complex and uncertain reasoning processes.

[0041] The credibility evaluator's rapid analysis of intermediate results is based on analyzing multiple meta-features in the data structure integrity, logical consistency, and the original confidence score provided by the precursor agent, rather than replicating the complete computational logic of the precursor agent. It's important to clarify that in its implementation, the credibility evaluator does not attempt to replicate the complete computational process performed by the precursor agent when executing subtasks; this is a core design aimed at saving computational resources and time. Its analysis focuses on examining several quantifiable meta-features presented by the intermediate result data packet itself. These meta-features mainly include data structure integrity (checking whether the packet format and field integrity conform to predefined interface specifications); logical consistency (analyzing whether there are logical conflicts or illogical relationships between elements within the data); and the original confidence score calculated and added by the precursor agent when outputting the result. The evaluator's built-in rule engine comprehensively considers the state of these meta-features and, through a weighted or logical combination, ultimately synthesizes a real-time score representing the overall credibility level of the current result.

[0042] This approach avoids the huge overhead of performing secondary reasoning on the complex model itself, enabling credibility assessment to be seamlessly embedded as a lightweight, low-latency checkpoint into the dynamic scheduling process, thus providing technical feasibility for real-time decision-making.

[0043] The expected credibility threshold for each node in the dynamic dependency graph is dynamically set based on the historical execution success rate of the model corresponding to each agent and the task complexity information. It should be further clarified that, in actual implementation, the expected credibility threshold is not a fixed global value, but rather a parameter set individually for each node and dynamically adjusted based on system operating experience. Its setting process mainly relies on two types of information: First, the historical execution success rate of the model corresponding to each agent, which is obtained by continuously recording the frequency with which the model's output results have been verified as correct in past tasks. The threshold for a node corresponding to a model with stable and reliable historical performance can be set at a relatively high level, indicating that the system has high expectations for its output quality. Second, task complexity information. For subtasks identified as having complex processing logic, many input variables, or being prone to ambiguity, the threshold for the corresponding node may be set more cautiously to reduce the risk of erroneous results being propagated downwards due to the difficulty of the task itself.

[0044] The system can periodically, or after completing a certain number of tasks, automatically calibrate the expected credibility thresholds of relevant nodes based on updated historical execution data and a reassessment of task complexity. This dynamic and personalized threshold setting mechanism ensures that the standards used by the scheduling center when making path decisions are aligned with actual operational conditions, thereby making the credibility-based scheduling strategy more accurate and effective.

[0045] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements steps such as those of a multi-model scheduling and inference optimization method supporting collaborative work among intelligent agents. It should be further noted that, in specific implementations, the electronic device includes necessary core computing components, including the memory and the processor. The memory is used to non-temporarily store the computer program instructions required to implement the method of this invention, as well as various types of data generated during operation, such as dynamic dependency graphs, confidence scores, and historical execution records. The processor is connected to the memory via a system bus and acts as a scheduling hub by reading and executing program instructions from the memory. During program execution, the processor specifically receives complex tasks, performs task parsing to generate a set of subtasks, and constructs and dynamically maintains a graph structure reflecting the dependencies between tasks.

[0046] Crucially, the processor, by running a program, implements a collaborative control mechanism centered on a dynamic credibility stream. This includes, but is not limited to, calling a credibility evaluator to quickly analyze the intermediate results generated by the agent to generate real-time scores, comparing the scores with preset thresholds for nodes in the graph, deciding on resource allocation and task execution order based on the comparison results, and initiating secondary verification or final screening processes under specific conditions. This electronic device embeds the methodology into the hardware, enabling this multi-model scheduling and inference optimization capability, which supports collaborative agent work, to be implemented on physical devices in a stable and repeatable manner.

[0047] A non-transitory computer-readable storage medium stores a computer program thereon. When executed by a processor, the computer program implements steps such as those of a multi-model scheduling and inference optimization method supporting agent collaboration. It should be further noted that, in specific implementations, the storage medium can physically manifest as a solid-state drive, optical disc, or flash drive, but its core feature is that the computer program stored thereon encodes the complete logic of the method of this invention. When the storage medium is connected to an electronic device and its processor reads and executes the program, the instructions in the program control the electronic device to perform a series of specific operations. These operations precisely correspond to the various steps in the method claims, including but not limited to: parsing and decomposing the complex input task into a set of subtasks; constructing an initial dynamic dependency graph structure; initiating a credibility assessment process and generating a real-time score when the agent produces intermediate results; dynamically deciding on task scheduling and resource pre-allocation based on the comparison result of the score and node thresholds; and triggering a secondary verification mechanism or a final screening process when specific conditions are met.

[0048] By transforming the method flow into executable program code and embedding it on a physical storage medium, the innovative multi-model scheduling and inference optimization technology solution provided by this invention becomes a technology product that can be independently traded, transferred, and deployed, ensuring that the invention can be implemented stably and consistently.

[0049] It's important to further explain that in the specific implementation process, an intelligent scheduling center is constructed, which can manage multiple different agents to collaboratively complete a complex task. The entire process begins with the scheduling center receiving a complex overall task. Subsequently, the scheduling center analyzes this overall task, breaking it down into multiple smaller and relatively independent subtasks. Each subtask is assigned to an agent best suited to handle that type of task; for example, one subtask might be handled by a language model, while another might be handled by an image recognition model.

[0050] To manage the execution order and dependencies between these subtasks, the scheduling hub constructs a structure called a dynamic dependency graph. In this graph, each node represents an agent and its subtasks to be executed. The connections between nodes represent dependencies between subtasks, meaning one task can only begin after another has completed. Importantly, these dependencies are not static. Each dependency connection has an associated logical strength coefficient, indicating the strength or degree of determinism of the dependency.

[0051] Once the agent begins executing its task, the dynamic trust flow control process comes into play. This process aims to address blocking and deadlock issues caused by static scheduling's inability to handle runtime uncertainties. Specifically, whenever an agent completes its subtask and produces an intermediate result, this result is not directly passed to the next agent. Instead, the system first initiates a trust evaluation process. A dedicated trust evaluator quickly checks this intermediate result. This check does not repeat the agent's complex calculations but analyzes the result's meta-features, such as the completeness of the data format, the presence of logical contradictions in the internal information, and the agent's own confidence score. Based on these analyses, the trust evaluator generates a real-time trust score.

[0052] Next, the scheduling center compares this real-time reliability score with the preset expected reliability thresholds for subsequent task nodes in the dynamic dependency graph. This comparison directly determines the scheduling behavior of subsequent tasks. If the score is higher than the threshold of a certain subsequent node, the scheduling center determines that the path is unobstructed, immediately pre-allocates the necessary computing resources for this subsequent task, and marks it as ready, so it can be executed immediately once the data is obtained. This method of pre-allocating resources greatly reduces waiting time. Conversely, if the score is lower than the threshold, it indicates that the path is risky, the scheduling center will suspend the execution of the path, and initiate a secondary verification mechanism.

[0053] The second-verification mechanism involves invoking a backup verification agent to re-examine questionable intermediate results. This verification agent may employ different methods or perspectives for analysis. Based on the second-verification result, the system updates its previous real-time credibility score. Subsequently, the scheduling center re-performs the aforementioned comparison and decision-making process based on the new score, forming a closed-loop feedback loop. This mechanism ensures that the system can dynamically adjust its behavior based on the latest information, rather than rigidly executing a pre-set plan.

[0054] In the final stage of the task flow, before an agent produces its final result and prepares to output it, the system performs a final screening procedure. The scheduling center considers information such as the confidence level of the result and the historical confidence records of all its upstream input results to make a final consistency judgment. If any inconsistency or doubt is found, the system will trigger a final arbitration process, such as having an agent designated as the authority model adjudicate, to ensure the reliability of the final output.

[0055] Furthermore, the expected credibility threshold for each node in the dynamic dependency graph is not a fixed value, but is dynamically set and adjusted based on each agent's historical performance and the complexity of the task it handles. Agents with stable and reliable performance can have their thresholds set higher; nodes handling complex tasks may have their thresholds set more cautiously.

[0056] Through the aforementioned multi-layered verification and adaptive scheduling mechanism driven by dynamic credibility flow, this invention successfully transforms a rigid static workflow into a flexible, resilient, and highly reliable collaborative system. This system effectively avoids resource deadlock and task flow blockage, fully leveraging the potential of multi-agent parallel operation, thereby achieving efficient and reliable inference optimization in complex and uncertain environments.

[0057] This invention also provides an electronic device, including a memory and a processor, wherein the processor implements the above-described method by executing a computer program stored in the memory. Simultaneously, this invention provides a non-transitory computer-readable storage medium storing a computer program, which, when executed by a processor, can also implement the above-described method. These devices and media provide concrete physical implementation carriers for the technical solutions of this invention.

[0058] Specifically, the dynamic credibility flow control process also includes the following: The scheduling center first receives a complex task to be processed and parses it, breaking it down into a series of logically related subtasks. Each subtask is assigned to a specific, specialized agent for processing. These agents are concrete instances of artificial intelligence models, such as large language models, image classification models, or decision tree models.

[0059] To clearly manage the execution order and dependencies between subtasks, the scheduling hub constructs and maintains a dynamic dependency graph. In this graph, nodes represent agents and their subtasks, and directed edges represent dependencies between tasks. Each directed edge is associated with a logical strength coefficient, the initial value of which is set based on the logical relationships during task decomposition, and is used to quantify the strength of the dependency. For example, an absolutely necessary dependency will have a logical strength coefficient close to its upper limit, while an optional or weak dependency will have a lower coefficient.

[0060] The purpose of dynamic reliability flow control is to address the problem that static workflows cannot adapt to runtime uncertainties. The specific implementation of this process includes the following steps: First, there is the reliability assessment of intermediate results. When the leading agent completes a subtask and outputs an intermediate result, a lightweight reliability evaluator is activated. This evaluator does not repeat the agent's complete computation, but rather analyzes the meta-features of the intermediate result. These meta-features include, but are not limited to: the completeness of the data structure, the inherent logical consistency of the data field values, and the confidence score of the agent's own output. The reliability evaluator synthesizes these meta-features into a real-time reliability score based on a predefined set of rules or a simple linear weighted model. This score can be represented as: S current =f(I,C,S)self ). Among them, S current Represents real-time credibility score; I represents the evaluation value for data structure integrity; C represents the evaluation value for logical consistency; S self represents the confidence score inherent to the agent itself; f represents a mapping function, which can be a weighted sum or a rule-based decision function.

[0061] Secondly, there is dynamic scheduling decision-making based on credibility scores. The scheduling center will calculate the real-time credibility score S. current The expected credibility threshold T is preset for all subsequent directly dependent nodes in the dynamic dependency graph. threshold Compare the threshold T. threshold Based on the historical success rate H of the agent corresponding to this node success The estimated complexity C of the current subtask task Dynamically set, via T threshold =g(H success C task ) calculate, where T threshold : The expected credibility threshold for subsequent nodes, a critical value used to determine whether upstream results meet downstream processing requirements; H success : The historical success rate of the corresponding intelligent agent reflects its past performance in handling similar tasks; C task g: The estimated complexity of the current subtask, which can be quantified by the task parameter size, logic chain length, etc.; g: Adjustment function, satisfying H success The higher, C task The lower the value, the higher the expected credibility threshold T. threshold The higher it can be set, for example: g(H) success C task )=k·H success -m·C task +b, where k and m are positive coefficients, and b is the baseline threshold. The comparison rule is as follows: If S current >=T threshold If S is found to be ready, the path to the subsequent node is determined to be ready. The scheduling center immediately pre-allocates computing resources, such as GPU memory, to the agent corresponding to that node and places it in a high-priority ready queue. current <T threshold If the path is deemed risky, it will be placed on hold. In this case, the system will not allocate resources but will instead initiate a secondary verification mechanism.

[0062] Third, there is a secondary verification mechanism. When a path is delayed, the scheduling center activates a backup verification agent. This verification agent independently verifies questionable intermediate results, and the verification method may differ from the original agent, for example, using a model that places greater emphasis on interpretability. The verification generates a new verification result, and the system updates the real-time credibility score based on this result, obtaining S. updated Subsequently, the dispatch center used S updated Reconnect with T threshold The system compares the results and determines the path state based on the new comparison. This process forms a closed-loop feedback loop, allowing the system to correct for deviations in the initial assessment.

[0063] Finally, there is the screening mechanism before the final output. Before any agent completes its final subtask and prepares to output the result, the scheduling center initiates the final screening. It integrates the result's own confidence level with the historical confidence scores of all its upstream inputs to perform a consistency fusion judgment. A simplified judgment logic is to check for contradictions: for example, the final result may have a high confidence level, but the historical confidence scores of its key upstream inputs may have been consistently low. If such inconsistencies are found, a final arbitration process is triggered, typically by appointing a more authoritative agent as the arbitrator to make a final ruling on the points of contention, and the ruling is used as the final output.

[0064] Through the aforementioned dynamic control mechanism driven by credibility flow and incorporating multi-layered verification and feedback, this invention achieves refined management of the task flow execution process, effectively avoiding resource deadlock and task blocking, and improving the collaborative efficiency and reliability of multi-agent systems in uncertain environments.

[0065] This method also provides an electronic device, including a memory and a processor. The memory stores computer program instructions that implement the above method, as well as runtime data, such as dynamic dependency graphs and confidence score records. The processor reads and executes the program in the memory to specifically implement the functions of the scheduling center and all the above steps.

[0066] Furthermore, this invention provides a non-transitory computer-readable storage medium, such as an SSD or optical disc, on which a computer program is stored. When this program is executed by the processor of a computing device, the aforementioned method can be implemented. This medium facilitates the distribution and deployment of the technical solution of this invention.

[0067] By introducing a dynamic credibility flow control mechanism, this method effectively overcomes the shortcomings of existing static workflow scheduling methods in dealing with the inherent uncertainties of agent collaborative tasks. This method can adjust task execution paths and resource allocation in real time based on the credibility of intermediate results during runtime, thereby avoiding system resource deadlock and task flow blocking problems caused by dynamic changes in task dependencies or fluctuations in model inference time, and improving the overall throughput and resource utilization efficiency of the multi-agent system.

[0068] By constructing a multi-level inspection process that includes initial assessment, secondary verification, and final screening, this invention enhances the reliability and consistency of the agent's output results. This quality control mechanism, which runs throughout the entire task execution process, not only reduces the risk of erroneous results being transmitted and amplified within the system, but also makes the entire collaborative system exhibit higher robustness and reliability in complex application scenarios.

[0069] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0070] 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 multi-model scheduling and inference optimization method supporting agent cooperative work, characterized in that, include: The method is executed by the scheduling center and includes the following steps: S1: Task parsing steps: Receive a complex task and parse it into multiple subtasks to be handled by different agents; S2: Dynamic Dependency Graph Construction Steps: Based on the logical relationships between the subtasks, construct a dynamic dependency graph. The dynamic dependency graph contains nodes and edges, where nodes represent the agent and its corresponding subtasks, and edges represent the dependency relationships between subtasks and are associated with logical strength coefficients. S3: Scheduling and execution steps: Based on the real-time state of the dynamic dependency graph, schedule the corresponding agents to execute subtasks; The scheduling execution steps include a dynamic credibility flow control process, which adjusts the execution order and resource allocation of nodes in the dynamic dependency graph by introducing and circulating credibility parameters to adapt to the uncertainties in the task execution process.

2. The multi-model scheduling and inference optimization method supporting agent cooperative work according to claim 1, characterized in that: The dynamic reliability flow control process includes: When the leading agent completes a subtask and produces intermediate results, the first verification mechanism is initiated. The first verification mechanism includes: calling a credibility evaluator to quickly analyze the intermediate results and output a real-time credibility score; The scheduling center compares the real-time credibility score with the preset expected credibility threshold of the subsequent dependent nodes in the dynamic dependency graph to obtain the comparison result; based on the comparison result, it determines the ready state of the subsequent dependent nodes and performs resource pre-allocation.

3. The multi-model scheduling and inference optimization method supporting agent cooperative work according to claim 2, characterized in that: The method for determining the ready state of subsequent dependent nodes based on the comparison results is as follows: If the real-time credibility score is higher than the expected credibility threshold of the subsequent node, the path to the node is determined to be ready, and computing resources are immediately pre-allocated to the node, triggering it to enter the execution queue. If the real-time credibility score is lower than the expected credibility threshold of the subsequent node, the path to that node is temporarily suspended, and the second verification mechanism is initiated.

4. The multi-model scheduling and inference optimization method supporting agent cooperative work according to claim 3, characterized in that: The second verification mechanism includes: The verification agent is initiated to perform a secondary verification of the output of the predecessor agent that generated the intermediate result; Based on the results of the secondary verification, the real-time credibility score is updated; Based on the updated real-time reliability score, the scheduling center re-executes the step of determining the readiness status of subsequent dependent nodes according to the comparison results.

5. The multi-model scheduling and inference optimization method supporting agent cooperative work according to claim 2, characterized in that: Before any agent completes its final subtask reasoning and prepares the output result, a final screening mechanism is also included: The scheduling center performs a consistency fusion judgment by integrating the agent's own output confidence level and the credibility flow history information of its input data. If the consistency fusion judgment finds a potential contradiction, a targeted secondary verification process is triggered. The disputed results are arbitrated before the final output is allowed.

6. The multi-model scheduling and inference optimization method supporting agent cooperative work according to claim 5, characterized in that: The targeted secondary verification process includes calling a preset authoritative model to make a final ruling on the disputed result, obtaining the ruling result, and using the ruling result as the final output.

7. The multi-model scheduling and inference optimization method supporting agent cooperative work according to claim 2, characterized in that: The credibility evaluator's rapid analysis of intermediate results is based on the analysis of the data structure integrity, logical consistency, and multinomial features in the original confidence scores provided by the precursor agent, rather than reproducing the complete computational logic of the precursor agent.

8. The multi-model scheduling and inference optimization method supporting agent cooperative work according to claim 1, characterized in that: The expected credibility threshold of each node in the dynamic dependency graph is dynamically set based on the historical execution success rate and task complexity information of the model corresponding to each agent.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images