Task execution method and system and storage medium

Abstract

The invention provides a task execution method and system and a storage medium. The task execution method comprises the steps that the task execution system can determine an intervention point in a processing flow based on preset intervention logic in the processing flow of calling a large language model to execute a target task by an intelligent agent system, and the processing flow of executing the target task by the large language model is stopped. And then, the task execution system can execute a notification corresponding to the intervention point at the intervention point to obtain an execution result, and the notification comprises an intervention process. And finally, the task execution system can decide a processing flow of the target task based on the execution result, and decisions comprise execution continuing, execution re-execution or execution termination. The intervention point can be determined in the processing flow that the intelligent agent system calls the large language model to execute the target task, and the intelligent agent system is intervened at the intervention point, so that the reliability of the intelligent agent system is improved.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to a task execution method, system, and storage medium. Background Technology

[0002] Intelligent agent systems have been widely used in scenarios where natural language interaction is used to automate processes. In these scenarios, intelligent agent systems can respond to instructions given via natural language and automatically execute corresponding tasks. These systems may include at least one agent driven by a Large Language Model (LLM).

[0003] Currently, intelligent agents driven by large language models may generate unverified content during reasoning, resulting in varying degrees of hallucination problems.

[0004] However, the illusion problem may cause large language models to output unreliable content, which in turn can lead to reduced security and decision-making errors in the agent system, affecting the reliability and security of the agent system.

[0005] The background information is merely information known only to the inventor and does not imply that such information had entered the public domain before the date of this application, nor does it imply that it could be considered prior art in this disclosure. Summary of the Invention

[0006] This specification provides a task execution method, system, and storage medium that can identify intervention points in the processing flow of an intelligent agent system calling a large language model to execute a target task, and intervene in the intelligent agent system at the intervention points to improve the reliability of the intelligent agent system.

[0007] To achieve the above objectives, the embodiments in this specification adopt the following technical solutions: Firstly, this specification provides a task execution method, the method comprising: in a processing flow in which an intelligent agent system calls a large language model to execute a target task, determining an intervention point in the processing flow based on preset intervention logic, and stopping the large language model from executing the processing flow of the target task; at the intervention point, executing a notification corresponding to the intervention point to obtain an execution result, the notification including: the intervention flow; and based on the execution result, deciding on the processing flow of the target task, the decision including: continuing execution, re-execution, or terminating execution.

[0008] In some embodiments, the preset intervention logic includes a first intervention logic at the level of a large language model and a second intervention logic at the level of an agent; the step of determining intervention points in the processing flow where the agent system calls the large language model to execute the target task, based on the preset intervention logic, includes: identifying intervention points at the level of the large language model based on the first intervention logic during the process of generating the output stream in the processing flow where the agent system calls the large language model to execute the target task; and identifying intervention points at the level of the agent based on the second intervention logic in the processing flow where the agent system calls the large language model to execute the target task.

[0009] In some embodiments, during the process of generating an output stream in the processing flow of the intelligent agent system calling the large language model to execute the target task, identifying intervention points in the dimension of the large language model based on the first intervention logic includes: during the process of generating an output stream in the processing flow of the intelligent agent system calling the large language model to execute the target task, matching at least one stopping sequence in the preset aspect configuration information with the output stream, wherein the stopping sequence includes one or more labels; when it is determined that the output stream includes a target label, determining the position of the target label as an intervention point, wherein the target label is any label in the stopping sequence.

[0010] In some embodiments, at least one stopping sequence in the preset aspect configuration information includes: a stopping sequence of the instruction framework; the stopping sequence of the instruction framework includes one or more of the following tags: before planning, after planning, before execution, after execution, before summary, and after summary.

[0011] In some embodiments, in the processing flow of an intelligent agent system calling a large language model to execute a target task, based on the second intervention logic, an intervention point in the intelligent agent dimension is identified, including: in the processing flow of an intelligent agent system calling a large language model to execute a target task, when it is detected that the processing flow has reached a preset stop node, the position of the stop node is determined as the intervention point.

[0012] In some embodiments, the stopping node includes: a stopping node for a multi-agent layer and / or a stopping node for an agent layer; wherein, the stopping node for a multi-agent layer includes one or more of the following nodes: process start, pre-execution verification of the tool, post-execution processing of the tool, pre-execution processing of the agent, post-execution processing of the agent, information verification and process check, and process end; the stopping node for an agent layer includes one or more of the following nodes: initialization processing, information verification and process check, pre-execution verification of the tool, post-execution processing of the tool, and termination processing.

[0013] In some embodiments, the notification includes: observation notification, modification notification, and / or interruption notification; the step of deciding the processing flow of the target task based on the execution result includes: if the notification is an observation notification, obtaining the execution result of the observation notification and continuing to execute the processing flow of the target task; if the notification is a modification notification, re-executing the processing flow of the target task based on the execution result; if the notification is an interruption notification, terminating the execution of the processing flow of the target task based on the execution result.

[0014] In some embodiments, when the intervention point is an intervention point at the level of a large language model, if the notification is a modification notification, then based on the execution result, the processing flow of the target task is re-executed, including: generating prompt data based on the execution result and the context information of the intervention point, injecting the prompt data into the large language model, and guiding the large language model to re-execute the processing flow of the target task.

[0015] In some embodiments, injecting the prompt data into the large language model and re-executing the processing flow of the target task includes: performing format validity verification on the prompt data; if the verification is successful, injecting the prompt data into the large language model and guiding the large language model to re-executing the processing flow of the target task; the method further includes: if the verification fails, terminating the execution of the processing flow of the target task.

[0016] In some embodiments, when the intervention point is an intervention point at the agent level, if the notification is a modification notification, then based on the execution result, the processing flow of the target task is re-executed, including: determining the target subtask to be re-executed in the processing flow of the target task according to the position of the stop node corresponding to the intervention point; generating input data for the target subtask based on the execution result and the context information of the intervention point; and re-executing the target subtask based on the input data of the target subtask.

[0017] In some embodiments, the observation notifications include one or more of the following: log recording, metric statistics, tracking points, important information extraction, information verification and process inspection; the modification notifications include one or more of the following: parameter rewriting, prompt enhancement, result post-processing, context compression, information verification and process inspection, pre-filled continuation; the interruption notifications include one or more of the following: security interception, permission denial, illegal input, risk exceeding limits, compliance failure.

[0018] In some embodiments, when the intervention point corresponds to multiple notifications, the step of executing the notification corresponding to the intervention point at the intervention point and obtaining the execution result includes: executing the multiple notifications corresponding to the intervention point in the order of observation notifications, modification notifications and interruption notifications, and obtaining the execution result.

[0019] In some embodiments, the method further includes: generating an audit log; the audit log includes at least: the target label or stop node corresponding to the intervention point, the notification type, the execution result, and the decision based on the execution result.

[0020] Secondly, this specification also provides a task execution system, comprising: at least one storage medium storing at least one instruction set for executing a target task using an intelligent agent system; and at least one processor communicatively connected to the at least one storage medium, wherein, when the task execution system is running, the at least one processor reads the at least one instruction set and implements the method provided in the first aspect according to the instructions of the at least one instruction set.

[0021] Thirdly, this specification also provides a computer-readable non-volatile storage medium, wherein the computer-readable non-volatile storage medium stores at least one instruction set, which, when executed by at least one processor, implements the method provided in the first aspect.

[0022] Other functionalities of the task execution methods, systems, and storage media provided in this specification are partially listed in the following description. The inventive aspects of the task execution methods, systems, and storage media provided in this specification can be fully understood through practice or use of the methods, apparatus, and combinations described in the detailed examples below. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A schematic diagram illustrating an application scenario of the task execution method provided according to embodiments of this specification is shown. Figure 2 A hardware structure diagram of a computing system provided according to an embodiment of this specification is shown; Figure 3 A flowchart of a task execution method provided according to an embodiment of this specification is shown; Figure 4A cross-sectional schematic diagram corresponding to the task execution method provided according to the embodiments of this specification is shown; Figure 5 A schematic diagram illustrating intervention points for identifying large language model dimensions provided according to embodiments of this specification is shown. Figure 6 A schematic diagram of the output stream for a large language model dimension provided according to an embodiment of this specification is shown; Figure 7 A schematic diagram of an intervention point utilizing a multi-agent layer of stop nodes is shown according to an embodiment of this specification; Figure 8 A schematic diagram of an intervention point utilizing a stop node of an agent layer, provided according to an embodiment of this specification, is shown; and Figure 9 This diagram illustrates a processing flow for determining the target task using the execution results corresponding to intervention points within a large language model dimension, according to an embodiment of this specification. Detailed Implementation

[0025] The following description provides specific application scenarios and requirements for this specification, intended to enable those skilled in the art to make and use the contents of this specification. Various partial modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of this specification. Therefore, this specification is not limited to the embodiments shown, but rather to the widest scope consistent with the claims.

[0026] The terminology used herein is for the purpose of describing particular exemplary embodiments only and is not restrictive. For example, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. When used in this specification, the terms “comprising,” “including,” and / or “containing” mean that the associated integers, steps, operations, elements, and / or components are present, but do not exclude the presence of one or more other features, integers, steps, operations, elements, components, and / or groups, or that other features, integers, steps, operations, elements, components, and / or groups may be added to the system / method.

[0027] Considering the following description, these and other features of this specification, as well as the operation and function of the related components of the structure, and the economy of assembly and manufacture of the parts, can be significantly improved. All of these form part of this specification with reference to the accompanying drawings. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of this specification. It should also be understood that the drawings are not drawn to scale.

[0028] The flowcharts used in this specification illustrate operations implemented according to some embodiments of this specification. It should be clearly understood that the operations in the flowcharts may not be implemented in a sequential order. Instead, the operations may be implemented in reverse order or simultaneously. Furthermore, one or more additional operations may be added to the flowcharts. One or more operations may be removed from the flowcharts.

[0029] For ease of description, the terms that will appear later in this manual will be explained first.

[0030] Term 1: High-Order Program. A High-Order Program (HOP) is a control system proposed to solve the illusion problem of large language models, enabling the automation of standard operating procedures (SOPs) in a target domain. HOPs possess the characteristics of being executable, verifiable, derivable, and programmable. Executable means that an HOP can be a script written in a Domain Specific Language (DSL), which can be executed by a task execution system to intervene in an intelligent agent system. Verifiable means that after an HOP intervenes in an intelligent agent system, the data generated by the intelligent agent system can be verified based on the target domain SOPs, domain knowledge, and tool invocation logic contained within the HOP to identify and improve potential illusion problems. Derivable means that an HOP can be derived from different scenarios within the same domain to create HOP copies suitable for different scenarios. Programmable means that each HOP can be customized according to the needs of actual applications, adjusting the execution order and logic of different modules within the HOP to provide a HOP that better fits the actual requirements.

[0031] Term 2: Aspect. An aspect is a modular language construct in Aspect-Oriented Programming (AOP) that encapsulates multiple potentially invoked functions, tools, or modules. These functions, tools, or modules can be centrally managed through aspects; that is, any function, tool, or module encapsulated within an aspect can be invoked simply by accessing the aspect.

[0032] Term 3: Pointcut. A pointcut, also known as a join point, is a point in the process of an intelligent agent system executing a target task that can be intervened by a facet.

[0033] Term 4: Advice. Advice refers to the specific operation performed at the point of contact through an aspect, that is, the intervention logic woven into the HOP.

[0034] Term 5: Instruction Framework. An instruction framework (IF) is a structured instruction execution system designed in an intelligent agent system to regulate and guide the behavior of a large language model. An instruction framework can be pre-arranged, templated instruction data, and can define the execution stages, process constraints, and intervention points of the large language model.

[0035] In this specification, the Large Language Model (LLM) may also be referred to simply as the Large Model. A Large Language Model is a natural language processing model based on deep learning techniques, typically with billions to hundreds of billions or even more parameters, possessing powerful language understanding and generation capabilities. Large Language Models can employ the Transformer architecture or its variants (such as GPT, BERT, etc.), which utilizes an attention mechanism to globally model sequential data, efficiently handling long-distance dependencies and thus performing excellently 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 excellent 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. Its 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.

[0036] 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 accomplish complex tasks. For example, the Vision Language Model (VLM) discussed in this specification is a branch of MLLMs; a VLM can analyze an image and generate descriptive text. In other examples, MLLMs can also generate corresponding images or videos based on text descriptions. This cross-modal understanding and generation capability makes MLLMs widely applicable in multiple fields.

[0037] 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.

[0038] The following section introduces the application scenarios of this manual.

[0039] The technical solutions provided in this specification are applicable to scenarios where intelligent agent systems are used to execute tasks. As an example, an intelligent agent system can be deployed within the task execution system, and this system can contain at least one intelligent agent driven by a large language model. Upon receiving a target task, the task execution system can invoke the intelligent agent system to execute the target task. When executing the target task, the intelligent agent system can utilize the large language model corresponding to at least one intelligent agent to perform steps such as task decomposition, task processing, and task summarization, and output the execution result of the target task.

[0040] However, large language models may suffer from the illusion problem, which involves outputting unverified content. When an agent is performing a target task, the illusion problem can cause the system to make incorrect decisions, leading to reduced reliability and security, and hindering its application in scenarios requiring high reliability.

[0041] To address this, this specification provides a task execution method that can be executed by a task execution system. The method includes: The task execution system, within the processing flow of a target task called by an intelligent agent system using a large language model, determines an intervention point in the processing flow based on preset intervention logic, stopping the large language model from executing the target task's processing flow. Then, at the intervention point, the task execution system executes a notification corresponding to the intervention point, obtaining an execution result. The notification includes the intervention process. Finally, based on the execution result, the task execution system decides on the processing flow of the target task, including: continuing execution, re-execution, or termination.

[0042] In the solution provided in this specification, the task execution system determines intervention points in the processing flow through preset intervention logic and stops the processing flow of the large language model. Then, the task execution system executes the corresponding intervention process at the intervention point to obtain the execution result, and decides whether to continue, re-execute, or terminate the processing flow of the target task based on the execution result. The task execution system can perform reliability verification on the execution process of the target task at the intervention point and correct any content that may have hallucination problems. Through this mechanism, the uncertainty in the execution process of the large language model can be effectively reduced, thereby improving the reliability and security of the intelligent agent system when executing tasks. This allows the intelligent agent system to be applied to scenarios requiring high reliability, further expanding the application scenarios of intelligent agent systems.

[0043] Figure 1 A schematic diagram illustrating an application scenario of the task execution method provided according to embodiments of this specification is shown. For example... Figure 1 As shown, the application scenario 100 may include a task execution system 11 and a terminal device 12.

[0044] refer to Figure 1 The task execution system 11 is a system for executing a target task, and an intelligent agent system is deployed in the task execution system 11. The task execution system 11 can be deployed on a device or device cluster with data processing capabilities. For example, the task execution system 11 can be deployed on physical devices such as servers, server clusters, and cloud servers. In this case, the physical device corresponding to the task execution system 11 can store data or instructions for executing the task execution method described in this specification, and can execute or be used to execute the data or instructions. In some embodiments, the physical device corresponding to the task execution system 11 may include a hardware device with data information processing capabilities and the necessary programs required to drive the hardware device to work. The intelligent agent system may include at least one intelligent agent driven by a large language model. The large language model corresponding to these intelligent agents may be deployed on the same device or device cluster as the task execution system 11, or on a different device or device cluster than the task execution system 11. This specification does not impose any restrictions on this.

[0045] As an example, terminal device 12 may be an electronic device capable of accessing a task execution system. For example, terminal device 12 may include mobile devices, tablets, laptops, built-in devices in motor vehicles, or similar content, or any combination thereof. In some embodiments, the mobile device may include wearable devices, camera devices, smart mobile devices, virtual reality devices, augmented reality devices, or similar devices, or any combination thereof. In some embodiments, wearable devices include smartwatches, smart bracelets, smart glasses, etc. In some embodiments, the smart mobile device may include smartphones, personal digital assistants, gaming devices, navigation devices, etc., or any combination thereof. In some embodiments, the virtual reality device or augmented reality device may include head-mounted displays, virtual reality headsets, virtual reality glasses, virtual reality patches, augmented reality headsets, augmented reality glasses, augmented reality patches, or similar content, or any combination thereof.

[0046] In some embodiments, the terminal device 12 may have one or more applications (APPs) installed. The APPs provide the ability or interface to access the task execution system 11. The APPs include, but are not limited to: web browser APPs, search APPs, chat APPs, shopping APPs, video APPs, financial management APPs, instant messaging tools, email clients, social media platform software, etc. In some embodiments, the terminal device 12 may have a target APP installed. The target APP is the application corresponding to the task execution system 11. In some embodiments, the target APP can respond to user operations by sending an execution request for a target task to the task execution system 11.

[0047] It should be understood that Figure 1 The number of task execution systems 11 and terminal devices 12 shown is merely illustrative. Depending on implementation needs, there can be any number of financial systems 11 and terminal devices 12.

[0048] Figure 2 A hardware structure diagram of a computing system provided according to an embodiment of this specification is shown. The computing system 200 can serve as... Figure 1 The task execution system 11 in the specification executes the task execution method described herein.

[0049] like Figure 2 As shown, the computing system 200 may include at least one storage medium 230 and at least one processor 220. In some embodiments, the computing system 200 may also include a communication port 250 and an internal communication bus 210. The computing system 200 may also include I / O components 260.

[0050] The internal communication bus 210 can connect to different system components. For example, the internal communication bus 210 can connect to storage medium 230, processor 220, communication port 250, and I / O component 260, etc.

[0051] I / O component 260 supports input / output between computing system 200 and other components.

[0052] Communication port 250 is used for data communication between computing system 200 and the outside world. For example, communication port 250 can be used for data communication between computing system 200 and a network. Communication port 250 can be a wired communication port or a wireless communication port.

[0053] Storage medium 230 may include a data storage device. The data storage device may be a non-transitory storage medium or a temporary storage medium. For example, the data storage device may include one or more of a disk 232, a read-only storage medium (ROM) 234, or a random access storage medium (RAM) 235. Storage medium 230 also includes at least one instruction set stored in the data storage device. The instruction set may include computer program code, which may include programs, routines, objects, components, data structures, procedures, modules, etc.

[0054] At least one processor 220 may be communicatively connected to at least one storage medium 230. When the computing system 200 is running, at least one processor 220 reads the at least one instruction set and executes the task execution method provided in this specification according to the instructions of the at least one instruction set. The processor 220 may perform the steps included in the task execution method. The processor 220 may be in the form of one or more processors. In some embodiments, the processor 220 may include one or more hardware processors, such as a microcontroller, microprocessor, reduced instruction set computer (RISC), application-specific integrated circuit (ASIC), application-specific instruction set processor (ASIP), central processing unit (CPU), graphics processing unit (GPU), physical processing unit (PPU), microcontroller unit, digital signal processor (DSP), field-programmable gate array (FPGA), advanced RISC machine (ARM), programmable logic device (PLD), any circuit or processor capable of performing one or more functions, or any combination thereof.

[0055] For illustrative purposes only, the accompanying drawings show only one processor 220 for the computing system 200. However, it should be noted that the computing system 200 may also include multiple processors; therefore, the operations and / or method steps disclosed herein may be executed by one processor or by multiple processors in combination. For example, if the processor 220 of the computing system 200 described in this specification executes steps A and B, it should be understood that steps A and B may also be executed jointly or separately by two different processors 220 (e.g., a first processor executes step A, a second processor executes step B, or the first and second processors jointly execute steps A and B).

[0056] Figure 3 A flowchart of a task execution method according to an embodiment of this specification is shown. As previously described, the computing system 200 can execute the task execution method of this specification.

[0057] like Figure 3 As shown, the task execution method may include: S310: In the process of the intelligent agent system calling the large language model to execute the target task, an intervention point is determined in the process based on the preset intervention logic to stop the process of the large language model executing the target task.

[0058] In some embodiments, the target task can be a task within a target domain that requires high reliability. For example, target domains may include areas such as financial compliance, medical diagnostics, industrial control, risk analysis, and autonomous driving. These target domains have high requirements for the reliability of the intelligent agent system. Therefore, the task execution system can intervene in the processing flow during the execution of the target task by the intelligent agent system to identify and correct unreliable aspects of the processing flow.

[0059] Figure 4 A cross-sectional schematic diagram corresponding to the task execution method provided according to an embodiment of this specification is shown.

[0060] refer to Figure 4 An intelligent agent system includes at least one intelligent agent, a corresponding large language model, and tools. During runtime, the intelligent agent system, the large language model, and tools can interact with each other, cooperate, and execute the processing flow of the target task. However, due to the illusion problem inherent in the large language model, unreliable content may appear during the execution of the target task by the intelligent agent system; therefore, the intelligent agent system can be considered to operate on an execution layer plane with inherent uncertainty.

[0061] Pre-defined intervention logic can be executed via HOPs to identify intervention points in the processing flow. Furthermore, HOPs can utilize notifications from these intervention points to execute intervention procedures, identifying and correcting unreliable content. Therefore, HOPs can be viewed as operating on the deterministic control plane, or as aspects corresponding to the intelligent agent system. When pre-defined intervention logic is considered an aspect of the intelligent agent system, each intervention point can be viewed as a cutoff point.

[0062] In some embodiments, when a task execution system uses an intelligent agent system to call a large language model to execute a target task, the processing flow of the target task may include steps such as calling the large language model for reasoning, calling tools, and calling other intelligent agents. The intervention point can be during the process of calling the large language model for reasoning, or it can be before or after each step.

[0063] To identify intervention points at different locations, the task execution system can configure corresponding intervention logic for intervention points of different dimensions. As an example, the preset intervention logic may include a first intervention logic at the large language model dimension and a second intervention logic at the agent dimension. All intervention logic in the preset intervention logic can be executed through HOP.

[0064] In some embodiments, the task execution system can utilize HOP to identify intervention points at the large language model dimension based on a first intervention logic during the process of generating the output stream when the agent system calls the large language model to execute the target task. Furthermore, the task execution system can utilize HOP to identify intervention points at the agent dimension based on a second intervention logic during the process of the agent system calling the large language model to execute the target task.

[0065] Figure 5 A schematic diagram illustrating intervention points for identifying large language model dimensions provided according to embodiments of this specification is shown. Figure 6 A schematic diagram of the output stream for a large language model dimension provided according to an embodiment of this specification is shown.

[0066] In some embodiments, reference Figure 5The task execution system can match at least one stop sequence from the preset aspect configuration information with the output stream during the process of the intelligent agent system calling the large language model to execute the target task. When the task execution system determines that the output stream includes the target label, it determines the position of the target label as the intervention point, and the target label is any label in the stop sequence. The stop sequence is preset and includes at least one label. When the task execution system detects that the output stream of the large language model outputs content that completely matches any label in the stop sequence, the task execution system can use HOP to terminate the generation process of the large language model and intervene in the content generated by the large language model.

[0067] In some embodiments, the task execution system may pre-configure an instruction framework, defining multiple stages within the instruction framework for the large language model to execute the target task, and instructing the large language model to output corresponding labels before and after each stage. For example, refer to... Figure 5 When a large language model performs a target task, it can include planning, execution, and conclusion.

[0068] For example, after receiving input instructions from agent 1, the large language model will begin parsing and reasoning about the input instructions, and plan the execution steps of the target task. During this process, the large language model will stream the results of the planning phase. A pre-configured instruction framework can instruct the large language model to output the results of the planning phase before starting the streaming output of the planning phase results. <plan> This marks the start of the planning phase. It also instructs the large language model to output the results after the planning phase has been fully completed.< / plan> The project concludes with the marking of the project phase.

[0069] Similarly, large language models can output at the beginning of the execution phase. <execute> The process begins with marking the execution phase. Then, the large language model executes each step of the execution plan, streaming the execution result for each step. Finally, it instructs the large language model to output the results after outputting the execution results for all steps.< / execute> The execution phase ends when the marker is displayed.

[0070] Furthermore, the large language model can output at the beginning of the summary phase. <conclusion> The process begins with a summary phase. The large language model then summarizes the results based on the execution results of each step, streaming the summary phase output. Finally, it instructs the large language model to output the complete summary result after outputting the final summary.< / conclusion> The process ends with a marked summary phase.

[0071] As an example, based on the above example, at least one stopping sequence in the preset aspect configuration information includes: a stopping sequence of the instruction frame. The stopping sequence of the instruction frame can be a string marking the start and end of each stage. For example, refer to... Figure 5 The stop sequence of the instruction frame may include one or more of the following tags: pre-plan ( <plan> ), after the plan (< / plan> ), before execution ( <execute> After execution ()< / execute> ), before the summary ( <conclusion> ), after summarizing (< / conclusion> ).

[0072] In some embodiments, reference Figure 6During this process, the large language model outputs the reasoning process and results of planning, execution, and summarization to agent 1 via an output stream. The task execution system can obtain the streaming output of the large language model and match the labels in the stopping sequence within the streaming output.

[0073] For example, suppose the streaming output of a large language model is: “ <plan> ……< / plan> <execute> ……< / execute> <conclusion> ……< / conclusion> " Task execution systems can utilize the label "" in the stop sequence of the instruction frame. <plan> 、< / plan> , <execute> 、< / execute> , <conclusion> 、< / conclusion> "Matching is performed in the streaming output results. (Reference)" Figure 6 When the task execution system identifies any tag in the stop sequence of the instruction frame in the streaming output, it can determine that the tag is the target tag, and the location of the target tag can be determined as the intervention point.

[0074] In this embodiment, the task execution system matches at least one stopping sequence from the preset aspect configuration information with the output stream. When the task execution system determines that the output stream includes a target label, it identifies the position corresponding to the target label as an intervention point, thus achieving accurate identification of intervention points in the reasoning process of the large language model. Furthermore, the task execution system can intervene in the execution process of the large language model in a timely and effective manner based on the identified intervention points, thereby monitoring and correcting the output process of the large language model and improving the reliability of the intelligent agent system when executing the target task.

[0075] In some embodiments, the task execution system can utilize HOP to determine the location of the stop node as an intervention point when the processing flow of the intelligent agent system calling the large language model to execute the target task is detected to have reached a preset stop node.

[0076] Among them, the stopping nodes include: the stopping nodes of the multi-agent layer and / or the stopping nodes of the agent layer.

[0077] In some embodiments, when an agent system includes multiple agents, the agent system can be viewed as a multi-agent layer (or multi-agent aspect). In a multi-agent layer, multiple agents can call upon each other during the processing flow of the agent system calling a large language model to execute the target task. In this case, the task execution system can utilize the stopping nodes of the multi-agent layer to identify intervention points when multiple agents call upon each other and execute the target task's processing flow.

[0078] When an agent system includes one agent, or when an agent system invokes one agent to execute a task, the agent system can be considered an agent layer (or agent aspect). Within the agent layer, the agent system can invoke the agent to execute the target task or subtasks decomposed from the target task. In this case, the task execution system can utilize the stopping nodes of the agent layer to identify intervention points when the agent performs reasoning and invokes tools to execute tasks.

[0079] Figure 7 A schematic diagram of an intervention point utilizing a multi-agent layer of stopping nodes is shown according to an embodiment of this specification. Figure 8 A schematic diagram of an intervention point utilizing a stop node of an intelligent agent layer, provided according to an embodiment of this specification, is shown.

[0080] Among them, reference Figure 7 The stopping node of a multi-agent layer includes one or more of the following nodes: process start, pre-execution verification of the tool, post-execution processing of the tool, pre-execution processing of the agent, post-execution processing of the agent, information verification and process check, and process end. (Reference) Figure 8 The stopping node of the intelligent agent layer includes one or more of the following nodes: initialization processing, information verification and process inspection, tool execution pre-verification, tool execution post-processing, and termination processing.

[0081] In some embodiments, when creating an agent system, multiple agent layers and agent layer stop nodes can be pre-set by embedding tracking points in the code and configuring corresponding trigger conditions. For example, a task execution system can monitor received instructions through tracking points, and when it receives an instruction to execute the target task, it can determine to trigger the stop node corresponding to "process start". Alternatively, the task execution system can also detect calling instructions output by a large language model (such as instructions to call a tool or other agents), and after detecting the calling instruction, determine to trigger the stop node of "agent execution preprocessing" or "tool execution preprocessing" based on the target (tool or other agent) corresponding to the calling instruction.

[0082] In this embodiment, the task execution system utilizes pre-configured stop nodes within the agent system and identifies these pre-defined stop nodes during the processing flow of the agent system calling the large language model to execute the target task. The task execution system determines the identified stop nodes as intervention points and intervenes in the target task's processing flow at these points, identifying and correcting any potentially unreliable aspects, thereby improving the reliability and security of the agent system when executing the target task.

[0083] S320: At the intervention point, execute the notification corresponding to the intervention point and obtain the execution result. The notification includes: the intervention process.

[0084] In some embodiments, each intervention point includes at least one corresponding advice, and an advice may include a corresponding intervention process, wherein the intervention process can be executed via HOP.

[0085] As an example, notifications can include: observation notifications, modification notifications, and / or interruption notifications. Different notifications correspond to different intervention processes.

[0086] In some embodiments, observation notifications are primarily used to obtain contextual information without intervening in the processing flow. For example, observation notifications may include one or more of the following intervention processes: logging, metric statistics, tracking points, important information extraction, information verification, and process checks. Modification notifications are primarily used to modify data in the context, such as input data, output data, or status data. For example, modification notifications may include one or more of the following intervention processes: parameter rewriting, prompt enhancement, result post-processing, context compression, information verification and process checks, and pre-filled continuation. Interruption notifications are primarily used to prematurely terminate the processing flow. For example, interruption notifications may include one or more of the following intervention processes: security blocking, permission denial, illegal input, risk exceeding limits, and compliance failure.

[0087] In some embodiments, one intervention point may correspond to multiple notifications. When one intervention point corresponds to multiple notifications, the task execution system can execute the multiple notifications corresponding to the intervention point in the order of observation notifications, modification notifications, and interruption notifications to obtain the execution result.

[0088] As an example, when at an intervention point, if the task execution system needs to execute the intervention process of the modification notification based on the execution result of the observation notification, the task execution system can concurrently issue observation notifications and modification notifications at that intervention point.

[0089] For example, suppose the task execution system concurrently sends observation-type notifications (metric statistics) and modification-type notifications (parameter rewriting) at the intervention point. In this case, the task execution system can first use HOP to perform metric statistics at the intervention point to obtain the statistical metric data. Then, the task execution system uses HOP to determine the parameters that need to be rewritten based on the statistical metric data, and executes the parameter rewriting intervention flow based on the parameters that need to be rewritten.

[0090] In this embodiment, the task execution system can execute intervention procedures based on corresponding notifications at intervention points. These notifications include observation notifications, modification notifications, and interruption notifications. The task execution system can use different notifications to intervene in the execution process of the target task and obtain the execution result after intervention. The task execution system can flexibly adopt corresponding intervention procedures according to actual needs, monitor and verify the processing flow during the execution of the target task by the intelligent agent system, thereby improving the reliability and security of the intelligent agent system when executing the target task.

[0091] S330: Based on the execution results, determine the processing flow of the target task. The decision includes: continue execution, re-execute, or terminate execution.

[0092] In some embodiments, the execution result of an observation-type notification is whether the reliability of the obtained context information meets the requirements. If the notification is an observation-type notification, the task execution system can instruct the agent system to continue executing the processing flow of the target task when the execution result indicates that the reliability of the obtained context information meets the requirements. Alternatively, the task execution system can instruct the agent system to continue modifying the intervention flow corresponding to the notification at that intervention point when the execution result indicates that the reliability of the obtained context information does not meet the requirements.

[0093] Reliability can be categorized into two dimensions: flow execution reliability and step execution reliability. Flow execution reliability refers to, at a macro level, whether the execution direction of the target task's processing flow is correct and whether the flow is complete. Step execution reliability refers to, at a micro level, whether the execution quality of each step / subtask meets the requirements and whether the execution results conform to standards.

[0094] For example, assuming an observation-type notification received at an intervention point is a log entry, the task execution system can use HOP to read and record the corresponding log from the intervention point. After recording the log, the task execution system can use HOP to perform reliability checks on the content recorded in the log to obtain the execution result. When the execution result indicates that the reliability of the content recorded in the log meets the requirements, the task execution system can use HOP to instruct the agent system to continue executing the target task's processing flow. When the execution result indicates that the reliability of the content recorded in the log does not meet the requirements, the task execution system can use HOP to instruct the agent system to continue modifying the intervention flow corresponding to the notification at that intervention point to correct the content that does not meet the reliability requirements.

[0095] In some embodiments, the execution result of a modification-type notification is the content in the context data that needs to be modified. For example, the context data can be input data, output data, or state data. The task execution system can use HOP to perform reliability verification on this context data, identifying content in the context data that may have reliability issues and needs modification. Based on the execution result of the modification-type notification and the context data, the task execution system can generate modified context data. The task execution system can then use the modified context data to re-execute the processing flow of the target task. Alternatively, if the execution result indicates that there is no content in the context data that needs modification, the task execution system can use HOP to instruct the agent system to continue executing the processing flow of the target task.

[0096] As an example, suppose the intervention point is at the level of a large language model, the notification is a modification notification, and the execution result indicates that there is content in the context data that needs to be modified. The task execution system can use HOP to generate prompt data based on the execution result and the context information of the intervention point. The prompt data includes the modified context information. Then, the task execution system can use HOP to inject the prompt data into the large language model, guiding the large language model to re-execute the processing flow of the target task. The prompt data can be prompt words described in natural language, or it can be structured instruction code or instruction vectors based on an instruction framework.

[0097] As an example, suppose the intervention point is at the agent level, the notification is a modification notification, and the execution result indicates that there is content in the context data that needs to be modified. The task execution system can use HOP (Host Pointer) to determine the target subtask to be re-executed in the processing flow of the target task based on the location of the stop node corresponding to the intervention point. Then, the task execution system can use HOP to generate the input data for the target subtask based on the execution result and the context information of the intervention point. Finally, the task execution system can re-execute the target subtask based on the input data.

[0098] The processing flow of the target task can include multiple subtasks, and the intervention point can be a node where two subtasks are switched. At this intervention point, if the output data of the previous subtask (i.e., the target subtask) contains content that needs modification, then the previous subtask becomes the target subtask to be re-executed. The task execution system can use the HOP to obtain the original input data of the target subtask, and modify the original input data based on the content that needs modification in the output data to generate the input data of the target subtask. Finally, the task execution system can use the HOP to instruct the intelligent agent system to re-execute the target subtask based on the input data of the target subtask.

[0099] In some embodiments, before re-executing the target task's processing flow using the modified context data, the task execution system can also validate the modified context data. For example, referring to the previous example, when the task execution system injects the prompt data into the large language model and re-executes the target task's processing flow, it can first use HOP to perform format validity validation on the prompt data. When the format validity validation result indicates successful validation, the task execution system can inject the prompt data into the large language model, guiding the large language model to re-execute the target task's processing flow. When the format validity validation result indicates failed validation, the task execution system can execute an interruption notification at that intervention point to terminate the execution of the target task's processing flow.

[0100] In some embodiments, the execution result of an interruption notification is whether to determine whether to terminate the processing flow of the target task. If the notification is an interruption notification, the task execution system may terminate the processing flow of the target task if the execution result of the interruption notification determines that the processing flow of the target task should be terminated.

[0101] As an example, referring to the previous example, when the task execution system uses HOP to validate the modified context data, and the validation result indicates a format validity failure, the intervention process corresponding to the illegal input in the interrupt notification can be executed. The task execution system can use HOP to re-examine the modified context data that failed validation to determine whether the modified context data can be modified to pass the format validity validation. If the review result indicates that the modified context data cannot pass the format validity validation after modification, the process of terminating the execution of the target task can be determined.

[0102] If the audit results indicate that the modified context data passes the format validity check, the task execution system can use HOP to update the modified context data and perform a second format validity check on the updated context data. After the updated context data passes the format validity check, the task execution system can use the updated context data to re-execute the target task's processing flow.

[0103] In this embodiment, the task execution system determines the processing flow of the target task based on the execution result of the intervention process at the intervention point. The task execution system can continue to execute reliable processing flows, correct unreliable processing flows, and terminate processing flows that cannot be corrected. Based on the above decisions, the probability of the intelligent agent system outputting unreliable results can be effectively reduced, significantly improving the overall reliability and security of the task execution system in executing the target task using the intelligent agent system.

[0104] Here, we will take the process of the task execution system using the execution results corresponding to the intervention points of the large language model dimension to determine the processing flow of the target task as an example to further explain the task execution method.

[0105] Figure 9 This diagram illustrates a processing flow for determining the target task using the execution results corresponding to intervention points within a large language model dimension, according to an embodiment of this specification.

[0106] refer to Figure 9 In the processing flow where the intelligent agent system calls the large language model to execute the target task, the large language model outputs the processing results in a streaming manner. The task execution system can use the labels in the stopping sequence to detect whether the target label exists in the streaming output processing results. When the task execution system detects the presence of the target label in the streaming output processing results, it can determine the location of the target label as an intervention point.

[0107] In some embodiments, reference Figure 9 Similar to the previous example of a stopping sequence, when the task execution system detects the presence of "", "" in the processing results of the streaming output <execute> ”、"< / execute> " <conclusion>When waiting for the target label, it can be determined Figure 9 The four intervention points shown are labeled "1 (intervention point 1)", "2 (intervention point 2)", "3 (intervention point 3)" and "4 (intervention point 4)". Intervention points 1 and 4 correspond to observation-type notifications, while intervention points 2 and 3 correspond to modification-type notifications.

[0108] Once the task execution system identifies intervention point 1 using the label "", it can use HOP to instruct the large language model to pause the processing flow. Then, the task execution system can use HOP to execute the intervention flow corresponding to the observation-type notification at intervention point 1, verifying the reliability of the step execution dimension at intervention point 1 and obtaining the execution result. When the reliability of the execution result representing the context information meets the requirements, the task execution system can use HOP to instruct the large language model to continue executing the processing flow.

[0109] When the task execution system uses the tag " <execute>"After identifying intervention point 2, the HOP can be used to instruct the large language model to pause the execution of the processing flow. Then, the task execution system can use the HOP to execute the intervention flow corresponding to the modification notification at intervention point 2, verify the reliability of the step execution dimension at intervention point 2, and obtain the execution result. When the reliability of the execution result representing the context information does not meet the requirements, the task execution system can use the HOP to instruct the large language model to re-execute the plan-related processes in the target task's processing flow."

[0110] When the task execution system uses the tag "< / execute> "After identifying intervention point 3, the HOP can be used to instruct the large language model to pause the processing flow. Then, the task execution system can use the HOP to execute the intervention flow corresponding to the modification notification at intervention point 3, verify the reliability of the process execution dimension at intervention point 3, and obtain the execution result. When the execution result includes content in the context data that needs to be modified, the task execution system can use the HOP to generate instruction data based on the execution result and context data, and weave the instruction data into the large language model from intervention point 3, guiding the large language model to re-execute the execution-related processes in the target task's processing flow."

[0111] When the task execution system uses the tag " <conclusion>"After identifying intervention point 4, the HOP can be used to instruct the large language model to pause the processing flow. Then, the task execution system can use the HOP to execute the intervention flow corresponding to the observation notification at intervention point 4, verify the reliability of the step execution dimension at intervention point 4, and obtain the execution result. When the reliability of the execution result representing the context information meets the requirements, the task execution system can use the HOP to instruct the large language model to continue executing the processing flow."

[0112] It should be noted that the process of the task execution system using the execution results corresponding to the intervention points at the agent dimension to determine the target task can be referred to in the above process of the task execution system using the execution results corresponding to the intervention points at the large language model dimension to determine the target task, and will not be elaborated here.

[0113] In some embodiments, the task execution system may record and generate audit logs during the processing flow of executing the target task and during intervention in the processing flow of executing the target task. As an example, the audit log includes at least: the target label or stop node corresponding to the intervention point, the notification type, the execution result, and the decision based on the execution result.

[0114] In this embodiment, the task execution system records and generates audit logs during the execution and intervention processes of the target task. The audit logs include at least the target tag or stop node corresponding to the intervention point, the notification type, the execution result, and the decision based on the execution result. By generating audit logs, the task execution system creates a traceable record of task execution and intervention. This approach not only fully preserves the basis for intervention and decision-making in the processing flow, providing data for subsequent compliance reviews and audit analyses, but also allows for rapid and accurate problem localization based on the audit logs after intervention or decision-making by the task execution system. This effectively enhances the reliability, security, and interpretability of the task execution system.

[0115] In summary, the task execution method and system provided in this specification involve a task execution system that determines intervention points in the processing flow through preset intervention logic and stops the processing flow of the large language model. Then, the task execution system executes the corresponding intervention process at the intervention point to obtain the execution result, and decides whether to continue, re-execute, or terminate the processing flow of the target task based on the execution result. The task execution system can perform reliability verification on the execution process of the target task at the intervention point and correct any content that may have illusion problems. This mechanism effectively reduces the uncertainty in the execution process of the large language model, thereby improving the reliability and security of the intelligent agent system when executing tasks. This allows the intelligent agent system to be applied to scenarios requiring high reliability, further expanding the application scenarios of intelligent agent systems.

[0116] This specification, in another aspect, provides a computer-readable non-transitory storage medium storing at least one set of instructions for performing a target task using an intelligent agent system. When the at least one set of instructions is executed by a processor, it instructs the processor to implement the steps of the task execution method described herein. In some possible embodiments, various aspects of this specification can also be implemented as a program product comprising program code. When the program product is run on a computing system 200, the program code causes the computing system 200 to perform the steps of the task execution method described herein. The program product for implementing the above method may employ a portable compact disk read-only memory (CD-ROM) containing program code and may run on the computing system 200. However, the program product of this specification is not limited thereto. In this specification, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system. The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media 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 fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can send, propagate, or transmit programs for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing the operations described herein can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on computing system 200, partially on computing system 200, as a standalone software package, partially on computing system 200 and partially on a remote computing device, or entirely on a remote computing device.

[0117] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0118] In summary, after reading this detailed disclosure, those skilled in the art will understand that the foregoing detailed disclosure is presented by way of example only and is not restrictive. Although not explicitly stated herein, those skilled in the art will understand that this specification requires various reasonable changes, improvements, and modifications to the embodiments. These changes, improvements, and modifications are intended to be made by this specification and are within the spirit and scope of the exemplary embodiments described herein.

[0119] Furthermore, certain terms in this specification have been used to describe embodiments of this specification. For example, "an embodiment," "an embodiment," and / or "some embodiments" mean that a particular feature, structure, or characteristic described in connection with that embodiment may be included in at least one embodiment of this specification. Therefore, it is to be emphasized and understood that two or more references to "an embodiment" or "an embodiment" or "alternative embodiment" in various parts of this specification do not necessarily refer to the same embodiment. Moreover, specific features, structures, or characteristics may be suitably combined in one or more embodiments of this specification.

[0120] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and to aid in understanding a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art, upon reading this specification, may readily identify some of the devices as separate embodiments. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. And the content of each secondary embodiment is valid even if it contains fewer than all the features of a single foregoing disclosed embodiment.

[0121] Every patent, patent application, publication of a patent application, and other material such as articles, books, specifications, publications, documents, articles, etc., cited herein, except for those inconsistent with or conflicting with this document, or those having a restrictive effect on the widest scope of the claims, may be incorporated herein by reference for all purposes now or hereafter associated with this document. Furthermore, in the event of any inconsistency or conflict between the description, definition, and / or use of relevant terms in any material and the description, definition, and / or use of relevant terms in this document, the terms in this document shall prevail.

[0122] Finally, it should be understood that the embodiments disclosed herein are illustrative of the principles of the embodiments described in this specification. Other modified embodiments are also within the scope of this specification. Therefore, the embodiments disclosed in this specification are merely examples and not limitations. Those skilled in the art can implement the applications described in this specification using alternative configurations based on the embodiments in this specification. Therefore, the embodiments in this specification are not limited to the embodiments precisely described in the applications.< / conclusion> < / conclusion>

Claims

1. A task execution method, the method comprising: In the process of an intelligent agent system calling a large language model to execute a target task, an intervention point is determined in the process based on a preset intervention logic to stop the large language model from executing the target task. At the intervention point, a notification corresponding to the intervention point is executed to obtain the execution result. The notification includes: the intervention process. Based on the execution result, a decision is made regarding the processing flow of the target task, including: continuing execution, re-execution, or termination of execution.

2. The method according to claim 1, wherein, The preset intervention logic includes a first intervention logic at the level of a large language model and a second intervention logic at the level of an agent. In the process of the intelligent agent system calling a large language model to execute a target task, the intervention point is determined in the process based on preset intervention logic, including: During the process of generating the output stream in the intelligent agent system calling the large language model to execute the target task, the intervention points in the dimension of the large language model are identified based on the first intervention logic. as well as In the process of the intelligent agent system calling the large language model to execute the target task, intervention points in the intelligent agent dimension are identified based on the second intervention logic.

3. The method according to claim 2, wherein, In the process of generating the output stream during the processing flow of the intelligent agent system calling the large language model to execute the target task, based on the first intervention logic, intervention points in the dimension of the large language model are identified, including: During the process of generating an output stream by calling a large language model to execute a target task in an intelligent agent system, at least one stop sequence in the preset aspect configuration information is matched with the output stream, and the stop sequence includes one or more labels; When it is determined that the output stream includes a target label, the position of the target label is determined as an intervention point, and the target label is any label in the stop sequence.

4. The method according to claim 3, wherein, At least one stop sequence in the preset section configuration information includes: a stop sequence of the instruction frame; The stop sequence of the instruction framework includes one or more of the following labels: before planning, after planning, before execution, after execution, before summary, and after summary.

5. The method according to claim 2, wherein, In the processing flow of the intelligent agent system calling the large language model to execute the target task, based on the second intervention logic, intervention points at the agent dimension are identified, including: In the process of an intelligent agent system calling a large language model to execute a target task, when the process is detected to have reached a preset stop node, the position of the stop node is determined as the intervention point.

6. The method according to claim 5, wherein, The stopping nodes include: stopping nodes of the multi-agent layer and / or stopping nodes of the agent layer; The stopping node of the multi-agent layer includes one or more of the following nodes: process start, tool execution pre-verification, tool execution post-processing, agent execution pre-processing, agent execution post-processing, information verification and process check, and process end. The stopping node of the intelligent agent layer includes one or more of the following nodes: initialization processing, information verification and process check, tool execution pre-verification, tool execution post-processing, and termination processing.

7. The method according to any one of claims 1-6, wherein, The notifications include: observation notifications, modification notifications, and / or interruption notifications; The process of determining the target task based on the execution result includes: If the notification is an observation notification, then obtain the execution result of the observation notification and continue to execute the processing flow of the target task; If the notification is a modification notification, then based on the execution result, the processing flow of the target task will be re-executed; If the notification is an interruption notification, the processing flow of the target task will be terminated based on the execution result.

8. The method according to claim 7, wherein, When the intervention point is a large language model dimension intervention point, if the notification is a modification notification, then based on the execution result, the processing flow of the target task is re-executed, including: Based on the execution result and the context information of the intervention point, prompt data is generated, and the prompt data is injected into the large language model to guide the large language model to re-execute the processing flow of the target task.

9. The method according to claim 8, wherein, The step of injecting the prompt data into the large language model and re-executing the processing flow of the target task includes: The format of the prompt data is validated. If the validation is successful, the prompt data is injected into the large language model, which then guides the large language model to re-execute the processing flow of the target task. The method further includes: If the verification fails, the processing flow of the target task will be terminated.

10. The method according to claim 7, wherein, When the intervention point is an intervention point at the agent level, if the notification is a modification notification, then based on the execution result, the processing flow of the target task is re-executed, including: Based on the location of the stop node corresponding to the intervention point, determine the target sub-task to be re-executed in the processing flow of the target task; Based on the execution result and the context information of the intervention point, the input data for the target sub-task is generated; and The target subtask is re-executed based on the input data of the target subtask.

11. The method according to claim 7, wherein, The observation-type notifications include one or more of the following: log recording, indicator statistics, tracking points, important information extraction, information verification and process inspection; The modification notifications include one or more of the following: parameter rewriting, prompt enhancement, result post-processing, context compression, information verification and process checking, and pre-filled continuation writing; The interruption notifications include one or more of the following: security block, permission denied, illegal input, risk exceeding limits, and compliance failure.

12. The method according to claim 7, wherein, When an intervention point corresponds to multiple notifications, the step of executing the notification corresponding to the intervention point at that intervention point and obtaining the execution result includes: The multiple notifications corresponding to the intervention point are executed sequentially in the order of observation notifications, modification notifications, and interruption notifications to obtain the execution results.

13. The method according to claim 7, wherein, The method further includes: generating an audit log; the audit log includes at least: the target label or stop node corresponding to the intervention point, the notification type, the execution result, and the decision based on the execution result.

14. A task execution system, comprising: At least one storage medium storing at least one instruction set for using an intelligent agent system to execute target tasks; as well as At least one processor is communicatively connected to the at least one storage medium, wherein, when the task execution system is running, the at least one processor reads the at least one instruction set and implements the method as described in any one of claims 1-13 according to the instructions of the at least one instruction set.

15. A computer-readable non-volatile storage medium, wherein, The computer-readable non-volatile storage medium stores at least one instruction set, which, when executed by at least one processor, implements the method as described in any one of claims 1-13.

Images