Cooperative task processing method and system of multi-level intelligent agent

Abstract

The invention provides a collaborative task processing method and system of a multi-level intelligent agent. In the cooperative processing method, a first agent deployed in a cloud server receives a first task created by a first user, and generates a first execution plan corresponding to the first task according to first user information corresponding to the first user. And the first agent issues the first execution plan to a second agent deployed in a first edge device associated with the first user. The second agent analyzes and splits the first execution plan, generates a plurality of sub-execution plans suitable for a plurality of third agents, and distributes the plurality of sub-execution plans to the plurality of third agents, and the plurality of third agents are deployed in at least one terminal device of the first user. And then, each third agent executes the corresponding sub-execution plan.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of artificial intelligence, and more particularly to methods and apparatus for multi-level collaborative task processing using multiple intelligent agents. Background Technology

[0002] With the development of Large LLMs and deep learning technologies, the semantic understanding, intent reasoning, and task planning capabilities of agents have reached unprecedented levels. However, while current agents are powerful, most are confined to software sandboxes on single terminals or purely cloud environments. For example, an agent on a mobile phone cannot obtain the working status of software running on a PC in real time with low latency, and vice versa. This results in the agent's "intelligence" being restricted to a local environment, unable to provide users with omniscient and continuous services.

[0003] We hope to find an improved solution that can better utilize the task planning capabilities of intelligent agents to provide users with comprehensive services. Summary of the Invention

[0004] This specification describes one or more embodiments of a collaborative task processing method and system for multi-level intelligent agents, which can provide users with comprehensive task planning and auxiliary processing capabilities.

[0005] According to a first aspect, a task processing method is provided, executed by a first intelligent agent deployed in a cloud server, the method comprising:

[0006] Receive the first task created by the first user;

[0007] Based on the first user information corresponding to the first user, generate the first execution plan corresponding to the first task;

[0008] The first execution plan is sent to the second intelligent agent deployed in the first edge device associated with the first user, so as to trigger the second intelligent agent to decompose the first execution plan into several sub-execution plans suitable for execution by the terminal device.

[0009] According to a second aspect, a task processing method is provided, executed by a second intelligent agent deployed in an edge device, the method comprising:

[0010] Receive the first execution plan for the first user issued by the first intelligent agent deployed in the cloud server;

[0011] The first execution plan is analyzed and decomposed to generate several sub-execution plans applicable to several third intelligent agents;

[0012] The plurality of sub-execution plans are distributed to the plurality of third intelligent agents, which are deployed in at least one terminal device of the first user.

[0013] According to a third aspect, a task processing method is provided, executed by a target intelligent agent deployed in a terminal device, the method comprising:

[0014] The system receives a first sub-execution plan from a second agent deployed in an edge device; the first sub-execution plan is a sub-execution plan corresponding to the target agent among several sub-execution plans generated by the second agent; the several sub-execution plans are generated by decomposing the execution plan sent from the cloud.

[0015] Execute the first sub-execution plan.

[0016] According to the fourth aspect, a task processing method is provided, including:

[0017] The first intelligent agent deployed in the cloud server receives the first task created by the first user, and generates the first execution plan corresponding to the first task based on the first user information corresponding to the first user.

[0018] The first intelligent agent sends the first execution plan to the second intelligent agent deployed in the first edge device associated with the first user;

[0019] The second intelligent agent analyzes and decomposes the first execution plan to generate several sub-execution plans applicable to several third intelligent agents, and distributes the several sub-execution plans to the several third intelligent agents, which are deployed in at least one terminal device of the first user;

[0020] Each third agent executes its corresponding sub-execution plan.

[0021] According to the fifth aspect, a task processing system is provided, comprising: a cloud server, a first edge device, and at least one terminal device;

[0022] The cloud server is equipped with a first intelligent agent, which is used to receive a first task created by a first user and generate a first execution plan corresponding to the first task based on the first user information corresponding to the first user; it is also used to send the first execution plan to a second intelligent agent deployed in the first edge device.

[0023] The second intelligent agent deployed in the first edge device is used to analyze and decompose the first execution plan, generate several sub-execution plans applicable to several third intelligent agents, and distribute the several sub-execution plans to the several third intelligent agents; the several third intelligent agents are deployed in the at least one terminal device;

[0024] Each third agent deployed in the at least one terminal device is used to execute the corresponding sub-execution plan.

[0025] According to a sixth aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in any one of the first to third aspects.

[0026] In the embodiments described in this specification, a three-tiered intelligent agent architecture is provided: edge intelligent agents, device-side intelligent agents, and cloud intelligent agents. Tasks set by the user in the cloud are transformed into specific sub-execution plans executed on the user's terminal device through the collaborative operation of each layer of intelligent agents. The sub-execution plans executed in each edge intelligent agent complement each other, collectively constituting the user's task execution plan. Furthermore, each edge intelligent agent executes its corresponding sub-execution plan on the user's terminal device and extracts and reports execution status information. The edge intelligent agent determines and tracks the user's intent state based on the execution status information of each edge intelligent agent and reports the user's intent state to the cloud intelligent agent at necessary nodes. Thus, through the collaboration of intelligent agents at each level, the user is assisted in completing the set task. Throughout the task processing process, the collaborative cooperation of the three layers of intelligent agents—edge intelligent agents, device-side intelligent agents, and cloud intelligent agents—enables a deep understanding and seamless continuation of user intent across applications and devices, comprehensively assisting the user in completing task processing. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A multi-level agent deployment framework according to one embodiment is illustrated;

[0029] Figure 2 This diagram illustrates a task processing procedure initiated in the cloud according to one embodiment.

[0030] Figure 3 This diagram illustrates a task processing scenario where an abnormal execution condition exists in one embodiment.

[0031] Figure 4 This diagram illustrates a task processing scenario where abnormal execution occurs in another embodiment. Detailed Implementation

[0032] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0033] As mentioned earlier, although intelligent agents possess powerful reasoning and task planning capabilities, current agents are either deployed in purely cloud environments or on individual terminal devices. This fragmented collaboration between devices prevents them from providing continuous, comprehensive, and omniscient services. Users expect AI assistants to deliver a unified intelligent experience across application and device boundaries. For example, suppose a user is researching information on a PC browser and then picks up their phone to go out. Unified intelligent planning across devices and applications would allow the user's phone to automatically sense and continue the reading progress and navigation intent from the PC.

[0034] Several cross-device collaboration solutions are offered in related technologies, but these solutions mainly rely on pure cloud synchronization, such as cloud storage and account synchronization. This model has several inherent drawbacks. First, there is context fragmentation and lag. Pure cloud synchronization can only handle files or simple data streams and cannot understand and synchronize the user's deep working context, such as the user's current cognitive load, the logical points of the document being processed, and the noise level of the environment. This leads to a break in the pure cloud synchronization solution's understanding of the user's intent when switching devices.

[0035] Furthermore, pure cloud-based synchronization solutions also bring bandwidth and privacy challenges. Uploading all device status data to the cloud consumes significant network bandwidth and poses serious user privacy risks. Moreover, pure cloud-based decision-making processes are lengthy and lack real-time decision-making capabilities, failing to meet the extremely high real-time requirements of scenarios such as autonomous driving and industrial control.

[0036] In view of this, the embodiments of this specification propose a framework for task processing by a three-layer intelligent agent consisting of terminal device, edge device, and cloud, aiming to achieve seamless continuity of user intent and provide users with continuous and comprehensive task planning capabilities.

[0037] Before describing the above framework, we will first give a brief introduction to the implementation of the agent.

[0038] Currently, intelligent agents are generally implemented based on large language models. Large language models (LLMs) are natural language processing models based on deep learning techniques. Their parameter count typically ranges from billions to hundreds of billions or even higher, possessing powerful language understanding and generation capabilities. Large language models can employ the Transformer architecture or its variants (such as GPT and BERT). This architecture utilizes an attention mechanism to globally model sequential data, efficiently handling long-distance dependencies and thus performing exceptionally well in natural language tasks. Large language models learn the statistical features and semantic relationships of language through pre-training on large-scale corpora, giving them outstanding generalization capabilities. The core capabilities of large language models include, but are not limited to: understanding contextual semantics, generating coherent and grammatically correct text, performing logical reasoning, and handling multi-task scenarios. Their usage typically includes two modes: direct inference and fine-tuning. In direct inference mode, the user guides the large language model to generate specific outputs by designing prompts. Cue words can be task descriptions or instructions in text form, used to stimulate the semantic understanding and generation capabilities of large language models. In fine-tuning mode, large language models are further trained on small-scale datasets in specific domains to optimize their performance on specific tasks. The powerful generalization ability and flexibility of large language models make them an important tool in the field of artificial intelligence, providing efficient and accurate solutions for automated text generation and understanding.

[0039] In some embodiments, large language models can also understand and generate data from other modalities (such as visual and audio data). In this case, large language models can also be called multimodal large language models (MLLMs). MLLMs provide a richer and more natural interactive experience by integrating multiple types of input and output, such as text, images, and sound. The core advantage of MLLMs lies in their ability to process and understand information from different modalities and fuse this information to complete complex tasks. For example, MLLMs can analyze an image and generate descriptive text, or generate a corresponding image based on a text description. This cross-modal understanding and generation capability makes MLLMs widely applicable across multiple fields.

[0040] It should be noted that the key technologies of large language models can be found in the detailed description in the paper "A Survey of Large Language Models" (paper number: arXiv:2303.18223v16, published on March 11, 2025, public link: https: / / doi.org / 10.48550 / arXiv.2303.18223), and will not be repeated here.

[0041] Building upon the reasoning capabilities of a large language model, the intelligent agent integrates additional modules, thereby acquiring comprehensive capabilities such as task planning and execution. These additional modules include an extended tool module, which can access resources such as search engines and databases, enabling the agent to overcome the limitations of prompt word content. Additional modules may also include a memory module, storing user preferences, historical interactions, and other information to achieve personalized services. Furthermore, the agent can include an action module or execution module, which can perform practical operations such as API calls, email sending, and command control. With the large language model as its core reasoning engine, and through the integration and expansion of these additional modules, the intelligent agent possesses comprehensive analysis, reasoning, and execution capabilities.

[0042] Leveraging the comprehensive capabilities of intelligent agents, the user experience is enhanced through a multi-level, multi-agent collaborative architecture in the embodiments of this specification.

[0043] Figure 1 This illustrates a multi-level agent deployment framework according to one embodiment. Figure 1 As shown, the multi-level, multi-agent architecture includes a three-layer agent deployment: terminal (device), edge (edge ​​device), and cloud (cloud). These are described below.

[0044] The cloud-based intelligent agent is deployed on a cloud server and includes a large model (LLM) and a global storage module. The LLM is used for core complex reasoning. The global storage module stores a global knowledge base that may be used for reasoning, which can include knowledge texts, knowledge graphs, and other data from various domains. The global storage module can also store user information for each user. User information may include short-term state information such as user schedules, as well as long-term memory information that reflects user preferences and characteristics. This long-term memory information is also known as user profile information. In some implementations, to protect user privacy, user profile information is stored and recorded in the form of abstract semantic embedding vectors.

[0045] Cloud-based intelligent agents can be intelligent agents with a large number of parameters and strong performance.

[0046] Edge agents are deployed in edge devices located at the edge of the network. These edge devices can include gateway devices, routers, and LAN servers, which can be home networks or internal networks set up by a company or organization.

[0047] The edge agent sits between the cloud agent and the edge agent, responsible for intermediate-level task processing. Specifically, it breaks down task execution plans issued by the cloud agent into sub-task plans suitable for edge agent execution and distributes them to the edge agents. For execution status and other data reported by edge agents, the edge agent performs context aggregation and fusion of heterogeneous data reported by various edge agents within the local area network, and makes local low-latency decisions. When it determines that a local decision is not possible, it reports to the cloud agent.

[0048] Compared to cloud-based intelligent agents, edge intelligent agents can be slightly smaller and have lower performance, and can be implemented as intelligent agents with medium parameters.

[0049] Edge-side intelligent agents are deployed on the user's terminal devices, such as on the user's personal computer (PC) or mobile devices (phones, tablets, etc.). Edge-side intelligent agents can be lightweight agents, responsible for data collection and simple interactions with the user.

[0050] Specifically, the client-side intelligent agent can include a browser agent deployed in the browser. This browser agent can take the form of a browser plugin or similar component, maintaining the context of the entire browsing system. More specifically, the browser agent can extract information such as the webpage DOM, browsing history, number of tabs, active tabs, and webpage content, and interact with the user through the browser.

[0051] Edge-side intelligent agents can also include mobile agents deployed on mobile devices. In specific instances, mobile agents can manifest as applications on smartphones or system programs embedded in the phone's operating system. Mobile agents can extract information such as GPS data, chat history, device status, network connectivity, and app running status, and interact with users through the mobile device.

[0052] The edge-side intelligent agent can also be a PC agent deployed on a PC. In specific instances, the PC agent can manifest as an application on the PC or a system program embedded in the PC operating system. The PC agent can extract information such as running process information, clipboard information, file system information, and app running status, and interact with the user through the PC.

[0053] In other embodiments, the edge-side intelligent agent can also be deployed in other forms of terminal devices and have other forms of manifestation, such as a micro intelligent agent deployed in a wearable device, an intelligent agent deployed in a home IoT device, and so on.

[0054] Understandably, multiple edge agents are connected to the edge device, receive sub-task plans issued by it, and report the collected status data to it.

[0055] pass Figure 1 The collaborative work of three levels of intelligent agents—mid-end intelligent agents, edge intelligent agents, and cloud intelligent agents—enables deep understanding and seamless continuity of user intent across applications and devices, providing users with continuous and comprehensive task planning capabilities.

[0056] It is understandable that users need to pre-register and associate their devices in order to utilize the capabilities and services of multi-level intelligent agents. For example, if the terminal devices of users U1 and U2 are in the same local area network (e.g., the same office), then the terminal devices of user U1 (e.g., mobile phone M1, laptop N1) and user U2 (e.g., mobile phone M2, PC P1) can both be associated with the edge device Edge1 (e.g., a gateway device) of that local area network. The association relationship between each terminal device (and the intelligent agents deployed therein) and the user (referred to as the first association relationship) is recorded in the edge device Edge1. Similarly, assuming that the terminal devices of user U3 (e.g., mobile phone M3, laptop N2, and PC P2) are associated with the edge device Edge2, this association relationship can be recorded in the edge device Edge2. In addition, the association relationships between each user and the edge device (referred to as the second association relationship) can be reported to the cloud server and recorded in the cloud intelligent agents. The second association may include, for example, user U1 and user U2 being associated with edge device Edge1, and user U3 being associated with edge device Edge2.

[0057] Based on this, each layer of intelligent agents can provide task assistant services to users. The following describes the process of task processing through the collaborative efforts of the three-layer intelligent agents (device, edge, and cloud).

[0058] Figure 2 This diagram illustrates a task processing procedure initiated in the cloud according to one embodiment.

[0059] like Figure 2 As shown, in step S201, the first user creates a first task in the cloud-based intelligent agent. It should be noted that the first user here refers to any specific user; for clarity, it will be referred to as user U1 below. The first task can be any task suitable for agent-assisted processing, hereinafter referred to as T1. Thus, in this step, the cloud-based intelligent agent receives task T1 created by user U1.

[0060] In different implementations, users can create tasks in the cloud through various entry points and methods. In one embodiment, the cloud-based intelligent agent has a dedicated client application on the user's terminal side. User U1 can create tasks through this client application. In one example, the terminal-side intelligent agent of user U1 also acts as a "client" of the cloud-based intelligent agent, through which the user can create tasks. In another example, the cloud-based intelligent agent can provide other interface programs, such as task planning applications or tools deployed on the terminal device, or intelligent assistant tools, such as the phone's "Calendar," "Notes," and voice assistant. These tools can be considered as interface applications of the cloud-based intelligent agent. Users can create task T1 by entering schedules or task descriptions in these tools.

[0061] Task T1 can be any task suitable for planning and assisted processing by the agent. For example, in one case, user U1 creates task T1 as: Attend a meeting at XXX company at 10:00 AM tomorrow.

[0062] After receiving the task T1, in step S202, the cloud-based intelligent agent generates the execution plan P1 corresponding to task T1 based on the user information D1 corresponding to user U1.

[0063] Specifically, user information D1 may include user U1's schedule / status information, as well as long-term memory information reflecting user U1's preferences / characteristics, i.e., user profile information. The task description of task T1 and user information D1 of user U1 can be input together into the inference large model LLM in the cloud intelligent agent to generate an execution plan P1 for executing task T1.

[0064] To protect user privacy, in one embodiment, user profile information is stored on a cloud server in the abstract form of semantic embedding vectors. During inference in the large LLM model, the task description text for task T1 is first embedded using an embedding layer, and the embedded task representation is then input into subsequent processing layers. For user profile information, the corresponding semantic embedding vector is directly fused with the task representation and input into subsequent processing layers without going through the embedding layer. After processing by the various network layers in the large model, an execution plan P1 is generated.

[0065] In one embodiment, the cloud-based agent generates execution plan P1 by referencing a global knowledge base stored in a cloud server. For example, the cloud-based agent can generate execution plan P1 based on the global knowledge base using a retrieval-augmented generation (RAG) method.

[0066] It should be noted that the execution plan P1 generated by the cloud-based intelligent agent is a relatively general, coarse-grained, high-level execution plan.

[0067] Continuing the previous example, assuming task T1 is "Attend a meeting at XXX company at 10:00 AM tomorrow," then the execution plan P1 generated by the cloud-based agent could be "Attend a meeting on December 30th at 10:00 AM; Transportation: By car; Location: Building A." The mode of transportation, "by car," can be determined by the cloud-based agent based on travel preferences in user information D1, and "Building A" can be determined by the cloud-based agent based on the address of XXX company in the global knowledge base.

[0068] Next, in step S203, the cloud agent sends the above execution plan P1 to the edge agent deployed in the edge device Edge1 associated with user U1.

[0069] As mentioned earlier, the cloud server can record the association relationships between each user and the edge device. Based on these relationships, the cloud agent can determine the edge device Edge1 associated with user U1, and then send the aforementioned execution plan P1 for user U1 to the edge device Edge1, inputting it into the edge agent deployed on Edge1. In practice, the data sent along with the plan also includes the user U1's identifier (i.e., UID) and optional context related to the aforementioned execution plan P1, such as other traffic suggestions.

[0070] After receiving the execution plan P1, in step S204, the edge agent deployed in the edge device Edge1 analyzes and decomposes the execution plan P1 to generate several sub-execution plans applicable to several edge agents.

[0071] Specifically, the edge agent can determine the user's status information based on the status data of various terminal devices associated with user U1. For example, based on data such as the network connection status and application usage status of each terminal device, it can be determined that user U1 is currently working at a specific location. The edge agent can then combine this user status information to analyze and decompose the coarse-grained execution plan P1, generating specific sub-execution plans suitable for execution by each edge agent. This analysis and decomposition process can be derived through reasoning from the larger model within the edge agent.

[0072] As mentioned earlier, the edge agents deployed in user terminal devices can include browser agents, mobile agents, PC agents, and so on. Correspondingly, the edge agents can generate sub-execution plans suitable for the characteristics of each edge agent. For example, for a mobile agent, the generated sub-execution plan may include navigation tasks and / or motion status monitoring tasks; for a PC agent, the generated sub-execution plan may include document processing tasks; and for a browser agent, the generated sub-execution plan may include information retrieval tasks.

[0073] Continuing the previous example, assume the current execution plan P1 is "Attend a meeting on the morning of December 30th, time: 10:00 AM; transportation: by car; location: Building A". After analysis and decomposition, the edge agent can generate the following sub-execution plans:

[0074] Sub-execution plan p11 for the browser agent: "Query parking availability around the meeting location and the meeting topic";

[0075] Sub-execution plan for PC intelligent agents p12: "Remind and assist users in preparing and reviewing meeting materials";

[0076] The sub-execution plan for the mobile intelligent agent (p13) states: "At 9:00, open the navigation software to monitor traffic conditions. Based on the traffic conditions, remind the user to depart around 9:30. At that time, prepare to receive navigation and monitor the movement status."

[0077] Next, in step S205, the edge agent distributes the aforementioned sub-execution plans to several end-side agents. In one implementation, the edge agent determines the terminal devices used by user U1 based on the association between the user and the terminal devices recorded in the edge device Edge1, and sends each sub-execution plan to the corresponding terminal device, thereby passing it to the end-side agent deployed in the terminal device. In another implementation, the edge agent directly pushes each sub-execution plan to the end-side agent logged in with user U1's identifier through an internal mechanism between multi-level agents.

[0078] Therefore, in step S206, each end-side agent receives the issued sub-execution plan and executes the corresponding sub-execution plan.

[0079] For example, after receiving the aforementioned sub-execution plan p11, the browser agent can automatically search in the background for information such as parking availability near the meeting location and the meeting topic. Alternatively, it can prompt the user in the search box whether they want to immediately search for information about the surrounding area and the meeting topic, and initiate the search after the user confirms.

[0080] After receiving the aforementioned sub-execution plan p12, the PC agent pops up a reminder box: "Please note the need to prepare and check meeting materials."

[0081] After receiving the aforementioned sub-execution plan p13, the mobile intelligent agent checks the status of navigation tools, sets the system to automatically open the navigation software at 9:00 to monitor traffic conditions, and sets an alarm clock reminder at 9:30 to notify the user to depart.

[0082] In this way, the tasks set by the user in the cloud are transformed into specific sub-execution plans that are executed on the user's terminal device through the collaborative operation of various intelligent agents. The sub-execution plans executed in each terminal intelligent agent complement each other, and together they constitute the execution plan of the user's task.

[0083] Furthermore, in one implementation, each end-side agent also reports the execution status of the sub-execution plan to the edge agent, enabling the edge agent to determine and track changes in the user's state.

[0084] Specifically, in step S207, each edge agent obtains the execution status information of the corresponding sub-execution plan and uploads the execution status information to the edge agent.

[0085] Specifically, the browser agent can extract information such as the webpage DOM, browsing history, number of tabs, active tabs, and webpage content to obtain execution status information. Continuing the previous example, the browser agent can use webpage DOM parsing technology to determine that the user is "viewing the surroundings of the meeting location," serving as execution status feedback for sub-execution plan p11.

[0086] The PC agent can extract one or more of the following to obtain execution status information: running processes, clipboard, file system, application running status, etc. Continuing the previous example, the PC agent can sense whether the user has clicked to confirm the notification box through the running processes, serving as execution status feedback for sub-execution plan p12.

[0087] Mobile intelligent agents can extract GPS data, chat history, device status, network connection, application running status, etc., from mobile devices to obtain execution status information. Continuing the previous example, the mobile intelligent agent can determine the following execution status information through GPS data and application running status: the user has been notified of the alarm reminder, the navigation tool is working normally, and the user is currently still at the work location and waiting to depart. This execution status information can be used as the execution status feedback for sub-execution plan p13 and reported to the edge intelligent agent.

[0088] It should be noted that although the above descriptions of execution status information are in text form, this is only for ease of understanding; in reality, execution status information can have various data formats depending on the object and content being extracted.

[0089] According to one implementation, to protect user privacy, the edge agent performs an anonymization processing on the execution status information before uploading it to the edge agent. This anonymization primarily involves detecting whether the execution status information contains specific user privacy-related information, such as phone numbers, ID card numbers, detailed addresses, etc. If such information is found, it can be obfuscated in various ways, such as removing parts of the content or replacing specific numbers with "Number 1," "Number 2," etc. Various existing anonymization techniques can be used to achieve information anonymization, and the specific anonymization method is not limited.

[0090] Next, in step S208, the edge agent performs fusion analysis on the execution status information reported by each end agent to determine the intention state of user U1.

[0091] It is understandable that since the edge agents come from different terminal devices, the data format and content of the execution status information they report may differ significantly. For example, the data reported by a browser agent may include images, the data reported by a PC agent may include text and structured data, and the data reported by a mobile agent may include numerical values ​​such as time and GPS location. Therefore, under typical circumstances, the edge agents receive heterogeneous data with different formats from the various edge agents.

[0092] Therefore, the preferred large model for edge agents is a multimodal large model, which can directly process heterogeneous data from multiple modalities and align such data to the same semantic space. Thus, the edge agent can semantically align and fuse heterogeneous execution state information from various edge agents to determine the user's intent state.

[0093] Continuing from the previous example, if the execution status information received by the edge agent includes: the user is viewing the surroundings of the meeting location, the user has clicked to confirm the reminder box, the user has been informed of the alarm reminder, the navigation tool is working normally, and the user is still at the work location waiting to depart (the above items can have different data formats), the edge agent can determine the intention state of user U1 after fusing this information: the user has been informed of and accepted the execution plan P1.

[0094] The above steps S207-S208 can be executed repeatedly. For example, in one embodiment, steps S207-S208 can be executed at certain time intervals. Alternatively, in another embodiment, when any edge agent detects a change and update in the execution state information, it uploads the updated execution state information, and the edge agent determines the update intention state of user U1 based on the updated execution state information.

[0095] When the edge agent detects a significant change in the intent state of user U1 and needs to report it to the cloud agent, in step S209, the edge agent uploads the intent state data corresponding to the intent state of user U1 to the cloud agent.

[0096] For example, during a certain execution step S208, the edge agent determines that the user U1's intention state is: the user has taken a taxi to the meeting. This intention state involves a change in a key node of task execution, so the edge agent uploads the corresponding intention state data to the cloud agent.

[0097] In one embodiment, to better protect user privacy, the edge agent extracts the schedule state portion from the user U1's intent state, abstracts the portion involving the user's personal data, determines its vectorized representation, and uploads the schedule state portion and the vectorized representation of the user's personal data to the cloud agent.

[0098] In step S210, the cloud-based agent updates user information D1 if it determines that user U1's intent state data is normal. In this step, the cloud-based agent first determines whether user U1 has deviated from the execution trajectory of execution plan P1 based on user U1's intent state data. If there is no deviation, it determines that the intent state data is normal and updates user information D1 accordingly.

[0099] As mentioned earlier, user information D1 for user U1 may include user U1's schedule / status information and user profile information. User profile information is generally stored as semantic embedding vectors. Accordingly, after the cloud agent receives the intent state data reported by the edge agent, it can update the schedule / status information in user information D1 based on the schedule / status portion, and update the user profile information based on the vectorized representation. Updates to the user profile information can be used to subsequently optimize future strategy formulation.

[0100] When user U1 executes according to the predetermined execution plan P1, the cloud-based intelligent agent can finally determine that task T1 has been completed after updating user information D1 several times, and thus mark task T1 as completed in the schedule / status information.

[0101] In the above process, each edge agent executes its corresponding sub-execution plan on the user's terminal device and extracts and reports execution status information. The edge agent, based on the execution status information from each edge agent, determines and tracks the user's intent state and, at necessary nodes, reports the user's intent state to the cloud agent. Thus, through the collaboration of agents at various levels, the system assists the user in completing the set task.

[0102] The following describes the anomaly reporting process initiated by the edge agent.

[0103] Figure 3 This diagram illustrates a task processing scenario where an abnormal execution condition exists in one embodiment. For example... Figure 3 As shown, in step S301, any edge agent detects that the current execution state deviates from the issued sub-execution plan, and the edge agent obtains abnormal execution information. Similarly, this abnormal execution information can be anonymized execution state information.

[0104] In step S302, the edge agent makes a decision based on the reported abnormal execution information. If it determines that the abnormal state can be resolved locally, the edge agent rearranges the sub-execution plan. In other words, the edge agent generates several updated sub-plans based on the reported abnormal execution information.

[0105] In step S303, the edge agent distributes several update sub-plans to several end-side agents. In step S304, each end-side agent executes its corresponding update sub-plan.

[0106] In a specific example, suppose the PC agent, while executing sub-execution plan p12 to help the user review meeting materials, detects a malfunction in the software tool used to open a particular meeting document through application runtime status monitoring. Therefore, the user cannot use this software tool to review the meeting materials. The PC agent can then report this execution anomaly to the edge agent.

[0107] The edge agent, by summarizing and analyzing the device status data of the terminal devices previously reported by various edge agents, discovered that the aforementioned software tool existed in a mobile version and that the mobile version was functioning normally. Therefore, the edge agent determined that the abnormal state could be resolved locally and generated several update sub-plans. The update sub-plan for the PC agent included sending the aforementioned meeting materials to the mobile device; the update sub-plan for the mobile agent included assisting the user in using the mobile version of the aforementioned software tool to review the meeting materials.

[0108] By resolving minor anomalies locally through edge agents, we can avoid uploading large amounts of data to the cloud and making decisions in the cloud, thereby reducing decision-making delays and improving decision-making efficiency.

[0109] Figure 4 This diagram illustrates a task processing scenario where abnormal execution occurs in another embodiment. For example... Figure 4 As shown, in step S401, any edge agent detects that the current execution state deviates from the issued sub-execution plan, and the edge agent obtains abnormal execution information. Similarly, this abnormal execution information can be anonymized execution state information.

[0110] In step S402, the edge agent makes a decision based on the reported abnormal execution information. Specifically, the edge agent integrates the execution status information of each end-side agent, including the reported abnormal execution information, to determine the abnormal state of user U1. If it is determined that the abnormal state cannot be resolved locally, the edge agent initiates anomaly reporting.

[0111] Therefore, in step S403, the edge agent reports the abnormal data corresponding to the user's abnormal state to the cloud agent. Similar to the previous reporting of the user's intent state, the edge agent can vectorize the portion involving the user's personal data to protect user privacy.

[0112] In step S404, the cloud agent updates the execution plan based on the reported abnormal data of user U1. In this step, the cloud agent inputs the abnormal data of user U1 and the context data of task T1 into the LLM for further decision-making and generates an updated execution plan P2.

[0113] In step S405, the cloud agent sends the updated execution plan P2 to the edge agent. The edge agent then analyzes, breaks down, and sends the execution plan P2 again; this process is similar to... Figure 2 The process after step S204 is similar and will not be repeated here.

[0114] Continuing with the previous example, suppose the mobile agent automatically opens its navigation software at 9:00 AM on the day of the meeting to monitor traffic conditions. It detects a "major traffic accident" on the only route to the meeting location, with real-time traffic showing the road completely blocked and an estimated clearing time of 3 hours. The mobile agent, through reasoning and calculation, concludes that the current physical path is unreachable, and the current execution state deviates from the original plan. Therefore, the mobile agent reports this abnormal execution information to the edge agent.

[0115] The edge agent first makes a decision to assess whether the anomaly can be resolved locally. Specifically, the LLM in the edge agent can perform comprehensive reasoning based on the current execution state information, conducting multi-path searches, for example, attempting to find alternative routes. However, the result of the multi-path search is that, due to the accidental blockade, the ETA (Estimated Time of Arrival) of all alternative paths is significantly later than the meeting start time. Therefore, the edge agent decides that the anomaly cannot be resolved locally.

[0116] Therefore, in step S403, the edge agent reports the abnormal data corresponding to the user's abnormal state to the cloud agent. Optionally, the edge agent may also report alternative execution plan proposals to the cloud agent for decision-making. For example, an alternative execution plan proposal could be to change an offline meeting to an online meeting.

[0117] After receiving reported abnormal data, the cloud-based agent infers an updated execution plan P2. During this inference process, the agent can reference not only the task T1 context based on user U1 and user information D1, but also leverage the global knowledge base to obtain more comprehensive information, such as the status of other participants in the aforementioned meeting. Therefore, the cloud-based agent can generate an updated execution plan P2, for example, "Create a video conference link and invite personnel from XXX company to join."

[0118] The cloud-based agent sends the updated execution plan P2 to the edge agent. The edge agent analyzes and breaks it down according to the aforementioned steps, so that each edge agent executes the update sub-plan sent by the edge agent to assist the user in completing the task T1.

[0119] Through the above process, when a major anomaly occurs during task execution, the edge agent and the terminal agent report the anomaly level by level to the cloud agent. The cloud agent, leveraging a global knowledge base, can then more accurately determine an alternative execution plan. Thus, through the collaborative operation of multiple layers of agents, the user's task execution status is continuously monitored, anomalies are eliminated, and comprehensive assistance is provided to the user in completing the intended task.

[0120] review Figures 2 to 4 As can be seen from the execution process, in the embodiments of this specification, through the collaborative cooperation of three levels of intelligent agents—end-side intelligent agents, edge intelligent agents, and cloud intelligent agents—a deep understanding and seamless connection of user intent can be achieved across applications and devices, comprehensively assisting users in completing task processing.

[0121] According to another embodiment, a task processing system is also provided, including: a cloud server, a first edge device, and at least one terminal device;

[0122] The cloud server is equipped with a first intelligent agent, which is used to receive a first task created by a first user and generate a first execution plan corresponding to the first task based on the first user information corresponding to the first user; it is also used to send the first execution plan to a second intelligent agent deployed in the first edge device.

[0123] The second intelligent agent deployed in the first edge device is used to analyze and decompose the first execution plan, generate several sub-execution plans applicable to several third intelligent agents, and distribute the several sub-execution plans to the several third intelligent agents; the several third intelligent agents are deployed in the at least one terminal device;

[0124] Each third agent deployed in the at least one terminal device is used to execute the corresponding sub-execution plan.

[0125] According to another embodiment, apparatus for implementing the execution logic of cloud-based intelligent agents, edge intelligent agents, and endpoint intelligent agents is also provided. These apparatuses include corresponding modules that implement the functions of the corresponding intelligent agents.

[0126] According to another embodiment, a computer system is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement an execution method / process of any one of a cloud-based intelligent agent, an edge intelligent agent, and an end-side intelligent agent.

[0127] This specification also provides a computer-readable non-volatile storage medium storing a computer program that, when executed by a processor, can be used to perform one or more steps of one or more methods described or illustrated herein, or to provide the functionality described or illustrated herein. In this document, a computer-readable non-volatile storage medium or medium may include one or more semiconductor or other integrated circuits (ICs) (e.g., field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs)), hard disk drives (HDDs), hybrid hard disk drives (HHDs), optical disks, optical disk drives (ODDs), magneto-optical disks, magneto-optical drives, floppy disks, floppy disk drives (FDDs), magnetic tape, solid-state drives (SSDs), RAM drives, secure digital cards or drives, any other suitable computer-readable non-volatile storage medium, or any suitable combination thereof where appropriate. A computer-readable non-volatile storage medium may be volatile, non-volatile, or a combination of volatile and non-volatile.

[0128] It should also be noted that 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 limitation, 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.

[0129] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0130] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this application.

Claims

1. A task processing method, executed by a first intelligent agent deployed in a cloud server, the method comprising: Receive the first task created by the first user; Based on the first user information corresponding to the first user, generate the first execution plan corresponding to the first task; The first execution plan is sent to the second intelligent agent deployed in the first edge device associated with the first user, so as to trigger the second intelligent agent to decompose the first execution plan into several sub-execution plans suitable for execution by the terminal device.

2. The method of claim 1, wherein, The first task is created through the interface application corresponding to the first intelligent agent.

3. The method of claim 1, wherein, The first user information includes the first profile information of the first user, which is stored in the cloud server in the form of a semantic embedding vector.

4. The method of claim 1, wherein, Generate the first execution plan corresponding to the first task, including: The first execution plan is generated based on the first user information and the global knowledge base stored in the cloud server.

5. The method according to claim 1, further comprising: Receive intent state data representing the intent state of the first user from the second intelligent agent; The intent state is determined by the second agent through fusion analysis of the execution state information of the several sub-execution plans; Update the first user information based on the intent state data.

6. The method according to claim 1, further comprising: Receive first abnormal data from the second intelligent agent, wherein the first abnormal data represents the first abnormal state of the first user; The first abnormal state is determined by the second intelligent agent based on the abnormal execution information of the plurality of sub-execution plans; Based on the first abnormal data and the first user information, a second execution plan is generated; The second execution plan is sent to the second intelligent agent to trigger the second intelligent agent to decompose the second execution plan into several update sub-plans suitable for execution by the terminal device.

7. A task processing method, executed by a second intelligent agent deployed in an edge device, the method comprising: Receive the first execution plan for the first user issued by the first intelligent agent deployed in the cloud server; The first execution plan is analyzed and decomposed to generate several sub-execution plans applicable to several third intelligent agents; The plurality of sub-execution plans are distributed to the plurality of third intelligent agents, which are deployed in at least one terminal device of the first user.

8. The method of claim 7, wherein, The edge device is one of the following: a local area network (LAN) gateway device, a router, or a LAN server.

9. The method of claim 7, wherein, The plurality of third agents includes one or more of the following: a browser agent deployed in a browser, a mobile agent deployed in a mobile device, and a PC agent deployed in a personal computer.

10. The method of claim 9, wherein, The sub-execution plans include one or more of the following: Navigation tasks and / or mobility status monitoring tasks suitable for mobile intelligent agents to perform; Document processing tasks suitable for PC-based intelligent agents; Suitable for information retrieval tasks performed by browser agents.

11. The method according to claim 7, wherein, The first execution plan is analyzed and decomposed to generate several sub-execution plans for several third agents, including: Based on the first user's first status information, the plurality of sub-execution plans are generated; the first status information is determined based on the status data of the at least one terminal device.

12. The method of claim 7, further comprising: Receive corresponding execution status information from the plurality of third intelligent agents; By integrating and analyzing the execution status information, the intent status of the first user can be determined.

13. The method of claim 12, further comprising: The intent state data corresponding to the intent state is uploaded to the first intelligent agent.

14. The method of claim 13, wherein, The execution status information is anonymized information, and the intent status data includes user personal data represented by vectorization.

15. The method of claim 7, further comprising: The first target intelligent agent among the plurality of third intelligent agents receives the first abnormal execution information reported by the first target intelligent agent; Based on the first abnormal execution information, several update sub-plans are generated; The aforementioned update sub-plans are distributed to the aforementioned third-party agents.

16. The method of claim 7, further comprising: The second target agent receives the reported second abnormal execution information from among the plurality of third agents; Based on the second abnormal execution information, the first abnormal state of the first user is determined; The first abnormal data corresponding to the first abnormal state is reported to the first intelligent agent, so that it updates the execution plan.

17. The method of claim 16, wherein, The second abnormal execution information is de-identified information, and the first abnormal data is a vectorized representation of the first abnormal state.

18. A task processing method, executed by a target intelligent agent deployed in a terminal device, the method comprising: Receive the first sub-execution plan issued by the second intelligent agent deployed in the edge device; The first sub-execution plan is the sub-execution plan corresponding to the target intelligent agent among several sub-execution plans generated by the second intelligent agent; the several sub-execution plans are generated by decomposing the execution plan sent from the cloud; Execute the first sub-execution plan.

19. The method of claim 18, further comprising: Obtain the execution status information of the first sub-execution plan; After the execution status information is anonymized, it is uploaded to the second intelligent agent.

20. The method according to claim 19, wherein, The target agent is a browser agent deployed in a browser; obtaining the execution status information of the first sub-execution plan includes: obtaining the execution status information by extracting one or more of the following: webpage DOM, browsing history, number of tabs, active tabs, and webpage content; or, The target intelligent agent is a mobile intelligent agent deployed in a mobile device; obtaining the execution status information of the first sub-execution plan includes: obtaining the execution status information by extracting one or more of the following: GPS, chat history, device status, network connection, application running status; or, The target intelligent agent is a PC intelligent agent deployed in a personal computer; obtaining the execution status information of the first sub-execution plan includes: obtaining the execution status information by extracting one or more of the following: running process, clipboard, file system, application running status.

21. The method of claim 18, further comprising: In response to the current execution state deviating from the first sub-execution plan, obtain abnormal execution information; The abnormal execution information is anonymized and then uploaded to the second intelligent agent.

22. A task processing method, comprising: The first intelligent agent deployed in the cloud server receives the first task created by the first user, and generates the first execution plan corresponding to the first task based on the first user information corresponding to the first user. The first intelligent agent sends the first execution plan to the second intelligent agent deployed in the first edge device associated with the first user; The second intelligent agent analyzes and decomposes the first execution plan to generate several sub-execution plans applicable to several third intelligent agents, and distributes the several sub-execution plans to the several third intelligent agents, which are deployed in at least one terminal device of the first user; Each third agent executes its corresponding sub-execution plan.

23. A task processing system comprising: Cloud server, first edge device, at least one terminal device; The cloud server is equipped with a first intelligent agent, which is used to receive a first task created by a first user and generate a first execution plan corresponding to the first task based on the first user information corresponding to the first user. It is also used to distribute the first execution plan to the second intelligent agent deployed in the first edge device; The second intelligent agent deployed in the first edge device is used to analyze and decompose the first execution plan, generate several sub-execution plans applicable to several third intelligent agents, and distribute the several sub-execution plans to the several third intelligent agents; the several third intelligent agents are deployed in the at least one terminal device; Each third agent deployed in the at least one terminal device is used to execute the corresponding sub-execution plan.

24. A computer system comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, When the processor executes the program, it implements the method of any one of claims 1-21.

25. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-21.

Images