Master-slave Agent collaboration method and device based on hybrid cloud architecture

Abstract

The invention discloses a master-slave Agent collaboration method and a master-slave Agent collaboration device based on a hybrid cloud architecture. The master-slave Agent collaboration method based on the hybrid cloud architecture comprises the following steps: converting user input into an atomization task through a master agent of a public cloud; distributing the atomization task to a slave agent of the private cloud; an atomization task is completed through the slave agent; and returning an atomization task execution result to the main agent. Therefore, according to the master-slave Agent collaboration method based on the hybrid cloud architecture, the master agent responsible for planning is deployed in the public cloud, and the slave agent responsible for atomization task processing is deployed in the private cloud, so that the data security can be ensured while the local deployment cost can be reduced.

Description

Technical Field

[0001] This application relates to the field of intelligent agent applications, and specifically relates to a master-slave Agent collaboration method and device based on a hybrid cloud architecture. Background Art

[0002] Large models are deep learning models trained based on massive amounts of data. For example, the Generative Pre-trained Transformer (GPT) series, the Large Language Model Meta Artificial Intelligence (LLama), DeepSeek, etc. are good at processing multimodal inputs (such as text, images, and speech) and generating context understanding. The core value of these models lies in their knowledge reserve, language generation, and pattern recognition capabilities, making them perform excellently in standardized tasks such as translation and question answering. With the progress of technology, the improvement of large models in aspects such as multimodal processing and reasoning ability directly enhances the decision-making accuracy and the scope of executable tasks of intelligent agents (Agents). Research by institutions such as Meta shows that large models provide an infrastructure similar to the human brain for the "cognition-perception-memory" module of Agents.

[0003] An Agent is an autonomous decision-making system with a large model as its "brain", capable of perceiving the environment, planning tasks, invoking tools, and executing actions. An Agent makes up for the limitations of large models in real-time data acquisition and physical operations by invoking external tools (such as databases, API interfaces). An Agent ensures that the output of the large model meets actual requirements and avoids deviation from the goal through a preset principle framework (such as role attributes, safety ethics). For example, in a multi-Agent collaboration scenario, different roles need to follow task assignment rules to achieve efficient collaboration.

[0004] Currently, the capabilities of Agents mainly rely on the large models they adopt. The more parameters a large model has, the stronger its capabilities. However, in order to make an Agent more capable, more large model parameters are required, which also means higher deployment costs. For example, the full version of Deepseek-R1 has 671 billion parameters and requires a server with 8 H20 141G GPUs for support, and the hardware cost for deployment alone is as high as 1 million yuan. Given this high cost, it is difficult for ordinary enterprises to bear the cost of locally deploying extremely large models, especially when enterprises do not have very high concurrent requirements, and it is uneconomical to locally deploy extremely large models. Therefore, considering the scale effect, for most enterprises, it is a more reasonable choice to deploy extremely large models on public clouds.

[0005] However, public cloud deployment brings new challenges, mainly including data security and response speed issues. On the one hand, uploading all data to the public cloud is relatively dangerous, especially for companies with high requirements for data confidentiality, because the business data of enterprises is often the core asset of the company. On the other hand, since the models on the public cloud are usually shared over the network and the models on the public cloud usually serve multiple users, network congestion will occur when the access volume is large, and queuing is required, which affects the response time. Especially in scenarios involving multi-Agent interaction, this kind of delay is particularly obvious.

[0006] Therefore, a new type of Agent application solution is needed. Summary of the Invention

[0007] One advantage of this application is to provide a master-slave Agent collaboration method and its device based on a hybrid cloud architecture. Among them, the master-slave Agent collaboration method based on the hybrid cloud architecture can balance the deployment cost while ensuring data security.

[0008] According to one aspect of this application, a master-slave Agent collaboration method based on a hybrid cloud architecture is provided, which includes: Converting user input into atomic tasks through the master agent in the public cloud; Allocating the atomic tasks to the slave agents in the private cloud; Completing the atomic tasks through the slave agents; Returning the execution results of the atomic tasks to the master agent.

[0009] In an embodiment of the master-slave Agent collaboration method based on the hybrid cloud architecture according to this application, allocating the atomic tasks to the slave agents in the private cloud includes: sending the atomic tasks to the local data gateway; performing data verification on the atomic tasks through the local data gateway; sending the atomized tasks after data verification to the local executor; and calling the slave agents in the private cloud through the local executor.

[0010] In an embodiment of the master-slave Agent collaboration method based on the hybrid cloud architecture according to this application, returning the execution results of the atomic tasks to the master agent includes: generating an initial execution result of the atomic task based on the completion of the atomic task by the slave agent; obtaining the result ID of the initial execution result of the atomic task to obtain an execution result with an ID; performing data desensitization on the execution result with an ID through the local data gateway to obtain a desensitized execution result of the atomic task, where the desensitized execution result of the atomic task does not include the result ID; and returning the desensitized execution result of the atomic task to the master agent.

[0011] In an embodiment of the master-slave Agent collaboration method based on a hybrid cloud architecture according to the present application, obtaining the ID of the initial execution result of the atomic task to obtain an execution result with an ID includes: writing the initial execution result of the atomic task into the local execution result library; generating a result ID through the local execution result library; and returning the result ID to the local executor to generate an execution result with an ID.

[0012] In an embodiment of the master-slave Agent collaboration method based on a hybrid cloud architecture according to the present application, the atomic task is completed by the slave agent, including: completing the atomic task by the slave agent calling tools and data.

[0013] In an embodiment of the master-slave Agent collaboration method based on a hybrid cloud architecture according to the present application, the tools include at least one of the following tools: natural language processing tool, computer vision tool, audio processing tool, machine learning tool, and content generation tool; the data includes at least one of the following data: database, office file, knowledge base, text file, and business API.

[0014] In an embodiment of the master-slave Agent collaboration method based on a hybrid cloud architecture according to the present application, the number of parameters of the large model adopted by the master agent is greater than the number of parameters of the large model adopted by the slave agent.

[0015] In an embodiment of the master-slave Agent collaboration method based on a hybrid cloud architecture according to the present application, the master-slave Agent collaboration method based on a hybrid cloud architecture further includes: Sending the initial execution result of the atomic task to the user side.

[0016] In an embodiment of the master-slave Agent collaboration method based on a hybrid cloud architecture according to the present application, sending the initial execution result of the atomic task to the user side includes: Receiving the result ID; Based on the result ID, calling the API, and then calling the initial execution result of the atomic task and sending the initial execution result of the atomic task to the user side.

[0017] According to another aspect of the present application, the present application proposes a master-slave Agent collaboration device based on a hybrid cloud architecture, which includes: An atomic task conversion module, configured to convert user input into an atomic task through the master agent in the public cloud; An atomic task allocation module, configured to allocate the atomic task to the slave agent in the private cloud; An atomic task execution module, configured to complete the atomic task through the slave agent; The execution result feedback module is used to return the execution results of the atomized tasks to the main agent.

[0018] Through the understanding of the subsequent description and the accompanying drawings, the further objectives and advantages of the present application will be fully reflected. Description of the Drawings

[0019] By describing the embodiments of the present application in more detail in combination with the accompanying drawings, the above and other objectives, features, and advantages of the present application will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0020] Figure 1 The flowchart of the master-slave Agent collaboration method based on the hybrid cloud architecture according to the embodiment of the present application is illustrated.

[0021] Figure 2 The flowchart of an example of the master-slave Agent collaboration method based on the hybrid cloud architecture according to the embodiment of the present application is illustrated.

[0022] Figure 3 The structural block diagram of the master-slave Agent collaboration device based on the hybrid cloud architecture according to the embodiment of the present application is illustrated. Detailed Embodiments

[0023] Next, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0024] It can be understood that the term "a" should be understood as "at least one" or "one or more". That is, in one embodiment, the number of an element can be one, while in other embodiments, the number of the element can be multiple. The term "a" cannot be understood as a limitation on the number. "Multiple" means greater than or equal to two.

[0025] Although ordinal numbers such as "first", "second", etc. will be used to describe various components, those components are not limited herein. The term is only used to distinguish one component from another. For example, the first component can be called the second component, and similarly, the second component can also be called the first component without departing from the teachings of the concept of the present application. The term "and / or" used herein includes any and all combinations of one or more of the associated listed items.

[0026] The terms used herein are for the purpose of describing various embodiments only and are not intended to be limiting. As used herein, the singular forms are also intended to include the plural forms unless the context clearly dictates otherwise. Additionally, it will be understood that the terms "comprising" and / or "having" when used in this specification specify the presence of the stated features, numbers, operations, components, elements, or combinations thereof, without precluding the presence or addition of one or more other features, numbers, operations, components, elements, or combinations thereof.

[0027] As previously mentioned, currently, the capabilities of an Agent mainly rely on the large model it adopts. The more parameters a large model has, the stronger its capabilities. However, to make the Agent more capable, more large model parameters are required, which also means higher deployment costs. Considering the scale effect, for most enterprises, deploying a super-large model on a public cloud is a more reasonable choice. Nevertheless, public cloud deployment brings new challenges, mainly including data security and response speed issues.

[0028] Based on this, the present application proposes a master-slave Agent collaboration method based on a hybrid cloud architecture, that is, a master-slave intelligent agent collaboration method based on a hybrid cloud architecture. Specifically, a powerful Agent is deployed on the public cloud to be responsible for high-level task planning and decomposition; at the same time, multiple small Agents are deployed locally to execute specific atomic tasks. This design not only reduces the hardware cost but also enhances the ability to protect sensitive data.

[0029] As Figures 1 to 2 shown, the master-slave Agent collaboration method based on a hybrid cloud architecture according to an embodiment of the present application is illustrated. The master-slave Agent collaboration method based on a hybrid cloud architecture includes: S110, converting user input into atomic tasks through the master intelligent agent on the public cloud; S120, allocating the atomic tasks to the slave intelligent agents on the private cloud; S130, completing the atomic tasks through the slave intelligent agents; S140, returning the execution results of the atomic tasks to the master intelligent agent.

[0030] Specifically, in step S110, the main agent in the public cloud converts the user input into atomic tasks. It should be understood that the main agent in the public cloud plays a crucial role. It is not only responsible for interacting with users, collecting user inputs and requirements, but also needs to intelligently disassemble these complex tasks and convert them into a series of specific and executable atomic tasks. Correspondingly, the large model adopted by the main agent in the public cloud has a large number of parameters. In an example of this application, the main agent in the public cloud relies on a powerful large language model (LLM), and the deployment cost is relatively high. Since the main agent in this application uses a large language model deployed in the public cloud, local deployment costs can be saved. In other words, placing the main agent with general planning capabilities in the public cloud and relying on a large-sized LLM to disassemble tasks reduces the local deployment cost. Moreover, the main agent does not involve private data during the planning process, protecting data security.

[0031] Specifically, the main agent in the public cloud first receives the input from the user through various interfaces (such as web interfaces, mobile applications, or voice assistants). After receiving the user input, the main agent in the public cloud uses a large model to parse the user's requirements and convert the task described in natural language into a form that can be understood and processed. After clarifying the user's requirements, the main agent in the public cloud disassembles the high-level task into several specific atomic tasks. Each atomic task is relatively independent and simple, suitable for execution by subordinate agents that rely on small-sized large models. The small-sized large model (sLLM) has a relatively small number of parameters. For example, 7 bytes, 14 bytes. The subordinate agent can even use a local small-sized large model (sLLM).

[0032] For example, in a smart home control system, a user wants to set an automation rule: when I get home at night, if the outdoor temperature is lower than 15 degrees, automatically turn on the heating and set the temperature to 22 degrees. The user enters this instruction through the mobile phone App. The main agent in the public cloud receives this task through the mobile phone App and uses a large model to parse the user's requirements, understanding that this is a task related to environmental monitoring and control, involving time conditions (night), location conditions (getting home), temperature threshold judgment, and device operation (turning on the heating). The task is disassembled into multiple atomic tasks: monitoring the current time and the user's geographical location; checking whether the outdoor temperature is lower than 15 degrees; if the condition is met, sending a command to the home temperature control system to adjust the settings.

[0033] In step S120, the atomization tasks are assigned to the slave agents in the private cloud. The slave agents that interact with private data more but only rely on small-sized models are deployed in the private cloud, which improves the performance and throughput compared to the models on the public cloud.

[0034] It is worth mentioning that in this application, the data gateway is used as the "gatekeeper" between the public cloud and the private cloud, responsible for filtering harmful instructions to ensure data security. The atomization tasks are sent to the local data gateway. The data gateway performs data verification to avoid executing risky local operations, such as modifying the core data of the database, etc. The verified data will be sent to the local executor. The local executor will execute the planned tasks in sequence and call multiple slave agents deployed on the private cloud.

[0035] Correspondingly, step S120 includes steps: S121, sending the atomization tasks to the local data gateway; S122, performing data verification on the atomization tasks through the local data gateway; S123, sending the atomization tasks after data verification to the local executor; S124, calling the slave agents in the private cloud through the local executor.

[0036] In step S122, the local data gateway performs data verification on the atomization tasks. Specifically, the local data gateway can check the validity of the tasks, confirm whether they involve sensitive or restricted operations (such as modifying the core data of the database), and verify whether the original tasks follow the preset security protocols. In this way, the data gateway can effectively prevent any operations that may endanger the system stability from being executed. For example, in a smart home control system, if an atomization task involves adjusting the settings of the home security system, the data gateway will pay special attention to checking whether the operation has legal authorization and evaluate its impact on the overall security. For example, for the task of detecting the outdoor temperature, since it does not involve any sensitive data or operations, the data gateway will directly pass it to the local executor; while for the task of turning on the heating and setting the temperature, although it is a regular operation, the data gateway still needs to confirm that the command comes from a trusted source and will not cause abnormally high energy consumption or other security risks.

[0037] In step S124, the local executor calls the slave agents in the private cloud. Specifically, after receiving the atomization tasks from the data gateway, the local executor selects the slave agents in the private cloud that are used to complete the atomization tasks. Each atomization task is assigned to the most suitable slave agent to execute it. For example, the location service slave agent is responsible for monitoring the user's geographical location, the temperature sensing slave agent is responsible for monitoring the outdoor temperature, and the temperature control slave agent is responsible for operating the temperature control equipment.

[0038] As Figure 2As shown, in an example of the present application, after the local actuator receives the atomized task from the data gateway, it invokes the slave agent 2-sLLM and the slave agent 3-sLLM of the private cloud.

[0039] In step S130, the atomized task is completed by the slave agent. Specifically, the atomized task is completed by the slave agent invoking tools and data. The tools include at least one of the following tools: natural language processing tool, computer vision tool, audio processing tool, machine learning tool, and content generation tool; the data includes at least one of the following data: database, office file, knowledge base, text file, and business application programming interface (API).

[0040] As Figure 2 shown, in an example of the present application, the slave agent 2-sLLM invokes the natural language processing tool and the knowledge base; the slave agent 3-sLLM invokes the machine learning tool business API.

[0041] In step S140, the execution result of the atomized task is returned to the master agent. It is worth mentioning that as a gatekeeper, the data gateway can not only ensure that harmful instructions do not enter the private cloud, but also ensure that sensitive data does not go to the public cloud, effectively protecting data security.

[0042] Specifically, after the slave agent completes the current atomic task, it returns the result of the current step to the local actuator. If the execution is correct, it continues to execute the next step until the execution ends; otherwise, it retries the failure, and terminates after multiple failures. The local actuator will write the intermediate step results and the final results of the execution into the local execution result library together. The execution result library returns the ID of this record to the local actuator, and the local actuator returns the execution result and the record ID to the data gateway. The data gateway performs business data desensitization (excluding the result ID), and then returns it to the master agent of the public cloud.

[0043] Correspondingly, step S140 includes: S141, generating an initial execution result of the atomized task based on the completion of the atomized task by the slave agent; S142, obtaining the result ID of the initial execution result of the atomized task to obtain an execution result with an ID; S143, performing data desensitization on the execution result with an ID through the local data gateway to obtain a desensitized execution result of the atomized task, where the desensitized execution result of the atomized task does not include the result ID; S144, returning the desensitized execution result of the atomized task to the master agent.

[0044] It is worth mentioning that business data desensitization refers to the process of processing sensitive information to protect personal privacy or business secrets without affecting business logic and functions. Through data desensitization technology, while retaining the data structure and format, the real sensitive data can be replaced, so that the desensitized data cannot be directly associated with specific individuals or entities, thus ensuring the security of data in scenarios such as data sharing, testing, and analysis.

[0045] For example, in a smart home control system, a user hopes to view the energy consumption report for a period of time, and these reports contain some sensitive information of the user, such as home address, peak electricity consumption time, etc. To protect user privacy while providing services, it is necessary to desensitize the relevant data.

[0046] For address information, the shielding method can be adopted to only retain the city-level information and omit the detailed block and house number. For electricity consumption records, if accurate time points are not required to analyze patterns, a disruption strategy can be selected. For example, the actual peak electricity consumption time can be offset by a few hours, or it can be blurred into a time period (for example, from 7 pm to 9 pm, which becomes the evening period after desensitization). The specific electricity consumption values can also be encrypted so that only authorized application programs can decrypt and read the real electricity consumption data.

[0047] Step S142, obtaining the result ID of the initial execution result of the atomic task to obtain the execution result with an ID, including: S1421, writing the initial execution result of the atomic task into the local execution result library; S1422, generating the result ID through the local execution result library; S1423, returning the result ID to the local executor to generate the execution result with an ID.

[0048] It is worth mentioning that the result ID is not uploaded to the public cloud, but can be sent to the user side. The user can obtain the execution result before data desensitization through the API of the local execution result library and the result ID, and have the next round of conversation with the Agent on the cloud.

[0049] Specifically, APIs can be predefined for users to call to query detailed execution results. These APIs are easy to integrate into various front-end applications. The design of the API needs to consider security. For example, through authentication to ensure that only legitimate users can access specific results. After receiving the result ID, the API provided by the local execution result library can be called with the result ID as a parameter, and the initial execution result of the atomic task can be called after the API call request is successful, and the initial execution result of the atomic task is sent to the user side.

[0050] Correspondingly, the master-slave Agent collaborative method based on the hybrid cloud architecture further includes: S150, sending the initial execution result of the atomized task to the user side. Step S150 includes: S151, receiving the result ID; S152, calling the API based on the result ID, and then calling the initial execution result of the atomized task and sending the initial execution result of the atomized task to the user side.

[0051] As Figure 3 shown, the present application also proposes a master-slave Agent collaborative device based on the hybrid cloud architecture. Specifically, the master-slave Agent collaborative device based on the hybrid cloud architecture includes an atomized task conversion module 10 for converting user input into atomized tasks through the master agent of the public cloud; an atomized task allocation module 20 for allocating atomized tasks to the slave agents of the private cloud; an atomized task execution module 30 for completing atomized tasks through the slave agents; and an execution result feedback module 40 for returning the execution result of the atomized task to the master agent.

[0052] In summary, the master-slave Agent collaborative method and its device based on the hybrid cloud architecture are elucidated. The master-slave Agent collaborative method based on the hybrid cloud architecture deploys the master agent responsible for planning in the public cloud and the slave agents responsible for processing atomized tasks in the private cloud, which can reduce the local deployment cost while ensuring data security.

[0053] The above describes the present application and its implementation manners. Such a description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present application, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments without creative efforts without departing from the creative purpose of the present application, they shall fall within the protection scope of the present application.

Claims

1. A master-slave Agent collaborative method based on a hybrid cloud architecture, characterized in that, including: converting user input into atomic tasks by the master agent in the public cloud; assigning the atomic tasks to the slave agents in the private cloud; completing the atomic tasks through the slave agents; returning the execution results of the atomic tasks to the master agent.

2. The master-slave Agent collaborative method based on a hybrid cloud architecture according to claim 1, wherein Assigning the atomic tasks to the slave agents in the private cloud includes: sending the atomic tasks to the local data gateway; performing data verification on the atomic tasks through the local data gateway; sending the atomic tasks after data verification to the local executor; invoking the slave agents in the private cloud through the local executor.

3. The master-slave Agent collaboration method based on the hybrid cloud architecture according to claim 2, wherein Returning the execution results of the atomic tasks to the master agent includes: generating an initial execution result of the atomic tasks based on the completion of the atomic tasks by the slave agents; obtaining the result ID of the initial execution result of the atomic tasks to obtain an execution result with an ID; performing data desensitization on the execution result with an ID through the local data gateway to obtain a desensitized execution result of the atomic tasks, where the desensitized execution result of the atomic tasks does not include the result ID; returning the desensitized execution result of the atomic tasks to the master agent.

4. The master-slave Agent collaboration method based on the hybrid cloud architecture according to claim 3, wherein, Obtaining the ID of the initial execution result of the atomic tasks to obtain an execution result with an ID includes: writing the initial execution result of the atomic tasks into the local execution result library; generating a result ID through the local execution result library; returning the result ID to the local executor to generate an execution result with an ID.

5. The master-slave Agent collaboration method based on a hybrid cloud architecture according to claim 4, characterized in that Completing the atomic tasks through the slave agents includes: completing the atomic tasks by the slave agents invoking tools and data.

6. The master-slave Agent collaboration method based on a hybrid cloud architecture according to claim 5, wherein The tools include at least one of the following tools: natural language processing tools, computer vision tools, audio processing tools, machine learning tools, and content generation tools; the data includes at least one of the following data: databases, office files, knowledge bases, text files, and business APIs.

7. The master-slave Agent collaboration method based on a hybrid cloud architecture according to claim 1, characterized in that The number of parameters of the large model adopted by the master agent is greater than the number of parameters of the large model adopted by the slave agent.

8. The master-slave Agent collaboration method based on the hybrid cloud architecture according to claim 4, characterized in that, The master-slave Agent collaboration method based on the hybrid cloud architecture further includes: sending the initial execution result of the atomic tasks to the user side.

9. The master-slave Agent collaboration method based on a hybrid cloud architecture according to claim 8, characterized in that, Sending the initial execution result of the atomic tasks to the user side includes: receiving the API and result ID of the execution result library; invoking the initial execution result of the atomic tasks based on the API and result ID and sending the initial execution result of the atomic tasks to the user side.

10. A master-slave Agent collaborative device based on a hybrid cloud architecture, characterized in that, including: an atomic task conversion module for converting user input into atomic tasks by the master agent in the public cloud; an atomic task assignment module for assigning atomic tasks to the slave agents in the private cloud; an atomic task execution module for completing atomic tasks through the slave agents; an execution result feedback module for returning the execution results of the atomic tasks to the master agent.

Images