Intelligent agent task execution method and system, medium and electronic equipment

Abstract

The embodiment of the invention relates to the technical field of intelligent assistants, and provides an intelligent agent task execution method and system, a medium and electronic equipment. The method comprises the steps that a first end sends a target task to a second end; the second end starts from the unified starting point state of the target task, parallel exploration is carried out through the multiple branch agents, and a target path corresponding to the target task is obtained; the second end sends the target path to the first end; and the first end executes the target task based on the target path by using the main agent. According to the method provided by the invention, the branch agent is utilized to perform path exploration at the second end to obtain the target path, and the main agent executes the target task at the first end, so that the task execution success rate of the main agent can be improved.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a method, system, medium, and electronic device for intelligent agent task execution. Background Technology

[0002] A graphical user interface (GUI) agent (hereinafter referred to as an agent) is an artificial intelligence program that can simulate human interaction with a software interface. It can perform specific tasks by executing operations (such as clicking, swiping, and typing text) on various software interfaces, thereby improving operational efficiency and automation levels. Currently, agents can be applied to scenarios requiring automated operation of graphical user interfaces, such as robotic process automation (RPA), automated software testing, and intelligent personal assistants. Summary of the Invention

[0003] When an intelligent agent performs a task, it can only make decisions based on the current real-time information. If it chooses the wrong path at a certain decision point, the task will fail.

[0004] To address the aforementioned technical problems, this disclosure provides an agent task execution method, system, medium, and electronic device. By utilizing a branch agent at the second end to explore the path to obtain the target path, and having the main agent execute the target task at the first end, the success rate of the main agent's task execution can be improved.

[0005] The first aspect of this disclosure provides an agent task execution method, the method comprising: a first end sending a target task to a second end; the second end starting from a unified starting state of the target task and using multiple branch agents to explore in parallel to obtain a target path corresponding to the target task; the second end sending the target path to the first end; and the first end using a main agent to execute the target task based on the target path.

[0006] A second aspect of this disclosure provides an intelligent agent task execution system, comprising: a main intelligent agent deployed in a first terminal for sending a target task to a second terminal; multiple branch intelligent agents deployed in the second terminal for parallel exploration starting from a unified starting point state of the target task to obtain the target path corresponding to the target task; a sending unit deployed in the second terminal for sending the target path to the first terminal; and the main intelligent agent is further configured to execute the target task based on the target path.

[0007] A third aspect of this disclosure is that embodiments of this disclosure provide a computer-readable storage medium storing a computer program for performing the intelligent agent task execution method provided in the first aspect.

[0008] In a fourth aspect of this disclosure, embodiments of this disclosure provide an electronic device comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to read executable instructions from the memory and execute the executable instructions to implement the intelligent agent task execution method provided in the first aspect.

[0009] Based on the agent task execution method provided in this disclosure, when the first end needs to utilize the main agent to execute the target task, multiple branch agents in the second end first explore in parallel to obtain the target path, and then the main agent in the first end executes the target task based on the target path. Since the target path is obtained by the branch agents in parallel exploration in the second end, the risk of single-point decision-making errors by the agent can be reduced, and the success rate of task execution can be improved. On this basis, the parallel exploration by multiple branch agents in the second end can evaluate and select the possible execution paths obtained by the multiple branch agents, thereby solving the balance problem between "exploration" (trying unknown operation sequences) and "utilization" (using known effective paths), and obtaining a better solution as the target path. Attached Figure Description

[0010] The above and other objects, features, and advantages of this disclosure will become more apparent from the more detailed description of the embodiments thereof in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this disclosure and form part of the specification. They are used together with the embodiments of this disclosure to explain the disclosure and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0011] Figure 1 This is a schematic diagram of an intelligent agent task execution system provided in an exemplary embodiment of this disclosure.

[0012] Figure 2 This is a flowchart illustrating an exemplary embodiment of the present disclosure of an agent task execution method.

[0013] Figure 3 This is a flowchart illustrating an intelligent agent task execution method provided in another exemplary embodiment of this disclosure.

[0014] Figure 4 This is a flowchart illustrating an intelligent agent task execution method provided in yet another exemplary embodiment of this disclosure.

[0015] Figure 5 This is a flowchart illustrating the intelligent agent task execution method provided in the fourth exemplary embodiment of this disclosure.

[0016] Figure 6 This is a flowchart illustrating the intelligent agent task execution method provided in the fifth exemplary embodiment of this disclosure.

[0017] Figure 7 This is a flowchart illustrating the intelligent agent task execution method provided in the sixth exemplary embodiment of this disclosure.

[0018] Figure 8 This is a flowchart illustrating the intelligent agent task execution method provided in the seventh exemplary embodiment of this disclosure.

[0019] Figure 9 This is a flowchart illustrating the intelligent agent task execution method provided in the eighth exemplary embodiment of this disclosure.

[0020] Figure 10 This is a flowchart illustrating the intelligent agent task execution method provided in the ninth exemplary embodiment of this disclosure.

[0021] Figure 11 This is a flowchart illustrating the intelligent agent task execution method provided in the tenth exemplary embodiment of this disclosure.

[0022] Figure 12 This is a flowchart illustrating the intelligent agent task execution method provided in the eleventh exemplary embodiment of this disclosure.

[0023] Figure 13 This is a schematic diagram of the structure of an intelligent agent task execution system provided in an exemplary embodiment of this disclosure.

[0024] Figure 14 This is a schematic diagram of the structure of an intelligent agent task execution system provided in another exemplary embodiment of this disclosure.

[0025] Figure 15 This is a structural diagram of an electronic device provided by an exemplary embodiment of the present disclosure. Detailed Implementation

[0026] Hereinafter, exemplary embodiments according to the present disclosure will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present disclosure, and not all embodiments of the present disclosure, and it should be understood that the present disclosure is not limited to the exemplary embodiments described herein.

[0027] It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of this disclosure.

[0028] Application Overview A graphical user interface (GUI) agent (hereinafter referred to as an agent) is an artificial intelligence program that can simulate human interaction with a software interface. It can perform specific tasks by executing operations (such as clicking, swiping, and typing text) on various software interfaces, thereby improving operational efficiency and automation levels. Currently, agents can be applied to scenarios requiring automated operation of graphical user interfaces, such as robotic process automation (RPA), automated software testing, and intelligent personal assistants.

[0029] When performing tasks, intelligent agents typically suffer from the following drawbacks: First, they can only make decisions based on current real-time information, meaning they can only try one operation path at any given time. If an incorrect path is chosen at a decision point (e.g., clicking the wrong button or entering an irrelevant menu), the agent may get stuck in an infinite loop or spend a significant amount of time backtracking and retrying, drastically reducing task efficiency and success rate. Second, in complex GUI environments, agents need to balance "exploration" (trying unknown operation sequences) and "exploitation" (using known effective paths). Over-exploration can lead to inefficiency, while over-exploitation may miss better solutions, making it difficult to achieve a balance. Finally, for long-running tasks requiring dozens or even hundreds of steps, a single incorrect decision can cause the entire task chain to fail. As task length increases, the probability of failure grows exponentially, making it difficult for agents to reliably complete complex workflows.

[0030] Therefore, there is an urgent need for a method that can ensure that intelligent agents can perform tasks accurately and reliably in complex and dynamic GUI environments, so as to improve the success rate of intelligent agents in task execution.

[0031] This disclosure proposes a novel solution to address the low success rate of agent task execution in complex, dynamic GUI environments. Specifically, it provides an agent task execution method. When a target task needs to be executed by a main agent at the first end, multiple branch agents at the second end explore in parallel to obtain the target path. Then, the main agent at the first end executes the target task based on the target path. Since the target path is obtained by the branch agents in parallel exploration at the second end, the risk of single-point decision-making errors by the agent is reduced, improving the success rate of task execution. Furthermore, the parallel exploration by multiple branch agents at the second end allows for the evaluation and selection of possible execution paths obtained by these branch agents, thereby resolving the balance between "exploration" (trying unknown operation sequences) and "utilization" (using known effective paths), resulting in a better solution as the target path.

[0032] Exemplary System Figure 1This is a schematic diagram of an intelligent agent task execution system provided in an exemplary embodiment of this disclosure.

[0033] like Figure 1 As shown, the intelligent agent task execution system provided in this disclosure includes a main intelligent agent 101 and branch intelligent agents 201, which are deployed at the first end 10 and the second end 20, respectively.

[0034] First, combine Figure 1 (1) exemplarily describes the intelligent agent task execution system provided in this disclosure from a hardware perspective.

[0035] At the hardware level, the first terminal 10 can be a device with a graphical user interface, such as an in-vehicle terminal, or a smartphone, tablet, laptop, etc.

[0036] The first end 10 is capable of running a lightweight main agent and allows the main agent to perform screen capture and simulated input operations. The second end 10 may include a processor, memory, and RAM (not shown in the figure).

[0037] The first end 10 may be equipped with a first communication module 102, including but not limited to: a wireless network module (such as a Wi-Fi module), a mobile communication module (such as a 4G / 5G module), and an Ethernet module. In this way, the first end 10 has network connectivity and can transmit data with the second end 20, such as sending target task-related information to the second end 20.

[0038] The second end 20 can be a device with high computing power, such as a server or workstation. For example, the second end 20 can be a cloud server.

[0039] The second end 20 may include a processor 202, a memory 203, and a memory 204.

[0040] The processor 202 can load and execute program instructions to control the agents to explore methods. In one example, the processor 202 may include a central processing unit (CPU), and the CPU may be a multi-core server-grade CPU, with the number of cores being proportional to the number of branch agents 201 running in parallel; the processor 202 may also include a graphics processing unit (GPU) to support the operation and inference of the inference model, as well as to support the parallel rendering of the virtual environment corresponding to multiple branch agents 201.

[0041] The memory 203 can be used to store processor-executable program instructions. The processor 202 can load and execute the program instructions in the memory to implement the method of the second terminal 20. The memory 203 can also be used to store data, such as environmental information of the virtual environment corresponding to the branch agent 201. The environmental information of the virtual environment refers to parameter information and state information used to describe the operating environment of the virtual environment. In one example, the memory 203 can be a high-speed solid-state drive (SSD) with a capacity of not less than 2TB.

[0042] Memory 204 can be used to provide high-speed data access space for the processor. In one example, memory 204 can be no less than 128GB to support the operation of multiple virtual environments and inference models.

[0043] The second end 20 is equipped with a second communication module 205, including but not limited to a wired network interface, such as a gigabit or 10-gigabit Ethernet interface. Thus, the second end 20 has network connectivity and can transmit data with the first end 10; for example, the sending module of the second end 20 can send target path related information to the first end 10.

[0044] Then combine Figure 1 (2) in this disclosure exemplarily describes the intelligent agent task execution system provided in this disclosure from a software perspective.

[0045] At the software architecture level, the first terminal 10 and the second terminal 20 provided in this disclosure can be deployed in environments such as Linux, Windows, macOS, Android, and iOS. The first terminal 10 and the second terminal 20 can be deployed in the same operating system environment or different operating system environments. For example, the first terminal 10 can be deployed in an Android environment, and the second terminal 20 can be deployed in a Linux environment.

[0046] In one example, the main agent 101 may be installed in the operating system environment deployed on the first end 10.

[0047] In one example, the second end 20 can be configured with a virtualization or container environment, including but not limited to Docker, Kubernetes, and KVM. In this way, the second end 20 can create independent virtual machines or containers for each branch agent 201 through virtualization or container technology, serving as a running sandbox for the branch agent 201.

[0048] In one example, the second end 20 can be configured with an environment cloning module 202, including but not limited to AndroidEmulator, Anbox, Xcode Simulator, or Scrcpy. Thus, the second end 20 can use the environment cloning module 202 to create virtual environments 203 for each branch agent within a runtime sandbox created through virtualization or container technology. For example, an Android simulation environment with an operating system, screen, and applications can be created in a virtual machine, allowing each branch agent to explore paths within the virtual environment 203.

[0049] In one example, the second end 20 can be equipped with a deep learning framework, such as PyTorch, JAX, TensorFlow, etc., for the operation and inference of the inference model.

[0050] In one example, the second end 20 can deploy an inference model 204, including but not limited to Transformer-based visual language models (such as GPT, CogVLM, etc.), object detection models, etc. In addition, the second end 20 can also integrate optional reinforcement learning libraries. In this way, each branch agent 201 can call the inference model 204 by calling the model interface, etc. The inference model 204 performs inference based on a deep learning framework to achieve parallel exploration of each branch agent 201.

[0051] In one example, the second end 20 can be configured with a communication protocol module (not shown in the figure), including but not limited to gRPC, WebSocket, or RESTful API. This allows data transmission between the first end 10 and the second end 20 via the communication protocol module.

[0052] In one example, the second end 20 can also be configured with a database (not shown in the figure), including but not limited to Redis, Neo4j, etc. In this way, the second end 20 can store the exploration history of the branch agent, evaluation information of explored paths, etc.

[0053] Exemplary methods Figure 2 This is a flowchart illustrating an exemplary embodiment of the present disclosure of an agent task execution method. Figure 2 As shown, the method includes the following steps: Step 100: The first end sends the target task to the second end.

[0054] In step 100, after determining that a target task needs to be executed, the first end can send the target task to the second end so that the second end can explore based on the target task and obtain the corresponding target path.

[0055] The following example uses the first end as the terminal and the second end as the cloud. For instance, when the terminal receives a user's input request to "order a cup of milk tea for delivery", it determines the target task "order a cup of milk tea with 70% sugar for delivery" based on the user's historical preference information, and sends the relevant information of the target task, such as text description information, to the cloud.

[0056] Step 200: The second end starts from the unified starting state of the target task and uses multiple branch agents to explore in parallel to obtain the target path corresponding to the target task.

[0057] In step 200, the second end starts from the unified starting state of the target task and explores in parallel using multiple branch agents. The unified starting state of the target task refers to information representing the initial state of the target task when it starts, such as the application's running state and interface display state. For example, for the target task "order a cup of milk tea delivery" in the previous example, the unified starting state could be "device displays the main screen interface, delivery app not started," etc.

[0058] The second end can comprehensively evaluate the exploration results of each branch agent to obtain the final target path. The target path is the sequence of operations from the task start point to the task end point. For example, for the target task "order a cup of milk tea" in the previous example, the target path P_best could be: "Open the food delivery app -> Click the search box -> Enter 'XX milk tea' -> Click the first search result -> Select '70% sugar' -> Click 'Buy Now' -> Click 'Submit Order' -> Click 'Pay'".

[0059] In some examples, the second end can select or merge the possible paths explored by multiple branch agents to obtain the final target path.

[0060] In some examples, the number of branch agents can be determined based on the complexity of the target task. For simple tasks, a small number of branch agents can be used for parallel exploration; for complex tasks, the number of branch agents can be increased to expand the breadth of exploration.

[0061] Step 300: The second end sends the target path to the first end.

[0062] In step 300, after determining the target path, the second end sends it to the first end so that the first end can follow the target path to complete the target task.

[0063] Step 400: The first end uses the main intelligent agent to execute the target task based on the target path.

[0064] In step 400, the main agent can parse and replay the target path, and execute the target task in the first end according to the target path.

[0065] For example, for the target task "order a cup of milk tea with 70% sugar" in the example above, the main intelligent agent can sequentially execute each click, input, and other operation in the target path P_best on the terminal to finally complete the takeout order.

[0066] As can be seen from the above technical solutions, the method provided in this disclosure, after the target task is triggered at the first end, the second end explores in parallel through multiple branch agents, evaluates the exploration results of each branch agent, obtains the final target path, and then the main agent at the first end executes each step in the target path sequentially to complete the target task. Compared with the conventional scheme of using the main agent to explore serially, this disclosure can avoid task execution identification due to errors in single path decision-making by using multiple branch agents to explore in parallel, thus improving the task execution success rate. In addition, in this disclosure, the first end only runs a lightweight main agent, thus reducing the resource consumption of the first end and minimizing the impact on the performance of the first end; and completing the path exploration at the second end, which has higher computing power, can make full use of the computing power resources of the second end, improve the path exploration efficiency, and shorten the overall task completion time.

[0067] Figure 3 This is a flowchart illustrating a smart agent task execution method provided in another exemplary embodiment of this disclosure. For example... Figure 3 As shown above, in the above Figure 2 Based on the illustrated embodiment, step 200 may include the following steps: Step 210: The second end determines at least one exploration phase based on the target task.

[0068] In step 210, the second end divides the complete exploration process into multiple exploration stages based on the target task. In this way, the branch agent can explore in stages, and proceed to the next exploration stage only after the previous exploration stage is completed.

[0069] In some examples, the number of exploration phases may be uncertain when dividing the exploration phases. Thus, after the previous exploration phase is completed, the next retrieval phase is defined, until the entire exploration process is completed. The completion marker for the complete exploration process can be the completion of the target task. For example, for the target task "order a 70% sugar milk tea delivery," the completion marker for the exploration process could be the completion of payment, such as the appearance of the message "Payment successful" on the interface.

[0070] Step 220: Based on the unified starting point state of the target task, the second end determines the unified starting point state corresponding to the first exploration phase.

[0071] In step 220, before starting the first exploration phase, the second end first needs to determine the phase start state S_1 corresponding to the first exploration phase so that each branch agent can start parallel exploration from a unified phase start state.

[0072] In one example, the starting state of the first exploration phase can be determined based on the unified starting state of the target task. For instance, if the unified starting state of the target task indicates that the food delivery app is currently running, then the starting state of the first exploration phase will also indicate that the food delivery app is running. Thus, in the first exploration phase, there is no need to explore the steps of "opening the food delivery app".

[0073] In one example, the unified starting state of the target task can be determined as the starting state of the first exploration phase.

[0074] Step 230: For each exploration phase, the second end starts from the unified starting state of the exploration phase, uses multiple branch agents to explore in parallel, obtains the target path segment corresponding to the exploration phase, and determines the unified starting state of the next exploration phase based on the endpoint state reached by the target path segment.

[0075] In step 230, within each exploration phase, multiple branch agents begin parallel exploration from the corresponding phase start state. The phase start state is information representing a unified baseline state at the start of a single exploration phase, such as information indicating the initial application running state or interface display state. For example, in the "Customize Specifications and Add to Cart" phase, the phase start state might include "Entered the milk tea product details page, specification options are at default, sweetness option not selected." Thus, each branch agent can start from this phase start state and explore in parallel the steps required to customize specifications and add to the cart.

[0076] During parallel exploration, each branch agent independently executes several steps. The second end can determine the target path segment p_best corresponding to the exploration phase based on the results of the independent execution of each branch agent. Here, the target path segment refers to the sequence of continuous operation steps formed by starting from the starting state of the phase within the exploration phase. Each target path segment only covers the operations of its corresponding exploration phase, and multiple target path segments can be spliced ​​in sequence to form a complete target path.

[0077] Subsequently, based on the endpoint state reached by the target path segment corresponding to the current exploration phase, the unified starting state of the next exploration phase can be determined. This ensures that the endpoint of the current exploration phase is consistent with the starting point of the next exploration phase, guaranteeing a seamless connection in the overall exploration process and ensuring that the final target path is continuous and executable.

[0078] For example, in the aforementioned case, if the endpoint status of the target path segment corresponding to the current exploration phase indicates that "XX milk tea" has been entered and a query result has been obtained, then the query result will also be obtained in the unified starting point status of the next exploration phase. Thus, in the next exploration phase, there is no need to re-enter "XX milk tea" and perform the query.

[0079] In one example, the endpoint state reached by the target path segment corresponding to the final exploration phase can be used as the endpoint state of the target task.

[0080] Step 240: The second end determines the target path corresponding to the target task based on the target path segments corresponding to each exploration stage.

[0081] In step 240, the second end concatenates the target path segments corresponding to each exploration stage in the order of the exploration stages to form the final target path P_success. In this way, a coherent sequence of operations from the unified starting state to the ending state of the target task can be obtained.

[0082] As can be seen from the above technical solutions, the method provided in this disclosure reduces exploration complexity and thus improves the success rate of exploration by breaking down long-term, complex target tasks into multiple consecutive exploration stages and having multiple agents conduct parallel exploration within each exploration stage. After the exploration is completed, the target path segments corresponding to each exploration stage are spliced ​​together to obtain a complete and executable operation sequence.

[0083] Figure 4 This is a flowchart illustrating a method for executing intelligent agent tasks, provided in yet another exemplary embodiment of this disclosure. For example... Figure 4 As shown above, in the above Figure 3 Based on the illustrated embodiment, step 240 may include the following steps: Step 241: The second end determines whether the completion conditions corresponding to the target task have been met based on the endpoint state reached by the target path segment corresponding to the last exploration stage of the sorting.

[0084] In step 241, considering that the endpoint state reached by the target path segment corresponding to the last exploration stage of sorting can be used as the endpoint state S_best of the target task, the endpoint state of the last exploration stage of sorting can be verified to determine whether the endpoint state reached by the target path segment corresponding to it meets the completion conditions of the target task.

[0085] In some examples, the completion condition for the target task can be the display of preset content on the interface, such as the display of "payment successful" on the interface of the branch agent's runtime environment; in other examples, the completion condition for the target task can also be that an application or process in the branch agent's runtime environment is in a preset state, such as the branch agent's runtime environment has successfully generated and persisted the order ID (orderId) and written it to the local order cache.

[0086] In some examples, after each parallel exploration phase is completed, the endpoint state of that exploration phase can be verified. If it is determined that the completion conditions corresponding to the target task have not been met, the next exploration phase is defined, and parallel exploration is carried out in the next exploration phase; if it is determined that the completion conditions corresponding to the target task have been met, the exploration stops, and the process proceeds to step 242.

[0087] Step 242: In response to the completion condition being met, the second end splices the target path segments corresponding to each exploration stage to obtain the target path corresponding to the target task.

[0088] In step 242, if the completion conditions corresponding to the target task are met, it is considered that the last exploration stage in the sequence can reach the endpoint of the target task's success. Therefore, the second end can concatenate the target path segments p_best corresponding to each exploration stage according to the order of each exploration stage to form the final target path P_success.

[0089] For example, for the aforementioned target task "order a cup of milk tea for delivery", the optimal segments of the "search and select milk tea" stage, the optimal segments of the "customize specifications and add to cart" stage, and the optimal segments of the "fill in address and complete payment" stage can be spliced ​​together to obtain a complete operation sequence from search to successful order placement.

[0090] As can be seen from the above technical solutions, the method provided in this disclosure, by performing endpoint status verification on the final exploration stage of sorting and splicing the target path segments after the verification is passed, can ensure that the final target path can reach the endpoint of the target task, that is, ensure that following the target path can complete the target task.

[0091] Figure 5 This is a flowchart illustrating the intelligent agent task execution method provided in the fourth exemplary embodiment of this disclosure. For example... Figure 5 As shown above, in the above Figure 3 Based on the illustrated embodiment, step 230, "starting from the unified stage starting state corresponding to the exploration stage, the second end utilizes multiple branch agents to explore in parallel and obtain the target path segment corresponding to the exploration stage," includes the following steps: Step 231: The second end utilizes multiple branch agents starting from a unified stage starting state and executes the target task in parallel in a virtual environment set to a unified stage starting state, thereby obtaining candidate path segments corresponding to each branch agent.

[0092] In step 231, during each exploration phase, multiple branch agents explore in parallel for that phase, obtaining candidate path segments corresponding to that phase. Specifically, each branch agent, in its own operating environment, starts from a unified phase starting state S_i and independently executes several steps to complete the target task within the exploration phase. For example, a branch agent j independently executes K steps, thus generating a candidate path segment p_j = {(S_{i+1}, a_1), (S_{i+2}, a_2), ..., (S_{i+K}, a_K)}. Here, a_1 refers to the first step of branch agent j, S_{i+1} refers to the state after executing the first step, and so on.

[0093] In this process, the operating environment of each branch agent is set to a unified phase start state before the start of each exploration phase, so that each branch agent can start its exploration from the phase start state.

[0094] In one example, the second end can be equipped with a memory module to store and manage information related to candidate path segments explored by each branch agent. For example, the information stored in the memory module can be in the form of a graph database (such as Neo4j), where each node represents a state, and the hash value or feature vector of the screenshot of the interface in that state can be used as the node ID; each edge represents an operation step, such as in the aforementioned candidate path segment p_j, where S_{i+1} can be a node and a_1 can be an edge. Alternatively, the information stored in the memory module can be in the form of key-value storage (such as Redis), where a key-value pair can be used to store a state, a composite key-value pair can be used to store an operation step, and an index can be used to indicate the association between states and operation steps.

[0095] Step 232: The second end evaluates each candidate path segment and determines the target path segment from multiple candidate path segments based on the evaluation results.

[0096] In step 232, after each exploration phase is completed, the evaluation and arbitration module at the second end can analyze the candidate path segments of each branch agent, as well as the complete information of their intermediate states, and then evaluate the merits of the candidate path segments submitted by each branch agent. For example, a state value evaluation function V(S) can be pre-designed. The state value evaluation function V(S) refers to a function model used to quantify the potential of a candidate path segment to complete the target task. For the aforementioned candidate path segment p_j, V(S) is used to verify the probability or expected reward of completing the target task starting from its last state S_{i+K}. In this way, the evaluation and arbitration module can determine the candidate path segment with the highest V(S) from the N candidate path segments {p_1, ..., p_n} obtained by the N branch agents as the target path segment p_best for this exploration phase.

[0097] V(S) can be implemented in several ways. For example, a value network can be trained to directly predict the V(S) score; or a large language model can be used for scoring. By providing the large language model with a screenshot of the current running environment of the branch agent, the task description information of the target task, and the execution path of the branch agent, the large language model can be prompted to analyze in a chain-of-thought (CoT) manner and give a success probability score between 0 and 1; or heuristic rules can be used, such as whether the interface of the running environment of the branch agent contains keywords related to the task target (such as a milk tea search list) or specific UI elements (such as the "70% sugar" checkbox being checked).

[0098] In one example, the memory module not only stores information related to candidate path segments, but also stores the V(S) corresponding to each candidate path segment, which serves as the basis for determining the target path segment.

[0099] As can be seen from the above technical solutions, the method provided in this disclosure evaluates multiple candidate path segments generated in parallel at each exploration stage and then selects the target path segment from them. This ensures that the target path segment is the best in that exploration stage, thereby improving the overall quality of the final target path and helping to improve the success rate of task execution.

[0100] Figure 6 This is a flowchart illustrating the intelligent agent task execution method provided in the fifth exemplary embodiment of this disclosure. For example... Figure 6 As shown above, in the above Figure 3 Based on the illustrated embodiment, step 230, "starting from the unified stage starting state corresponding to the exploration stage, the second end explores in parallel using multiple branch agents," includes the following steps: Step 233: The second end starts from the unified starting state corresponding to the exploration phase, and explores in parallel using multiple branch agents based on the exploration strategies corresponding to each branch agent.

[0101] In step 233, at each exploration phase, the agents in the second branch can employ different exploration strategies for parallel exploration to increase the diversity of exploration. For example, some agents can use zero-shot reasoning based on a large language model to generate the next action, while others can use reinforcement learning-based strategies, such as... The -greedy strategy balances the use of known optimal operations with the exploration of new operations.

[0102] As can be seen from the above technical solutions, the method provided by the embodiments of this disclosure can avoid problems such as the target path segment being a local optimum due to a single exploration strategy by configuring different exploration strategies for each branch agent during parallel exploration. Therefore, it is beneficial to improve the success rate and efficiency of task execution in complex GUI environments.

[0103] Figure 7 This is a flowchart illustrating the intelligent agent task execution method provided in the sixth exemplary embodiment of this disclosure. For example... Figure 7 As shown above, in the above Figure 5 Based on the illustrated embodiment, step 231, "execute the target task in parallel to obtain candidate path segments corresponding to each branch agent," may include the following steps: Step 2311: Based on the current state information of the target task and the virtual environment, execute at least one operation step in parallel iteratively until the branch agent reaches the termination condition corresponding to the exploration phase.

[0104] In step 2311, within each exploration phase, the branch agent analyzes the current state information of the virtual environment. For example, based on the gap between the current state and the state required for the target task, it determines the next operational step to be executed based on the corresponding exploration strategy. After the next operational step is executed, it re-analyzes the current state information of the virtual environment and the target task to determine the subsequent operational steps. This process continues until the branch agent meets the termination condition corresponding to the exploration phase. At this point, the branch agent is considered to have completed the exploration of that exploration phase.

[0105] The current state information of the virtual environment refers to comprehensive information used to represent the current running state and interface interaction state of the virtual environment in which the branch agent is located, such as information used to represent the interface layout, control state, and running parameters of the virtual environment.

[0106] In one example, the current state information of the virtual environment may include the interface information of the virtual environment. For example, the branch agent can analyze the current screenshot of the virtual environment's interface and extract key visual information. Advanced visual language models (VLMs), including but not limited to Google Gemini Pro and OpenAI GPT, can be used to directly generate descriptions and understandings of interface elements from images; or classic computer vision models, including but not limited to object detection models such as YOLO and Faster R-CNN, can be used to recognize icons, buttons, and other controls, and then combined with OCR technology to read the text.

[0107] For example, in the "search and select products" stage, a branch agent can analyze the current state information and determine that "XX milk tea" has already been entered into the search box. Therefore, based on the current state information and the target task, it can decide the next step to be performed: click the "search" button.

[0108] Step 2312: The second end determines the candidate path segment corresponding to each branch agent based on at least one operation step performed by each branch agent.

[0109] In step 2312, for each branch agent, the second end concatenates the operation steps it performs in the current exploration phase to obtain the candidate path segment corresponding to that branch agent. For example, the candidate path segment p_j corresponding to branch agent j is p_j = { (S_{i+1}, a_1), (S_{i+2}, a_2), ..., (S_{i+K}, a_K)}. Here, a_1 to a_K refer to the first to the Kth operation steps performed by branch agent j in the current exploration phase.

[0110] As can be seen from the above technical solutions, the method provided in this disclosure iteratively generates and executes the next operation step by analyzing the current state information and target task of the virtual environment. In this way, the branch agent considers both the current state and the target to be completed during exploration, enabling flexible adjustments when the current state changes and ensuring that the exploration direction always points towards completing the target task. Therefore, it is beneficial to shorten exploration time and improve the exploration success rate.

[0111] Figure 8 This is a flowchart illustrating the intelligent agent task execution method provided in the seventh exemplary embodiment of this disclosure. For example... Figure 8 As shown above, in the above Figure 7 Based on the illustrated embodiment, step 2311, "the branch agent reaches the termination condition corresponding to the exploration phase," may include at least one of the following steps 231a and 231b: Step 231a: The branch agent executes a preset number of operation steps during the exploration phase.

[0112] In step 231a, the termination condition for the exploration phase can be: the branch agent executes a preset number of operation steps during the exploration phase. The preset number can be set according to the actual application scenario or historical experience, for example, it can be set to 5. Thus, within an exploration phase, each branch agent is considered to have met the termination condition after executing 5 operation steps, and the current exploration phase is completed.

[0113] It should be noted that although each branch agent executes a preset number of operation steps within an exploration phase, the time required for each branch agent to execute the preset number of operation steps may vary due to different exploration strategies, different interfaces encountered during the exploration process, and other reasons.

[0114] Step 231b: The branch agent performs the target task for a preset duration during the exploration phase.

[0115] In step 231b, the termination condition for the exploration phase can be: the branch agent executes the target task for a preset duration during the exploration phase. The preset duration can be set based on the actual application scenario or historical experience, for example, it can be set to 1 second. Thus, within an exploration phase, each branch agent considers the termination condition met after 1 second of exploration, and completes the current exploration phase.

[0116] It should be noted that although each branch agent executes for a preset duration within an exploration phase, the number of operation steps performed by each branch agent within the preset duration may differ due to different exploration strategies, different interfaces encountered during the exploration process, and other reasons.

[0117] As can be seen from the above technical solutions, the method provided in this disclosure, by setting termination conditions for the exploration phase, can break down the complex exploration process into multiple short exploration phases, which helps to simplify the exploration difficulty, improve exploration efficiency and success rate, and is especially suitable for scenarios with complex target tasks. In addition, termination conditions also help to control the exploration cost of each exploration phase and improve the utilization rate of exploration resources.

[0118] Figure 9 This is a flowchart illustrating the intelligent agent task execution method provided in the eighth exemplary embodiment of this disclosure. For example... Figure 9 As shown above, in the above Figure 2 Based on the illustrated embodiment, the following steps may be included before step 200: Step 500: The first end sends the current device status information to the second end.

[0119] In step 500, to ensure that the operating environment of the branch agent in the second terminal is consistent with the actual state of the first terminal when it begins exploration, the first terminal can synchronize its current device state information to the second terminal before the branch agent begins parallel exploration. Here, the current device state information of the first terminal refers to information used to represent the current operating state of the first terminal, such as information indicating the application currently running on the first terminal, the content currently displayed on the screen, etc.

[0120] For example, after receiving the user's task execution request "order a cup of milk tea for delivery", the main intelligent agent can obtain the current device status information of the first end, such as the screenshot of the first end, the application currently running in the first end, etc., and send it to the second end through the first communication module (such as the Wi-Fi module) in the first end.

[0121] Step 600: The second end determines the unified starting point state of the target task based on the current device status information.

[0122] In step 600, the second end receives and parses the current device status information sent by the first end, and determines the unified starting point state of the target task based on the current device status information, which serves as the common starting point for all branch agents to begin parallel exploration.

[0123] For example, regarding the task request of "ordering a cup of milk tea for delivery," the second end parses the current device status information and determines that the first end is currently displaying the home screen interface and the food delivery app is not running. At this point, the second end can define the unified starting point state of the target task as "the device is on the home screen interface and the food delivery app is not running." In this way, each branch agent starts its exploration from this unified starting point state, for example, finding the food delivery app icon on the home screen interface, opening the app by clicking the icon, and then completing the subsequent operation steps.

[0124] As can be seen from the above technical solutions, the method provided in this disclosure, by sending current device state information from the first end to the second end before the branch agents at the second end begin parallel exploration, ensures that all branch agents start from a unified starting state consistent with the real state of the first end and conduct parallel exploration. In this way, the first operation step in the final target path can be seamlessly connected with the current state of the first end, improving the availability of the target path.

[0125] Figure 10 This is a flowchart illustrating the intelligent agent task execution method provided in the ninth exemplary embodiment of this disclosure. For example... Figure 10 As shown above, in the above Figure 9 Based on the illustrated embodiment, step 500 may include the following steps: Step 510: The first device captures the screenshot information and interface layout information from the current device status information.

[0126] In step 510, the current device status information sent by the first end may include screenshot information and interface layout information. The screenshot information may be image data, such as a screenshot, which can be obtained through the operating system's screenshot interface; the interface layout information may be structured data, such as JSON or XML format, which can be obtained by calling interfaces such as AccessibilityService or Accessibility API.

[0127] For example, the main agent on the first end first takes a screenshot of the screen of the first end, and obtains an image that includes the current interface of the first end (such as the main screen interface); at the same time, the main agent on the first end obtains the interface layout information by calling the Accessibility Service interface. The interface layout information indicates that the main screen interface of the first end contains the icon of the food delivery APP, and indicates the specific coordinates of the icon in the main screen interface.

[0128] Step 520: The first end sends the screenshot information and interface layout information to the second end.

[0129] In step 520, the first end encapsulates the captured screenshot and the parsed interface layout information and sends them to the second end.

[0130] As can be seen from the above technical solutions, the method provided in this embodiment allows the first end to upload screenshots and interface layout information to the second end. In this way, the screenshots can provide a visual reference, and the interface layout information can provide accurate data. The combination of the screenshots and the interface layout information enables the second end to accurately determine the current state of the first end.

[0131] Figure 11 This is a flowchart illustrating the intelligent agent task execution method provided in the tenth exemplary embodiment of this disclosure. Figure 11 As shown above, in the above Figure 5 Based on the illustrated embodiment, the following steps may be included before step 231: Step 234: The second end synchronizes the virtual environment where each branch agent is located to the unified starting state of the exploration phase.

[0132] In step 234, before the start of each exploration phase, the second end synchronizes the virtual environment where each branch agent resides to the unified starting state corresponding to that exploration phase, so that each branch agent can start from the same exploration progress and conduct a new round of parallel exploration. For example, the starting state of the exploration phase can be synchronized by restoring a virtual machine snapshot for the running sandbox established in the second end; reading and writing application Shared Preference data in the virtual environment; resetting data of applications such as food delivery apps through adb commands; or accurately replaying the operation sequence based on the target path segment of the previous phase in each virtual environment.

[0133] Specifically, for the first exploration phase, before the phase begins, the second terminal can synchronize the virtual environments Env_1, ..., Env_n of the N branch agents to a unified starting state corresponding to the target task. For example, it can synchronize to the state where "the device is on the main screen and the food delivery app is not running". In this way, each branch agent can start from the same state as the first terminal and independently explore steps such as opening the food delivery app.

[0134] After each exploration phase is completed, a unified starting state for the next exploration phase is determined, and all virtual environments are synchronized to this unified starting state before the next exploration phase begins. For example, after the "search and select milk tea" phase is completed, if the unified starting state for the next exploration phase ("customize specifications and add to cart" phase) is determined to be "successfully selected milk tea and entered the product details page," then all virtual environments can be synchronized to the state of "successfully selected milk tea and entered the product details page." In this way, during the "customize specifications and add to cart" phase, each branch agent independently explores steps such as selecting "70% sugar" and adding to the cart based on the same product details page.

[0135] As can be seen from the above technical solutions, the method provided in this disclosure, by synchronizing the state of the virtual environment where all branch agents reside before the start of each exploration phase, enables all branch agents to start exploring in parallel from the same state within the same exploration phase. This eliminates state differences caused by inconsistencies in the operation steps of each branch agent in the previous exploration phase, making the candidate path segments formed by all branch agents in the current exploration phase comparable. Furthermore, if a unified starting state for the current exploration phase is determined based on the endpoint state reached by the target path segment in the previous exploration phase, all branch agents in the current exploration phase can start from the optimal state of the previous exploration phase. Therefore, they can inherit the exploration results of the previous exploration phase, avoid continuing exploration on invalid branches, and ensure that the whole progresses in the direction most likely to complete the target task. This is beneficial for improving the task execution success rate and shortening the task completion time.

[0136] Figure 12 This is a flowchart illustrating the intelligent agent task execution method provided in the eleventh exemplary embodiment of this disclosure. Figure 12 As shown above, in the above Figure 5 Based on the illustrated embodiment, the following steps may be included before step 231: Step 235: The first end sends the device operating environment information corresponding to the first end to the second end.

[0137] In step 235, to enable the second end to pre-build a virtual environment consistent with the operating environment of the first end, the first end can first synchronize its device operating environment information to the second end. The device operating environment information of the first end refers to information describing the operating environment of the first end, such as hardware version and configuration. The operating environment information of the first end may include, but is not limited to, the device model of the first end, the installed operating system version, and the name and version number of the food delivery app installed on the first end.

[0138] In some examples, considering that the device operating environment information does not change frequently, the device operating environment information can be sent only when the operating environment is synchronized for the first time, and then uploaded only when the device operating environment changes; if the device operating environment does not change, it will not be uploaded again, and the synchronized device operating environment information can be reused when the second end executes the target task.

[0139] Step 236: Based on the device operating environment information, the second end establishes multiple virtual environments that match the device operating environment of the first end, and deploys at least one branch agent in each virtual environment.

[0140] In step 236, the second end receives the device runtime environment information sent by the first end and establishes multiple virtual environments matching the device runtime environment of the first end using virtualization technology. For example, multiple virtual machines or containers can be created on the second end firstly using technologies such as Docker, Kubernetes, and KVM to serve as runtime sandboxes; then, tools such as Android Emulator, Anbox, or Scrcpy can be used to start multiple instances of the same operating system (such as Android / iOS / desktop operating system) as the first end in the virtual machines or containers on the second end, serving as virtual environments for running each branch of the intelligent agent. Each virtual environment has the same applications and configurations installed as the first end.

[0141] As can be seen from the above technical solutions, the method provided in this disclosure involves the first end sending device operating environment information to the second end, and the second end establishing a virtual environment that matches the device operating environment information. In this way, the target path obtained by the second end through parallel exploration using branch agents is compatible with the actual operating environment of the first end, thus improving the availability of the target path and the success rate of the target task execution.

[0142] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention. For example, the steps described in the above embodiments can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in the present invention can be achieved, the present invention does not limit this.

[0143] Exemplary device The above describes the intelligent agent task execution method provided by the embodiments of this disclosure. It is understood that, in order to implement the various functions of this intelligent agent task execution method, an integrated circuit system may include corresponding hardware and software for implementing the hardware functions.

[0144] Those skilled in the art will readily recognize that the steps of the intelligent agent task execution method described in conjunction with the embodiments of this disclosure can be implemented in hardware or in a combination of software-driven hardware. Whether a function is executed in hardware or software-driven hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0145] Figure 13 This is a schematic diagram of the structure of an intelligent agent task execution system provided in an exemplary embodiment of this disclosure. For example... Figure 13 As shown, in one embodiment, the intelligent agent task execution system 1300 includes: a main intelligent agent 1310, multiple branch intelligent agents 1320, and a sending unit 1330.

[0146] The main intelligent agent 1310 is deployed in the first end and is used to send the target task to the second end; Multiple branch agents 1320 are deployed in the second end to explore in parallel from the unified starting state of the target task and obtain the target path corresponding to the target task. The sending unit 1330, deployed in the second end, is used to send the target path to the first end; The main intelligent agent 1310 is also used to execute target tasks based on target paths.

[0147] In an exemplary embodiment, the intelligent agent task execution system 1300 includes an initialization module (not shown in the figure), deployed in a second terminal, for: determining at least one exploration phase based on the target task; and determining a unified phase start state corresponding to the first exploration phase based on a unified start state of the target task.

[0148] Figure 14 This is a schematic diagram of the structure of an intelligent agent task execution system provided in another exemplary embodiment of this disclosure. Figure 14 Taking the vehicle-mounted system as the first end and the cloud as the second end as an example.

[0149] like Figure 14 As shown, in an exemplary embodiment, multiple branch agents 1320 are used to: start parallel exploration from a unified stage start state corresponding to the exploration stage, obtain the target path segment corresponding to the exploration stage, and determine the unified stage start state corresponding to the next exploration stage based on the end state reached by the target path segment. The intelligent agent task execution system 1300 includes an evaluation and arbitration module 1340, which is deployed in the second terminal and is used to: determine the target path corresponding to the target task based on the target path fragments corresponding to each exploration stage.

[0150] like Figure 14 As shown, in an exemplary embodiment, the evaluation arbitration module 1340 is used to: determine whether the completion condition corresponding to the target task has been met based on the endpoint state reached by the target path segment corresponding to the last exploration stage; and, in response to meeting the completion condition, splice the target path segments corresponding to each exploration stage to obtain the target path corresponding to the target task.

[0151] like Figure 14 As shown, in an exemplary embodiment, multiple branch agents 1320 are used to: start from a unified stage starting state, execute the target task in parallel in a virtual environment set to a unified stage starting state, and obtain candidate path segments corresponding to each branch agent 1320. The evaluation and arbitration module 1340 is used to: evaluate each candidate path segment and determine the target path segment from among the multiple candidate path segments based on the evaluation results.

[0152] In one exemplary embodiment, multiple branch agents 1320 are used to: start from a unified stage starting state corresponding to the exploration stage and explore in parallel based on the exploration strategy corresponding to each branch agent 1320.

[0153] In an exemplary embodiment, multiple branch agents 1320 are used to: perform at least one operation step in parallel iteratively based on the current state information of the target task and the virtual environment until the termination condition corresponding to the exploration phase is reached; The evaluation and arbitration module 1340 is used to: determine the candidate path segments corresponding to each branch agent 1320 based on at least one operation step performed by each branch agent 1320.

[0154] In an exemplary embodiment, the termination conditions corresponding to the exploration phase include: the branch agent 1320 performing a preset number of operation steps during the exploration phase; or, the branch agent 1320 performing the target task during the exploration phase for a preset duration.

[0155] In an exemplary embodiment, the main intelligent agent 1310 is used to: send the current device status information to the second terminal; The initialization module 1340 is used to: determine the unified starting point state of the target task based on the current device status information.

[0156] In an exemplary embodiment, the main intelligent agent 1310 is configured to: capture screenshot information and interface layout information from the current device status information; and send the screenshot information and interface layout information to the second terminal.

[0157] like Figure 14 As shown, in an exemplary embodiment, the agent task execution system 1300 includes an environment cloning module 1350, deployed in the second terminal, for: synchronizing the virtual environment where each branch agent 1320 is located to the unified stage starting state corresponding to the exploration stage.

[0158] like Figure 14 As shown, in an exemplary embodiment, the main intelligent agent 1310 is used to: send the device operating environment information corresponding to the first end to the second end; The environment cloning module 1350 is used to: establish multiple virtual environments that match the operating environment of the first end based on the device operating environment information, and deploy at least one branch agent 1320 in each virtual environment.

[0159] Exemplary electronic devices Figure 15 This is a structural diagram of an electronic device provided in an exemplary embodiment of this disclosure. Figure 15 As shown, the electronic device 1500 includes at least one processor 1510 and a memory 1520.

[0160] The processor 1510 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 1500 to perform desired functions.

[0161] The memory 1520 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 1510 may execute one or more computer program instructions to implement the intelligent agent task execution methods and / or other desired functions of the various embodiments of this disclosure described above.

[0162] In one example, the electronic device 1500 may also include an input device 1530 and an output device 1540, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0163] The input device 1530 may also include, for example, a keyboard, a mouse, etc.

[0164] The output device 1540 can output various information to the outside, including, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0165] Of course, for the sake of simplicity, Figure 15 Only some of the components of the electronic device 1500 relevant to this disclosure are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device 1500 may include any other suitable components depending on the specific application.

[0166] Exemplary computer program products and computer-readable storage media In addition to the methods and apparatus described above, embodiments of this disclosure may also be computer program products comprising computer program instructions that, when executed by a processor, cause the processor to perform the steps of the bandwidth control methods according to various embodiments of this disclosure as described in the "Exemplary Methods" section of this specification.

[0167] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of embodiments of this disclosure. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on a user's computing device, partially on a user's computing device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0168] Furthermore, embodiments of this disclosure may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the bandwidth control methods according to various embodiments of this disclosure as described in the "Exemplary Methods" section above.

[0169] Computer-readable storage media may take the form of any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0170] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0171] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context explicitly states otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0172] It should also be noted that in the apparatus, devices, and methods of this disclosure, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions to this disclosure.

[0173] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0174] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for executing a task by an intelligent agent, the method comprising: The first end sends the target task to the second end; The second end starts from the unified starting state of the target task and uses multiple branch agents to explore in parallel to obtain the target path corresponding to the target task; The second end sends the target path to the first end; The first end utilizes the main intelligent agent to execute the target task based on the target path.

2. The intelligent agent task execution method according to claim 1, wherein, The second end starts from the unified starting state of the target task and uses multiple branch agents to explore in parallel to obtain the target path corresponding to the target task, including: The second end determines at least one exploration phase based on the target task; The second end determines the unified starting state of the first exploration phase based on the unified starting state of the target task. For each exploration phase, the second end starts from the unified phase start state corresponding to the exploration phase, uses multiple branch agents to explore in parallel, obtains the target path segment corresponding to the exploration phase, and determines the unified phase start state corresponding to the next exploration phase based on the endpoint state reached by the target path segment. The second end determines the target path corresponding to the target task based on the target path segments corresponding to each of the exploration stages.

3. The intelligent agent task execution method according to claim 2, wherein, The second end determines the target path corresponding to the target task based on the target path segments corresponding to each of the exploration stages, including: The second end determines whether the completion condition corresponding to the target task has been met based on the endpoint state reached by the target path segment corresponding to the last exploration stage of the sorting. In response to the completion condition being met, the second end splices the target path segments corresponding to each of the exploration stages to obtain the target path corresponding to the target task.

4. The intelligent agent task execution method according to claim 2, wherein, The second end starts from the unified stage starting state corresponding to the exploration stage, and uses multiple branch agents to explore in parallel to obtain the target path segment corresponding to the exploration stage, including: The second end utilizes the multiple branch agents to start from the unified stage starting state and execute the target task in parallel in a virtual environment set to the unified stage starting state, thereby obtaining candidate path segments corresponding to each branch agent. The second end evaluates each of the candidate path segments and determines the target path segment from among the candidate path segments based on the evaluation results.

5. The agent task execution method according to any one of claims 2 to 4, wherein, The second end starts from the unified stage starting state corresponding to the exploration stage, and explores in parallel using multiple branch agents, including: The second end starts from the unified stage starting state corresponding to the exploration stage, and explores in parallel using the multiple branch agents based on the exploration strategies corresponding to each branch agent.

6. The intelligent agent task execution method according to claim 4, wherein, The parallel execution of the target task yields candidate path segments corresponding to each branch agent, including: Based on the current state information of the target task and the virtual environment, at least one operation step is executed in parallel and iteratively until the branch agent reaches the termination condition corresponding to the exploration phase. The second end determines the candidate path segment corresponding to each branch agent based on the at least one operation step performed by each branch agent.

7. The intelligent agent task execution method according to claim 6, wherein, The branch agent reaches the termination condition corresponding to the exploration phase, including: The branch agent executes a preset number of the operation steps during the exploration phase; or... The branch agent performs the target task for a preset duration during the exploration phase.

8. The agent task execution method according to any one of claims 1 to 4, wherein, The second end starts from a unified starting state of the target task. Before utilizing multiple branch agents for parallel exploration, the method further includes: The first terminal sends the current device status information to the second terminal; The second end determines the unified starting point state of the target task based on the current device status information.

9. The intelligent agent task execution method according to claim 8, wherein, The first terminal sends the current device status information to the second terminal, including: The first terminal captures screenshot information and interface layout information from the current device status information; The first terminal sends the screenshot information and the interface layout information to the second terminal.

10. The intelligent agent task execution method according to claim 4, wherein, Before the second end utilizes the multiple branch agents to execute the target task in parallel from the unified stage starting state in a virtual environment set to the unified stage starting state, the method further includes: The second end synchronizes the virtual environment where each branch agent is located to the unified starting state of the exploration phase.

11. The intelligent agent task execution method according to claim 4, wherein, The second end starts from a unified stage starting state corresponding to the exploration stage. Before utilizing multiple branch agents for parallel exploration, the method further includes: The first terminal sends the device operating environment information corresponding to the first terminal to the second terminal; Based on the device operating environment information, the second end establishes multiple virtual environments that match the device operating environment of the first end, and deploys at least one branch agent in each virtual environment.

12. An intelligent agent task execution system, the system comprising: The main intelligent agent, deployed on the first end, is used to send the target task to the second end; Multiple branch agents are deployed in the second terminal to explore in parallel from the unified starting state of the target task and obtain the target path corresponding to the target task. A sending unit, deployed in the second end, is used to send the target path to the first end; The main intelligent agent is also used to execute the target task based on the target path.

13. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the method as described in any one of claims 1 to 11.

14. An electronic device comprising: One or more processors, and a memory; the memory storing computer instructions; the computer instructions, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 11.

Images