Multi-agent interaction method and system and readable storage medium

Abstract

The embodiment of the invention provides a multi-agent interaction method and system and a readable storage medium, and the method comprises the steps: configuring a first agent, a second agent and a third agent of which the parameter scales are gradually increased, and adapting to the stage characteristics and complexity of an interaction task; in the initial reply stage, the first intelligent agent with the minimum parameter is called for quick response, the second intelligent agent with the medium parameter is used for processing a real-time instruction when the deep portrait is not generated, and the third intelligent agent with the maximum parameter is only started after the portrait is generated, so that resource waste caused by homogenized calling of a large model is avoided, unnecessary computing power consumption is reduced, and the computing cost is strictly controlled. Meanwhile, generation of a deep task portrait is dynamically triggered and is injected into a third agent as an enhanced context, and the decision accuracy is remarkably improved on the premise of not increasing extra cost by relying on the stronger reasoning capability and combining instant feedback and progressive optimization of staged calling; and the contradiction between decision precision and resource consumption in multi-agent interaction is effectively balanced.

Description

Technical Field

[0001] This application relates to the field of intelligent agent interaction technology, and in particular to a multi-agent interaction method, system, and readable storage medium. Background Technology

[0002] With the rapid development of Large Language Models (LLM), LLM-based multi-agent systems have become the mainstream technical solution for handling complex, multi-step tasks. Currently, the most common architecture is a centralized, synchronous, serial agent call chain. After a user inputs a request through a dialogue interface, a central routing agent first calls a large model to complete intent recognition and task decomposition, and then passes the request to subsequent agents in a fixed sequence. Regardless of the task complexity, all agents uniformly call the same large model, and the system must wait for all agents to complete their processing before returning the final result. This architecture has the following drawbacks: On the one hand, homogeneous model calls lead to high computational resource costs. Lightweight tasks that only require simple intent classification share expensive large models with business tasks that require complex reasoning, resulting in resource waste and making large-scale deployment difficult. On the other hand, the system architecture is rigid and lacks fault tolerance. The synchronous blocking feature means that the failure or timeout of any agent will cause the entire service link to fail. At the same time, the mode of central routing understanding and transfer to business agents for execution prevents business agents from obtaining the original dialogue context. They can only rely on intermediate results that may be simplified or distorted, which poses a risk of error propagation and information attenuation, affecting the accuracy of decision-making. Ultimately, this leads to the current multi-agent interaction having a problem of balancing the accuracy of interactive decision-making and the cost of computational resources. Summary of the Invention

[0003] To address the aforementioned issues and balance the decision-making accuracy and computational resource costs of multi-agent interaction, this application provides a multi-agent interaction method, system, and readable storage medium.

[0004] The embodiments of this application disclose the following technical solutions:

[0005] In a first aspect, embodiments of this application provide a multi-agent interaction method, including:

[0006] In response to the initial agent interaction command, the first agent is invoked to output the initial response content based on the initial agent interaction command, and the interaction task is analyzed according to the interaction task data associated with the initial agent interaction command to determine whether the current interaction task needs to introduce the corresponding deep task profile.

[0007] If it is determined that the current interaction task requires the introduction of the corresponding deep task profile, the deep task profile generation task is executed based on the interaction task data.

[0008] If the in-depth task profile is not yet fully generated, the second intelligent agent responds to the real-time interactive commands received in the current interaction process.

[0009] Once the deep task profile is generated, the generated deep task profile data is used as the enhanced context of the third agent, and the third agent responds to the real-time interaction commands received in the current interaction process; wherein, the model parameter size of the third agent is larger than that of the second agent, and the model parameter size of the second agent is larger than that of the first agent.

[0010] In one possible implementation, the step of performing interaction task analysis based on the interaction task data in the initial intelligent agent interaction instruction to determine whether the current interaction task needs to introduce a corresponding deep task profile includes:

[0011] Based on the interaction task data, a task complexity analysis is performed to obtain a task complexity score for the current interaction task.

[0012] If the task complexity score is greater than the first threshold, it is determined that the deep task profile needs to be introduced for the current interaction task.

[0013] If the task complexity score is not greater than the first threshold, it is determined that the current interaction task does not need to introduce the deep task profile.

[0014] In one possible implementation, the method for acquiring the interactive task data includes:

[0015] During the process of calling the first intelligent agent to output the initial response content, an independent asynchronous processing process collects associated historical interaction records, behavioral preference data and requirement details according to the interaction requirement description in the initial intelligent agent interaction instruction;

[0016] The historical interaction records, the behavioral preference data, and the requirement details are determined as the interaction task data.

[0017] In one possible implementation, the step of responding to the initial agent interaction instruction and invoking the first agent to output an initial response based on the initial agent interaction instruction includes:

[0018] The first agent performs coarse intent segmentation on the initial agent interaction command to obtain the initial response content; the response time of the first agent to the initial agent interaction command is less than a second threshold.

[0019] In one possible implementation, after performing interaction task analysis based on the interaction task data in the initial agent interaction instructions, the method further includes:

[0020] If it is determined that the current interaction task does not require the introduction of the corresponding deep task profile, the first agent or the second agent shall be used as the responding agent in the current interaction task.

[0021] In one possible implementation, the method further includes:

[0022] The real-time interactive commands are processed in a decentralized manner.

[0023] Secondly, embodiments of this application provide a multi-agent interaction system, including:

[0024] The first instruction response module is used to respond to the initial agent interaction instruction, call the first agent to output the initial response content based on the initial agent interaction instruction, and perform interaction task analysis based on the interaction task data associated with the initial agent interaction instruction to determine whether the current interaction task needs to introduce the corresponding deep task profile.

[0025] The profile generation module is used to perform a deep task profile generation task based on the interaction task data when it is determined that the current interaction task requires the introduction of the corresponding deep task profile.

[0026] The second instruction response module is used to respond to real-time interaction instructions received in the current interaction process through the second intelligent agent when the deep task profile has not been fully generated.

[0027] The third instruction response module is used to use the generated deep task profile data as the enhanced context of the third agent after the deep task profile is generated, and to respond to the real-time interaction instructions received in the current interaction process through the third agent; wherein the model parameter scale of the third agent is larger than that of the second agent, and the model parameter scale of the second agent is larger than that of the first agent.

[0028] In one possible implementation, the first instruction response module is specifically used for:

[0029] Based on the interaction task data, a task complexity analysis is performed to obtain a task complexity score for the current interaction task.

[0030] If the task complexity score is greater than the first threshold, it is determined that the deep task profile needs to be introduced for the current interaction task.

[0031] If the task complexity score is not greater than the first threshold, it is determined that the current interaction task does not need to introduce the deep task profile.

[0032] In one possible implementation, the method for acquiring the interactive task data includes:

[0033] During the process of calling the first intelligent agent to output the initial response content, an independent asynchronous processing process collects associated historical interaction records, behavioral preference data and requirement details according to the interaction requirement description in the initial intelligent agent interaction instruction;

[0034] The historical interaction records, the behavioral preference data, and the requirement details are determined as the interaction task data.

[0035] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any possible multi-agent interaction method in the first aspect.

[0036] Compared to existing technologies, this application offers the following advantages: This application provides a multi-agent interaction method, system, and readable storage medium. In this method, based on the stage characteristics and complexity differences of the interaction task, first, second, and third agents with progressively increasing parameter sizes are configured. Specifically, the first agent with the smallest parameter size is invoked for a rapid response during the initial response phase; a second agent with a medium parameter size processes real-time instructions before a deep task profile is generated; and the third agent with the largest parameter size is activated only after the deep task profile is acquired. This avoids the resource waste caused by homogeneous calls to large models for all tasks in related technologies, rationally allocating simple tasks to small- to medium-sized models, reducing unnecessary computational power consumption, and effectively controlling computational resource costs. On the other hand, the solution dynamically triggers the generation of deep task profiles by analyzing interactive task data, and injects them as enhanced context into the third agent after the profiles are completed. This enables the third agent to respond based on more comprehensive interactive task data and deep features. Combined with its stronger reasoning ability brought by its larger parameter scale, the decision accuracy of real-time interactive command response is significantly improved. At the same time, the phased agent invocation mode not only ensures the immediate feedback of the initial interaction, but also achieves the gradual optimization of decision accuracy through the synchronous upgrade of subsequent model capabilities and context quality. Ultimately, the solution achieves the improvement of decision accuracy without increasing additional computational costs, effectively balancing the contradiction between decision accuracy and resource consumption costs in multi-agent interaction. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 A flowchart illustrating a multi-agent interaction method provided in an embodiment of this application;

[0039] Figure 2 This is a schematic diagram of the structure of a multi-agent interaction system provided in an embodiment of this application. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and accompanying drawings. It should be particularly noted that the embodiments described in this application are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0041] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0042] As described earlier, with the rapid development of Large Language Models (LLMs), LLM-based multi-agent systems have become the mainstream technical solution for handling complex, multi-step tasks. Currently, the most common architecture is a centralized, synchronous, serial agent call chain. After a user inputs a request through a dialogue interface, a central routing agent first calls a large model to complete intent recognition and task decomposition, and then passes the request to subsequent agents in a fixed sequence. Regardless of the task complexity, all agents uniformly call the same large model, and the system must wait for all agents to complete their processing in sequence before returning the final result. This architecture has the following drawbacks: On the one hand, homogeneous model calls lead to high computational resource costs. Lightweight tasks that only require simple intent classification share expensive large models with business tasks that require complex reasoning, resulting in resource waste and making large-scale deployment difficult. On the other hand, the system architecture is rigid and lacks fault tolerance. The synchronous blocking feature means that the failure or timeout of any agent will cause the entire service link to fail. At the same time, the mode of central routing understanding and transfer to business agents for execution prevents business agents from obtaining the original dialogue context. They can only rely on intermediate results that may be simplified or distorted, which poses a risk of error propagation and information attenuation, affecting the accuracy of decision-making. Ultimately, this leads to the current multi-agent interaction having a problem of balancing the accuracy of interactive decision-making and the cost of computational resources.

[0043] Based on this, embodiments of this application provide a multi-agent interaction method, system, and readable storage medium. In this method, first, second, and third agents with progressively increasing parameter sizes are configured according to the stage characteristics and complexity differences of the interaction task. Specifically, the first agent with the smallest parameter size is invoked for a quick response during the initial response stage; real-time instructions are processed by the second agent with a medium parameter size before a deep profile is generated; and the third agent with the largest parameter size is activated only after the deep task profile is obtained. This avoids the resource waste problem caused by homogeneous calls to large models for all tasks in related technologies, rationally allocating simple tasks to small and medium-sized models, reducing unnecessary computing power consumption, and effectively controlling computing resource costs. On the other hand, the solution dynamically triggers the generation of deep task profiles by analyzing interactive task data, and injects them as enhanced context into the third agent after the profiles are completed. This enables the third agent to respond based on more comprehensive interactive task data and deep features. Combined with its stronger reasoning ability brought by its larger parameter scale, the decision accuracy of real-time interactive command response is significantly improved. At the same time, the phased agent invocation mode not only ensures the immediate feedback of the initial interaction, but also achieves the gradual optimization of decision accuracy through the synchronous upgrade of subsequent model capabilities and context quality. Ultimately, the solution achieves the improvement of decision accuracy without increasing additional computational costs, effectively balancing the contradiction between decision accuracy and resource consumption costs in multi-agent interaction.

[0044] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0045] See Figure 1 The figure is a flowchart illustrating a multi-agent interaction method provided in an embodiment of this application, specifically including the following steps:

[0046] S101: In response to the initial agent interaction command, the first agent is invoked to output the initial response content based on the initial agent interaction command, and the interaction task is analyzed according to the interaction task data associated with the initial agent interaction command to determine whether the current interaction task needs to introduce the corresponding deep task profile.

[0047] Responding to the initial agent interaction command is the starting point of the entire multi-agent interaction process. When the user inputs an initial request through the dialogue interface, the response system used in this method will immediately capture the initial agent interaction command and trigger the preset process scheduling logic. At this time, the system prioritizes calling the first agent to execute the immediate response task. The first agent belongs to the immediate response layer in the layered asynchronous pipeline architecture. It is essentially a lightweight agent, and its role is to meet the high real-time requirements and avoid the user from waiting. According to the task-aware hybrid model dynamic scheduling mechanism, the initial command parsing and intent coarse segmentation tasks handled by the first agent are of the lightweight and high real-time type. Therefore, the system will allocate a small to medium-sized model with a parameter scale of 7B to it. Although the computing power of this type of model is not as high as that of large models, it is efficient in fast semantic recognition, keyword extraction and simple intent matching, and can complete the core processing in a short time (specifically, the response time of the first agent needs to be limited to less than a second threshold, which can be 300ms). The initial response output by the first agent is not the final solution, but rather an instantaneous feedback containing intent classification results and a waiting prompt. For example, the initial response might be, "Hello, we are analyzing and selecting suitable retirement savings plans for you based on your retirement planning needs. Please wait..." This initial response is determined by the first agent through a rough intent classification of the initial agent interaction command. In this way, responding to the agent interaction command in a timely manner avoids excessive user waiting, aligning with the immediate feedback principle of real-time interaction systems.

[0048] While the first agent generates its initial response, the system simultaneously initiates the interaction task analysis process. This process runs concurrently with the initial response output, ensuring undiminished efficiency. The core of the interaction task analysis is to comprehensively analyze the core characteristics and potential demands of user needs based on the interaction task data associated with the initial agent's interaction commands. This interaction task data includes both the literal information initially input by the user and the user's basic interaction records and scenario-related data retrieved by the system. During the analysis, the system makes judgments based on three dimensions: first, the real-time requirement of the task, determining whether an immediate definitive response is needed; second, the computational complexity level, assessing whether fulfilling the requirement requires time-consuming operations such as deep reasoning and multi-dimensional data integration; and third, the business weight, considering the importance of the interaction task to user conversion and demand fulfillment. For example, when a user request involves a scenario that requires combining multiple dimensions of information such as personal risk preferences, asset status, and longevity risk tolerance, such as recommending pension products, the system will determine that the computational complexity is high and the business weight is high, and thus determine that the corresponding deep task profile should be introduced for the current interaction task. However, if the user only asks a simple query such as "Does a pension insurance product exist?", the system will determine that there is no need to start generating a deep task profile, and will directly complete a quick answer through the first or second intelligent agent to avoid unnecessary computational consumption.

[0049] The interactive task data is mainly obtained through the following two steps:

[0050] Step 1: In the process of calling the first intelligent agent to output the initial response content, through an independent asynchronous processing process, collect the associated historical interaction records, behavioral preference data and requirement details according to the interaction requirement description in the initial intelligent agent interaction instruction;

[0051] Step 2: Determine the historical interaction records, the behavioral preference data, and the requirement details as the interaction task data.

[0052] Specifically, the acquisition of interactive task data is synchronized with the process of the first agent outputting its initial response. This is achieved through an independent asynchronous processing process, avoiding any interference with the instantaneous response. Once the system captures the initial agent's interaction command, while scheduling a 7B parameter-scale model to support the first agent in generating immediate feedback, the asynchronous processing process initiates data collection based on the interaction requirements described in the command. For example, if a user requests a pension insurance product with stable returns and inflation protection for retirement next year, the process first parses the core keywords in the request (retirement, stable returns, inflation protection, pension insurance), using these as an index to retrieve the user's past interaction records with the system. Simultaneously, it collects the user's behavioral preference data within the platform and extracts details of the requirements (such as retirement timeframe, specific demands for return stability, and whether they are interested in additional benefits). The entire data collection process does not wait for the initial response to complete, nor does it consume the main process's computing power; parallel processing ensures high efficiency in data acquisition. Ultimately, the system integrates the collected historical interaction records, behavioral preference data, and requirement details to determine the complete interaction task data. This provides data support for subsequent judgments on whether to generate a deep task profile, ensuring low latency in the initial response and building a data foundation for the accuracy of subsequent interactions.

[0053] Furthermore, based on the interaction task data, interaction task analysis is performed to determine whether the current interaction task needs to incorporate the corresponding deep task profile process. This is achieved through the following three steps:

[0054] Step 1: Perform task complexity analysis based on the interaction task data to obtain a task complexity score for the current interaction task.

[0055] The prerequisite for determining whether a deep task profile is needed for a given interaction task is to conduct a comprehensive task complexity analysis based on the interaction task data and generate a quantitative score. The interaction task data, serving as the basis for analysis, encompasses the user's initial interaction instructions, related historical interaction records, accumulated behavioral preference data within the platform, and detailed information extracted from the requirements. During the analysis, the system constructs a multi-dimensional evaluation system, conducting quantitative analysis from three core dimensions: requirement clarity, data dependency, and decision logic complexity. The requirement clarity dimension assesses whether the user's needs are specific; for example, "requiring stable returns and inflation-resistant pension insurance for retirement next year" is a clearer requirement and receives a higher score than "recommending pension insurance." The data dependency dimension determines whether the task requires multi-dimensional, in-depth data support, such as whether it involves non-surface information like the user's asset status, risk preferences, and health condition. The decision logic complexity dimension considers whether the response task requires complex reasoning, multi-factor matching, or long-term planning analysis; for example, personalized product recommendations require multiple conditional filtering, while simple queries only require keyword matching. The system sets reasonable weight coefficients for each dimension, scores the performance of each dimension through a preset algorithm model, and finally summarizes the results to obtain a comprehensive task complexity score. The score results intuitively reflect the degree of task support for deep data, providing an objective basis for subsequent judgment.

[0056] Step 2: If the task complexity score is greater than the first threshold, determine that the current interaction task needs to introduce the deep task profile.

[0057] When the task complexity score exceeds a preset first threshold, the system will explicitly determine that the current interaction task requires the introduction of a corresponding deep task profile. The first threshold is dynamically calibrated based on specific business scenarios, system performance indicators, and user experience requirements. For example, in the question-and-answer scenario of an intelligent insurance advisor, for tasks involving long-term interests and long decision-making chains, such as retirement planning and large-sum insurance policy configuration, the threshold will be set within a reasonable range based on historical business data. This avoids overlooking tasks requiring in-depth analysis and also prevents excessive triggering of profile generation, which would lead to resource waste. At this time, the core characteristics of the task are that the requirement involves personalized matching, requires comprehensive decision-making based on multiple user characteristics, or cannot form an accurate response based solely on shallow data. For example, if a user asks, "Based on my assets and risk tolerance, recommend suitable retirement annuity insurance," its complexity score will inevitably exceed the first threshold. After the system makes a determination, it will immediately trigger the deep task profile generation task in the asynchronous processing layer. According to the task-aware hybrid model dynamic scheduling mechanism, a large model with 72B parameters is allocated to the task. Relying on its powerful reasoning and data integration capabilities, it mines key dimension information such as user risk preferences, existing assets, and health predictions from the interaction task data to construct a structured deep task profile.

[0058] Step 3: If the task complexity score is not greater than the first threshold, determine that the current interaction task does not need to introduce the deep task profile.

[0059] When the task complexity score is no greater than the first threshold, the system determines that the current interaction task does not require the introduction of a deep task profile. These types of tasks typically have simple and clear requirements, require no complex reasoning, and can be completed using only general rules or shallow information. For example, when a user asks questions like "Does the pension insurance product support online application?", the requirement does not require combining the user's deep personal characteristics; an accurate answer can be given simply by querying the basic knowledge base or matching general business rules. In this case, the system does not need to initiate a time-consuming deep profile generation task, but instead directly identifies the first or second agent as the responding agent and assigns an appropriate model based on the task's real-time requirements and simplicity.

[0060] S102: If it is determined that the current interaction task does not require the introduction of the corresponding deep task profile, the first agent or the second agent shall be used as the responding agent in the current interaction task.

[0061] When the system determines through interaction task analysis that a deep task profile is not required for the current interaction task, it means that the task is a simple query with low computational complexity, and can be answered based solely on shallow requirements and general rules. In this case, the system will select one of the first and second agents as the responding agent for the current interaction based on the task-aware hybrid model dynamic scheduling mechanism, avoiding unnecessary waste of computing power caused by calling large models with high parameter scales. If the task is a scenario with the highest priority of ultra-lightweight and immediate response, such as a user asking a simple confirmation question like "Are there any pension insurance products?", the first agent will be selected directly. Its 7B parameter small model can quickly complete keyword matching and basic semantic parsing, and output a clear answer within the response time of the second threshold (300 milliseconds). If the task does not require in-depth profiling but needs basic requirement breakdown or broad recommendations, such as "recommend several common pension insurance products," then the second intelligent agent is selected. Its 32-parameter medium-sized model can generate response content containing basic product information based on general business rules and users' shallow needs. This ensures both the maintenance of response efficiency and the satisfaction of users' basic needs. At the same time, by allocating small- and medium-sized models on demand, the consumption of computing resources is greatly reduced, avoiding the resource waste problem of homogeneous model calls in existing technologies, and achieving the optimal balance between efficiency and cost in simple task processing.

[0062] S103: If it is determined that the current interaction task needs to introduce the corresponding deep task profile, execute the deep task profile generation task according to the interaction task data.

[0063] Conversely, when it is determined that a deep task profile is required for a given interaction task, the system will immediately initiate a deep task profile generation task based on the acquired interaction task data. This task is executed through an independent asynchronous process, proceeding in parallel with the main interaction flow, without blocking the initial response feedback or subsequent real-time interaction responses. During the task profile generation process, the system will fully utilize historical interaction records, behavioral preference data, and requirement details from the interaction task data, combined with the business characteristics of the task's scenario (such as retirement planning needs in an intelligent insurance advisor scenario), to perform multi-dimensional data correlation and mining. For example, it extracts basic information such as user age, retirement time, and asset status, analyzes their past product consultation preferences, risk tolerance, and other behavioral characteristics, and then combines this with current needs to construct a multi-dimensional structured data framework. Simultaneously, according to the task-aware hybrid model dynamic scheduling mechanism, because deep task profile generation is a computationally complex task, the system will allocate a large model with 72 bytes of parameters. Leveraging the model's powerful reasoning and data integration capabilities, it will generate a deep task profile containing key dimensions such as user risk preference, existing assets, health prediction, and longevity risk tolerance, ensuring the comprehensiveness and accuracy of the profile data.

[0064] S104: Determine whether the deep task profile has been generated;

[0065] S105: If the deep task profile has not been fully generated, the second intelligent agent responds to the real-time interaction instructions received in the current interaction process.

[0066] After initiating the deep task profile generation task, the asynchronous processing layer continuously feeds back progress signals to the core business layer, including key node information such as data acquisition progress, model inference status, and structured data integration status. The state machine built into the core business layer receives these signals in real time and identifies whether the deep profile generation is complete through preset judgment rules (such as whether a complete multi-dimensional structured data object has been generated and whether data validation has passed). The purpose of real-time judgment on whether the deep task profile has been generated is to provide real-time basis for the selection of subsequent intelligent agents without interrupting the main interaction flow. This avoids the waste of computing power caused by blindly calling high-parameter models when the profile is not ready, and also prevents the main process from being blocked due to waiting for the profile to be generated, thereby ensuring that the main interaction flow can continuously respond to user needs and maintain a good real-time interactive experience.

[0067] Before the deep task profile is fully generated, step S105 uses a task-aware hybrid model dynamic scheduling mechanism. At this point, the real-time interactive task lacks deep profile support and falls into the "medium complexity, incomplete context" category. Therefore, the system allocates a medium-sized model with 32B parameters to the second agent. This model possesses stronger semantic parsing and basic reasoning capabilities than the first agent, without consuming the high computing power of a large model, achieving an optimal balance between resource cost and response effect. When the second agent responds, it generates a practical preliminary response result based on the acquired interactive task data and preset general business rules. For example, in the intelligent insurance advisor scenario, if a user asks "What are the coverage areas of the recommended products?" during the deep profile generation process, the second agent will not wait for a precise profile but will quickly return a basic explanation based on a general pension insurance framework, or a broad list containing the core coverage highlights of multiple products, and simultaneously prompt the user that "the recommendations will be optimized based on your specific situation." This response method avoids leaving users in a state of unresponsive waiting, while maintaining interaction through initial and valuable feedback. It lays the foundation for switching to a third intelligent agent after the subsequent deep profile is generated, ultimately serving the user experience goal of instant response and progressive and accurate interaction. It effectively solves the technical problems of high cost and high latency caused by all tasks relying on large models in existing technologies.

[0068] S106: Once the deep task profile is generated, the generated deep task profile data is used as the enhanced context of the third agent, and the third agent responds to the real-time interaction instructions received in the current interaction process.

[0069] Correspondingly, once the deep task profile is generated, the system immediately initiates a deep context injection mechanism. This mechanism integrates structured profile data, encompassing multiple dimensions such as user risk preferences, existing assets, health assessments, and longevity risk tolerance, into an enhanced context for the third-party agent in real time. This process is achieved through message middleware or a memory-sharing mechanism, ensuring efficient and real-time data transmission without interrupting the ongoing interaction. Unlike existing technologies where business agents can only obtain simplified intermediate results from a central routing mechanism, the third-party agent, through this injection mechanism, can simultaneously grasp the complete dialogue history, initial interaction requirements, and deep task profile data. This effectively solves the problems of error propagation and information attenuation, providing sufficient and direct decision-making basis for subsequent accurate responses. This context enhancement method deeply binds the user's deep characteristics with the interaction scenario, enabling the third-party agent to quickly locate the user's core needs.

[0070] As the agent receiving subsequent real-time interaction commands, the third agent has a larger model parameter scale than the second and first agents. Following a task-aware hybrid model dynamic scheduling mechanism, it utilizes a large model with 72 bytes of parameters, possessing stronger logical reasoning capabilities compared to the second agent's 32-byte medium-sized model and the first agent's 7-byte small-to-medium-sized model. When receiving real-time user interaction commands, the third agent, relying on enhanced context and its powerful model capabilities, can complete complex business processing. This combination of data augmentation and model upgrades avoids the resource waste caused by homogeneous calls to large models in existing technologies, while ensuring the accuracy of decisions in core business scenarios. It achieves an optimal balance between computational resource costs and interaction accuracy, allowing users to experience a gradual improvement in response quality as the interaction progresses, significantly optimizing the real-time interaction experience.

[0071] On the other hand, in one possible implementation, the multi-agent interaction method in this application embodiment can also perform decentralized processing on the received real-time interaction commands. Specifically, when a user issues a real-time interaction command during the interaction process (such as asking how the guaranteed receipt terms of a recommended product are calculated), the corresponding business agent taking over the dialogue does not need to rely on the intent parsing results of the central router. Instead, it accurately locates the associated business objects in the dialogue history through the built-in product identification module, and then optimizes the user's colloquial and ambiguous commands into standard knowledge base query statements with the help of the question rewriting module, ensuring the accuracy of command interpretation. This processing method allows the business agent to directly obtain the most original dialogue context, avoiding simplification or misinterpretation that may occur in intermediate results, effectively solving the problems of error propagation and information attenuation, while giving the system greater flexibility. Even if there is a fluctuation in a certain link, it will not affect the accuracy of the overall interaction, further improving the decision-making quality and user experience of multi-agent interaction.

[0072] This application provides a multi-agent interaction method. In this method, based on the stage characteristics and complexity differences of the interaction task, first, second, and third agents with progressively increasing parameter sizes are configured. Specifically, the first agent with the smallest parameter size is invoked for a rapid response during the initial response phase; a second agent with a medium parameter size handles real-time instructions before a deep task profile is generated; and the third agent with the largest parameter size is activated only after the deep task profile is acquired. This avoids the resource waste problem of homogeneously calling large models for all tasks in related technologies, rationally allocating simple tasks to small- to medium-sized models, reducing unnecessary computational power consumption, and effectively controlling computational resource costs. On the other hand, the solution dynamically triggers the generation of deep task profiles by analyzing interactive task data, and injects them as enhanced context into the third agent after the profiles are completed. This enables the third agent to respond based on more comprehensive interactive task data and deep features. Combined with its stronger reasoning ability brought by its larger parameter scale, the decision accuracy of real-time interactive command response is significantly improved. At the same time, the phased agent invocation mode not only ensures the immediate feedback of the initial interaction, but also achieves the gradual optimization of decision accuracy through the synchronous upgrade of subsequent model capabilities and context quality. Ultimately, the solution achieves the improvement of decision accuracy without increasing additional computational costs, effectively balancing the contradiction between decision accuracy and resource consumption costs in multi-agent interaction.

[0073] The following describes a multi-agent interaction system provided by an embodiment of this application. The multi-agent interaction system described below can be referred to in correspondence with the multi-agent interaction method described above.

[0074] See Figure 2 The figure is a schematic diagram of the structure of a multi-agent interaction system provided in an embodiment of this application, which specifically includes the following modules:

[0075] The first instruction response module 100 is used to respond to the initial intelligent agent interaction instruction, call the first intelligent agent to output the initial response content based on the initial intelligent agent interaction instruction, and perform interaction task analysis based on the interaction task data associated with the initial intelligent agent interaction instruction to determine whether the current interaction task needs to introduce the corresponding deep task profile.

[0076] The profile generation module 200 is used to perform a deep task profile generation task based on the interaction task data when it is determined that the current interaction task needs to introduce the corresponding deep task profile.

[0077] The second instruction response module 300 is used to respond to real-time interaction instructions received in the current interaction process through the second intelligent agent when the deep task profile has not been fully generated.

[0078] The third instruction response module 400 is used to, when the deep task profile is generated, use the generated deep task profile data as the enhanced context of the third agent, and respond to the real-time interaction instructions received in the current interaction process through the third agent; wherein, the model parameter scale of the third agent is larger than that of the second agent, and the model parameter scale of the second agent is larger than that of the first agent.

[0079] In one possible implementation, the first instruction response module 100 is specifically used for:

[0080] Based on the interaction task data, a task complexity analysis is performed to obtain a task complexity score for the current interaction task.

[0081] If the task complexity score is greater than the first threshold, it is determined that the deep task profile needs to be introduced for the current interaction task.

[0082] If the task complexity score is not greater than the first threshold, it is determined that the current interaction task does not need to introduce the deep task profile.

[0083] In one possible implementation, the method for acquiring the interactive task data includes:

[0084] During the process of calling the first intelligent agent to output the initial response content, an independent asynchronous processing process collects associated historical interaction records, behavioral preference data and requirement details according to the interaction requirement description in the initial intelligent agent interaction instruction;

[0085] The historical interaction records, the behavioral preference data, and the requirement details are determined as the interaction task data.

[0086] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a computer-readable storage medium storing computer instructions for causing the computer to execute the multi-agent interaction method as described in any of the above embodiments.

[0087] The computer-readable media in this application embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0088] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the multi-agent interaction method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0089] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the system, method, and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The system, method, and medium embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0090] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-agent interaction method, characterized in that, include: In response to the initial agent interaction command, the first agent is invoked to output the initial response content based on the initial agent interaction command, and the interaction task is analyzed according to the interaction task data associated with the initial agent interaction command to determine whether the current interaction task needs to introduce the corresponding deep task profile. If it is determined that the current interaction task requires the introduction of the corresponding deep task profile, the deep task profile generation task is executed based on the interaction task data. If the in-depth task profile is not yet fully generated, the second intelligent agent responds to the real-time interactive commands received in the current interaction process. Once the deep task profile is generated, the generated deep task profile data is used as the enhanced context of the third agent, and the third agent responds to the real-time interaction commands received in the current interaction process; wherein, the model parameter size of the third agent is larger than that of the second agent, and the model parameter size of the second agent is larger than that of the first agent.

2. The method according to claim 1, characterized in that, The step of performing interaction task analysis based on the interaction task data in the initial intelligent agent interaction instructions to determine whether the current interaction task needs to introduce a corresponding deep task profile includes: Based on the interaction task data, a task complexity analysis is performed to obtain a task complexity score for the current interaction task. If the task complexity score is greater than the first threshold, it is determined that the deep task profile needs to be introduced for the current interaction task. If the task complexity score is not greater than the first threshold, it is determined that the current interaction task does not need to introduce the deep task profile.

3. The method according to claim 1, characterized in that, The methods for acquiring the interactive task data include: During the process of calling the first intelligent agent to output the initial response content, an independent asynchronous processing process collects associated historical interaction records, behavioral preference data and requirement details according to the interaction requirement description in the initial intelligent agent interaction instruction; The historical interaction records, the behavioral preference data, and the requirement details are determined as the interaction task data.

4. The method according to claim 1, characterized in that, The step of responding to the initial agent interaction command and invoking the first agent to output an initial response based on the initial agent interaction command includes: The first agent performs coarse intent segmentation on the initial agent interaction command to obtain the initial response content; the response time of the first agent to the initial agent interaction command is less than a second threshold.

5. The method according to claim 1, characterized in that, After performing interaction task analysis based on the interaction task data in the initial agent interaction instructions, the method further includes: If it is determined that the current interaction task does not require the introduction of the corresponding deep task profile, the first agent or the second agent shall be used as the responding agent in the current interaction task.

6. The method according to claim 1, characterized in that, The method further includes: The real-time interactive commands are processed in a decentralized manner.

7. A multi-agent interaction system, characterized in that, include: The first instruction response module is used to respond to the initial agent interaction instruction, call the first agent to output the initial response content based on the initial agent interaction instruction, and perform interaction task analysis based on the interaction task data associated with the initial agent interaction instruction to determine whether the current interaction task needs to introduce the corresponding deep task profile. The profile generation module is used to perform a deep task profile generation task based on the interaction task data when it is determined that the current interaction task requires the introduction of the corresponding deep task profile. The second instruction response module is used to respond to real-time interaction instructions received in the current interaction process through the second intelligent agent when the deep task profile has not been fully generated. The third instruction response module is used to use the generated deep task profile data as the enhanced context of the third agent after the deep task profile is generated, and to respond to the real-time interaction instructions received in the current interaction process through the third agent; wherein the model parameter scale of the third agent is larger than that of the second agent, and the model parameter scale of the second agent is larger than that of the first agent.

8. The system according to claim 7, characterized in that, The first instruction response module is specifically used for: Based on the interaction task data, a task complexity analysis is performed to obtain a task complexity score for the current interaction task. If the task complexity score is greater than the first threshold, it is determined that the deep task profile needs to be introduced for the current interaction task. If the task complexity score is not greater than the first threshold, it is determined that the current interaction task does not need to introduce the deep task profile.

9. The system according to claim 7, characterized in that, The methods for acquiring the interactive task data include: During the process of calling the first intelligent agent to output the initial response content, an independent asynchronous processing process collects associated historical interaction records, behavioral preference data and requirement details according to the interaction requirement description in the initial intelligent agent interaction instruction; The historical interaction records, the behavioral preference data, and the requirement details are determined as the interaction task data.

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 interaction method as described in any one of claims 1-6.

Images