Agent collaborative decision-making method and system based on hierarchical decoupling and adaptive switching

Abstract

The embodiment of the application provides a kind of based on layered decoupling and adaptive switching agent collaborative decision-making method and system, the agent collaborative decision-making method based on layered decoupling and adaptive switching, comprising: receiving global target, the global target is decomposed into executable subtask set, based on the executable subtask set to microcosmic execution layer sends subtask execution instruction;Wherein, the subtask execution instruction includes: task ID, target, constraint and expected output format;Receive the subtask execution instruction, the subtask execution instruction is converted into execution event, and the execution event is distributed to corresponding microcosmic layer execution unit to be executed accordingly;Wherein, the microcosmic layer execution unit includes: the agent based on API call, the function or API interface based on LLM encapsulation, algorithm model for executing preset function.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence multi-agent cooperative technology and robust control, specifically to an agent cooperative decision-making method and system based on hierarchical decoupling and adaptive switching. Background Technology

[0002] With the development of artificial intelligence technology, especially the rapid development of large language models and reinforcement learning, large model intelligent agents, as a new type of intelligent decision-making and execution entity, have shown great application potential in many complex fields such as autonomous driving, industrial automation, intelligent logistics, military command, intelligent customer service, and automated programming. These large model intelligent agent systems are typically composed of one or more LLM-driven agents that collaboratively complete complex task objectives through complex logical chains, tool calls, and interactions with the environment.

[0003] However, in practical applications, existing large-scale intelligent agent collaborative decision-making and control methods face several severe challenges, especially in environments with high uncertainty and strong dynamic changes, where their performance is often unsatisfactory. Summary of the Invention

[0004] To address the existing technical problems, this invention provides a method and system for collaborative decision-making of intelligent agents based on hierarchical decoupling and adaptive switching, which specifically relates to achieving efficient and robust collaboration of intelligent agents based on large language models in complex dynamic environments through hierarchical decoupling and adaptive switching of decision paradigms.

[0005] To achieve the above objectives, the technical solution of this invention is implemented as follows: Firstly, a collaborative decision-making method for intelligent agents based on hierarchical decoupling and adaptive switching is provided, including: The system receives a global objective, decomposes the global objective into a set of executable subtasks, and sends subtask execution instructions to the micro-execution layer based on the set of executable subtasks; wherein, the subtask execution instructions include: task ID, objective, constraints, and expected output format; The system receives the subtask execution instruction, converts the subtask execution instruction into an execution event, and assigns the execution event to the corresponding micro-level execution unit for execution. The micro-level execution unit includes: an agent based on API calls, a function or API interface based on LLM encapsulation, and an algorithm model that executes a preset function.

[0006] Secondly, a collaborative decision-making system for intelligent agents based on hierarchical decoupling and adaptive switching is provided, including: The macro-planning layer receives the global objective, decomposes the global objective into a set of executable subtasks, and sends subtask execution instructions to the micro-execution layer based on the set of executable subtasks; wherein, the subtask execution instructions include: task ID, objective, constraints, and expected output format; The micro-execution layer includes multiple micro-execution units, which receive the subtask execution instructions, convert the subtask execution instructions into execution events, and allocate the execution events to the corresponding micro-execution units for execution. The micro-execution units include: an agent based on API calls, a function or API interface based on LLM encapsulation, and an algorithm model that executes a preset function.

[0007] The above embodiments provide a collaborative decision-making method and system for intelligent agents based on hierarchical decoupling and adaptive switching, proposing a technical concept of hierarchically dividing the decision-making and execution processes. By decoupling decision-making and execution, the decision layer can provide a stable, high-level task framework and constraints, effectively constraining the behavioral space of the execution layer agents and preventing them from losing sight of the global objective in local optimization. Furthermore, the decision layer is responsible for handling resource conflicts and collaboration issues across large model agents, ensuring optimal resource allocation and task distribution through global optimization algorithms, avoiding inefficiencies caused by local agent competition. After receiving macro-level instructions from the decision layer, the execution layer has a certain degree of autonomy to make decisions and adjustments within a local scope to adapt to real-time environmental changes or tool calls. This design enables large model intelligent agent systems to maintain the continuity of global tasks while exhibiting rapid adaptability to local changes when facing complex, dynamic, and uncertain environments. In the application of large model intelligent agent systems, the decision layer refers to the macro-level planning layer, and the execution layer refers to the micro-level execution layer; the two interact and transmit instructions through a clear interface.

[0008] The agent collaborative decision-making system based on hierarchical decoupling and adaptive switching provided in the above embodiments belongs to the same concept as the corresponding agent collaborative decision-making method embodiments based on hierarchical decoupling and adaptive switching, and thus has the same technical effect as the corresponding agent collaborative decision-making method embodiments based on hierarchical decoupling and adaptive switching, which will not be repeated here. Attached Figure Description

[0009] Figure 1 This is a flowchart of an agent collaborative decision-making method based on hierarchical decoupling and adaptive switching in one embodiment.

[0010] Figure 2 This is an architecture diagram of an intelligent agent collaborative decision-making system based on hierarchical decoupling and adaptive switching in one embodiment. Detailed Implementation

[0011] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0012] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to limit the ways in which the invention may be implemented. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0013] In the following description, the expression “some embodiments” is used, which describes a subset of possible embodiments. However, it should be understood that “some embodiments” can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0014] In response to the numerous severe challenges faced by collaborative decision-making and control methods for large-scale intelligent agents, particularly the unsatisfactory performance in environments with high uncertainty and strong dynamic changes, the inventors of this application have conducted the following research and analysis.

[0015] First, large-scale model agents with a single decision-making paradigm lack robustness. Many existing large-scale model agent systems tend to adopt a single decision-making paradigm. For example, agents purely based on long-chain reasoning and planning-execution can achieve high efficiency and logical consistency when the environment is stable and the task is clear. However, when faced with sudden events, unexpected environmental changes, missing information, or the "illusion" phenomenon of large-scale model reasoning paths, their pre-planned plans may quickly fail, leading to deviations in agent behavior, getting stuck in infinite loops, or task failure. They are unable to effectively cope with abnormal situations and exhibit poor robustness. On the other hand, purely reactive agents (such as those based on direct responses to a few-shot prompting) excel in local agility and can quickly respond to environmental changes, but they lack a global perspective and long-term planning capabilities. They are prone to getting trapped in local optima and may even deviate from the overall goal in complex tasks, making it difficult to guarantee global consistency. Furthermore, they may generate inaccurate or irrelevant responses without contextual support.

[0016] Secondly, balancing the contradiction between global consistency and local agility in large-scale model agents is difficult. In collaborative scenarios involving large-scale model agents, the agent system needs to coordinate its reasoning and actions to achieve a common macro-level goal, requiring global consistency in decision-making. However, complex dynamic environments demand that agents be able to react instantly and agilely to local changes. Traditional collaborative control methods often struggle to find a dynamic and effective balance between global optimization and local response. Overemphasizing global planning may lead to sluggish agent responses, an inability to adapt to rapid changes, and even logical drift during long-term reasoning; while overemphasizing local agility may disrupt global coordination, resulting in task failure or poor quality of generated content. For example, in a project where multiple large-scale model agents collaborate to complete a complex software development, the global architecture planning agent needs to ensure consistency in code structure and module interfaces. However, if a development agent encounters unforeseen library conflicts or API changes while implementing a specific function, a lack of immediate response capability may hinder the entire development process or even introduce new systemic errors.

[0017] Furthermore, large model agents are ill-suited to high-uncertainty scenarios. Current large model agent systems typically perform well in relatively structured or predictable environments (such as sandbox environments or specific datasets), but in the real world, environments are often noisy, information-incomplete, and subject to unpredictable events. Existing methods often lack effective mechanisms for situational awareness, risk assessment, and decision adjustment when dealing with these high uncertainties, resulting in large model agent systems failing to meet practical requirements in terms of task success rate and reliability under high-uncertainty scenarios. For example, in LLM-based intelligent customer service scenarios, when faced with complex questions from users that are vague, emotional, or contain multiple intents, a single-paradigm agent struggles to accurately understand the user's intent while simultaneously ensuring a smooth dialogue experience and efficient problem-solving.

[0018] Therefore, the field of collaborative decision-making and control of large model agents urgently needs a new method and system that can effectively integrate the advantages of planning and reaction to achieve high robustness, high agility, and high task success rate of large model agents in complex dynamic environments. This would overcome the technical bottlenecks of traditional single-paradigm large model agents in terms of robustness, balance between global consistency and local agility, and adaptability to high uncertainty scenarios.

[0019] In view of this, the inventors of this application propose a collaborative decision-making control method and system for large model intelligent agents based on hierarchical decoupling and adaptive switching, to solve the problems of poor robustness, difficulty in balancing global consistency and local agility, and insufficient adaptability to high uncertainty scenarios in existing single-paradigm large model intelligent agents under complex dynamic environments. This invention introduces a vertical hierarchical decoupling mechanism to decompose complex tasks into two independent yet collaborative layers: macro-planning and micro-execution. Based on this, an environmental situational awareness module drives the decision-making paradigm to adaptively switch and merge between "Forward Planning-Execution (FLPE)" and "Contextual Response (CR)," thereby establishing a dynamic balance between global optimization and immediate response, significantly enhancing the robustness, agility, and task success rate of large model intelligent agent systems in high uncertainty scenarios.

[0020] Please see Figure 1 The following is an embodiment of the intelligent agent collaborative decision-making method based on hierarchical decoupling and adaptive switching, which includes the following steps.

[0021] S101, Receive the global target, decompose the global target into a set of executable subtasks, and send subtask execution instructions to the micro-execution layer based on the set of executable subtasks; wherein, the subtask execution instructions include: task ID, target, constraints and expected output format.

[0022] S103, receive the subtask execution instruction, convert the subtask execution instruction into an execution event, and allocate the execution event to the corresponding micro-level execution unit for corresponding execution; wherein, the micro-level execution unit includes: an agent based on API calls, a function or API interface based on LLM encapsulation, and an algorithm model for executing preset functions.

[0023] The agent collaborative decision-making based on hierarchical decoupling and adaptive switching provided in the above embodiments proposes a technical concept of hierarchically separating the decision-making and execution processes. By decoupling decision-making and execution, the decision layer can provide a stable, high-level task framework and constraints, effectively constraining the behavioral space of the execution layer agents and preventing them from losing sight of the global objective in local optimization. Furthermore, the decision layer is responsible for handling resource conflicts and collaboration issues across large model agents, ensuring optimal resource allocation and task distribution through global optimization algorithms, avoiding inefficiencies caused by competition among local agents. After receiving macro-level instructions from the decision layer, the execution layer has a certain degree of autonomy to make decisions and adjustments within a local scope to adapt to real-time environmental changes or tool calls. This design enables large model agent systems to maintain the continuity of global tasks while exhibiting rapid adaptability to local changes when facing complex, dynamic, and uncertain environments. In the application of large model agent systems, the decision layer refers to the macro-level planning layer, and the execution layer refers to the micro-level execution layer; the two interact and exchange instructions through a clear interface.

[0024] Please refer to the following: Figure 2 This application proposes a vertically layered decoupling architecture that decouples the decision-making and execution processes of a large model intelligent agent system into a macro-planning layer 201 and a micro-execution layer 203.

[0025] A global objective refers to the ultimate goal set by the system at the macro level. It is typically unique and overarching, serving as the final target of the decomposed set of executable subtasks. Global objectives can be mainly classified into three categories: data-driven, operational, and strategic. Different types of global objectives have different characteristics. The agent-based collaborative decision-making method based on hierarchical decoupling and adaptive switching further includes employing corresponding checking strategies for data-driven, operational, and strategic global objectives to ensure that each global objective contains preset components, thereby ensuring the validity of the global objective input.

[0026] For data-driven global objectives (such as "optimize the production schedule of a product line to improve efficiency by 20% and reduce costs by 15%), it can be determined whether the currently input global objective contains the three major components of "quantifiable indicators, time nodes, and benchmark reference values," and whether the result meets the requirements of "quantifiable indicators being statistically verifiable, time nodes being actionable, and benchmark values ​​being traceable." If not (e.g., only "improving efficiency" is stated without specifying concrete values), the macro-planning layer 201 returns a supplementary request to the objective input party, clearly indicating the quantitative dimensions and reference standards that need to be supplemented, until the requirements are met. A qualified data-driven global objective typically needs to include the following indicator-related content: quantifiable target value, benchmark value, time frame, and core influencing factors (such as equipment utilization rate and manpower quota in the production schedule).

[0027] For operational global objectives (such as "completing the requirements analysis and architecture design of a complex software development project"), we can determine whether they include the three main components of "operation scope, execution boundary, and acceptance criteria," and whether the results satisfy the requirements of "unambiguous operation scope, clear and non-overlapping boundaries, and verifiable acceptance criteria." If they do not meet these requirements (e.g., ambiguous boundaries leading to multi-agent task conflicts in the decomposed executable sub-task set), the macro-planning layer 201, in conjunction with the domain agent, will define the boundaries and output a clear operation scope specification. A qualified operational global objective typically includes the following indicators: operation object, execution step framework, boundary definition, acceptance criteria, and responsible entity.

[0028] For strategic global objectives (such as "building a highly robust multi-agent collaborative decision-making system"), it can be determined whether they include the three main components of "strategic direction, phased objectives, and evaluation dimensions," and whether the results meet the criteria of "clear strategic direction, decomposable phased objectives, and quantifiable evaluation dimensions." If they do not meet these criteria (e.g., the lack of phased objectives makes implementation impossible), the macro-planning layer 201 breaks down the initial phased objectives, provides feedback to the target input party for confirmation and adjustment, and uses human-computer interaction until a qualified strategic global objective is obtained. A qualified strategic global objective typically includes the following indicators: core strategic positioning, phased milestones, evaluation indicators (such as robustness and collaborative efficiency), and resource allocation requirements.

[0029] As the "brain" of the large-scale intelligent agent collaborative decision-making system, the macro-planning layer 201's primary responsibility is to receive and deeply analyze high-level input information from users, upper-level business systems, environmental sensors, or other intelligent decision-making entities. After verifying and correcting the global goal based on the aforementioned global goal acquisition standards, it provides targeted decomposition strategies based on the type of global goal, ensuring that the decomposed sub-tasks can be correctly executed by the micro-execution layer 203. For example, for data-driven global goals, based on the three essential components they inevitably contain—"quantitative indicators, benchmark reference values, and time nodes"—it is broken down into three core sub-execution units: First, the data acquisition and benchmark analysis unit, responsible for collecting benchmark data and verifying data accuracy, corresponding to calling data query APIs or data processing algorithm models; second, the indicator decomposition and allocation unit, which breaks down the total quantitative goal by module and time (e.g., breaking down "efficiency improvement of 20%" into improvement ratios for each stage of each production line), corresponding to calling task allocation agents; and third, the process monitoring and calibration unit, responsible for tracking the achievement of indicators and timely calibrating deviations, corresponding to calling monitoring algorithms and feedback adjustment agents. For operational global goals, based on their three main components—"operational scope, execution boundaries, and acceptance criteria"—the objectives are broken down into process-oriented sub-execution units. For example, "software development requirements gathering and architecture design" is broken down into four sub-units: requirements gathering, requirements analysis, architecture selection, and architecture review. These are assigned to the research agent, analysis agent, architecture design agent, and review agent, respectively. Each sub-unit has clearly defined operational specifications and output requirements. For strategic global goals, based on their three main components—"strategic direction, phased goals, and evaluation dimensions"—the objectives are broken down into a set of phased sub-tasks. Each phased task is further broken down into strategic implementation units and effect evaluation units. For example, "building a highly robust collaborative system" is broken down into three phased tasks: "architecture building," "agent debugging," and "robustness testing." Each phased task is further broken down into specific execution units to ensure that strategic goals are implemented at every level. In addition to the decomposition strategies mentioned above, the macro planning layer 201 also receives other input information such as task priority and available resource constraints. After verification, it finally sends subtask execution instructions to the micro execution layer 203 based on the set of executable subtasks. The macro planning layer 201 is usually composed of core large model agents (or agent clusters) and focuses on high-level semantic reasoning and long-term goal management.

[0030] Micro-Execution Layer 203: This layer consists of multiple atomic execution units, which can be specialized large-scale model agents, function call tools, API interfaces, or small models encapsulating specific algorithms. They are responsible for receiving instructions from the macro-planning layer and translating them into specific, fine-grained operational instructions. These atomic execution units focus on real-time environmental awareness and local action execution, exhibiting high agility and immediate responsiveness to environmental changes. For example, when the macro-planning layer assigns a "material selection agent" to evaluate the performance of a composite material, the micro-execution layer is responsible for a series of specific operations, such as calling the material database API, running the finite element analysis model, or generating text reports with specific parameters. This vertical decoupling ensures the independence of large-scale model agent decisions and operations at different levels of abstraction, while also facilitating information exchange through clear interfaces, enabling the system to maintain a global strategic perspective while also considering local tactical flexibility.

[0031] Optionally, a technical concept of Dual-Paradigm Fusion and Adaptive Switching is further proposed.

[0032] To further enhance the adaptability of the large model intelligent agent system to complex dynamic environments, the large model intelligent agent system incorporates an environmental situation awareness module and drives the decision-making paradigm based on real-time monitored environmental indicators to adaptively switch and fuse between two complementary modes.

[0033] The agent-based collaborative decision-making method based on hierarchical decoupling and adaptive switching further includes: monitoring key environmental indicators and switching between a forward-looking planning-execution paradigm and a situational response paradigm based on these indicators. Simultaneously, weight coefficients are assigned to each key environmental indicator. , , , The determination follows clear rules and can be calculated by the Analytic Hierarchy Process (AHP). Combined with the priority of the impact of each key environmental indicator on the environmental status assessment, it is determined by a three-step process of constructing a judgment matrix, consistency check, and weight vector calculation. Alternatively, the baseline weight can be preset according to the actual application scenario, and then dynamically iterated and calibrated in combination with the historical switching effect to ensure that the weight allocation fits the decision-making needs of the specific scenario and avoids the assessment deviation caused by subjective assignment. Environmental Situational Awareness Module: Used to continuously monitor key environmental indicators in real time, including but not limited to environmental entropy, task complexity, sudden event indicators, and the confidence of the large model intelligent agent's own reasoning. The specific implementation schemes for monitoring each indicator are as follows: (1) Monitoring environmental entropy requires the necessary hardware support such as environmental sensors, data acquisition modules, and communication status monitoring equipment. It detects the external data stream transmission rate, data fluctuation amplitude, environmental parameters collected by sensors (such as temperature and signal strength), communication channel packet loss rate and delay fluctuation, etc. The Shannon entropy algorithm is used to perform probability distribution statistics on the collected multi-dimensional data, and H = -ΣP(x)log2P(x) formula, calculate the environmental entropy (P(x) is the probability value of each data dimension); (2) The task complexity needs to be semantically decomposed by the task parsing unit of macro planning layer 201, extract the core information such as operation steps, tool call type, number of collaborative agents, and constraints, and adopt the multi-dimensional scoring method to obtain the quantitative value of task complexity by weighted summation from four dimensions: number of operation steps, tool call complexity, number of collaborative agents, and severity of constraints; (3) Sudden event index, sudden events are monitored by anomaly detection algorithm through micro execution layer feedback, external system returned data, and user input instructions. When the data deviates from the normal range, the execution deviation exceeds the threshold, an emergency instruction or system interruption occurs, the sudden event will be detected. When it is determined to be a sudden event, the sudden event index adopts a combination of binary quantification and level scoring. When it does not occur, it is 0. After it occurs, it is divided into 1-5 levels according to the urgency and scope of impact, corresponding to different quantification values; (4) The confidence of the large model intelligent agent's own reasoning. Here, the large model intelligent agent specifically refers to the core large model intelligent agent (or intelligent agent cluster) in the macro planning layer 201 that receives the global goal and is responsible for task decomposition and paradigm decision-making. Its reasoning confidence is obtained by weighting three parameters: output probability distribution, reasoning path consistency, and matching degree with the known knowledge base. The output probability is the probability value of the final result. The consistency is compared by the overlap of multiple reasoning paths. The matching degree is obtained by semantic similarity comparison. The comprehensive analysis of these indicators provides an objective and quantitative basis for decision paradigm switching. Specifically, the environmental situation awareness module will comprehensively calculate an environmental state assessment value. For example, these metrics can be weighted and aggregated to obtain: in, Represents environmental entropy. Indicates task complexity. Indicators representing sudden events, while This represents the confidence level of the large model agent's inference. This reflects its uncertainty. These are the corresponding weighting coefficients, reflecting the importance of each indicator to the environmental status assessment. When When the threshold is exceeded, the system will tend to switch to a context-based response paradigm. The preset threshold is typically a dynamic value, determined using an "initial static preset + dynamic iterative calibration" model. In an optional example, the formula and rules for determining the preset threshold are as follows: the initial threshold can be determined through statistical analysis based on historical data, task requirements, and security boundaries of the specific application scenario; the formula can be "initial threshold = historical mean of the indicator ± 2 standard deviations," with adjustments made based on scenario tolerance requirements; dynamic iterative calibration combines data such as historical paradigm switching effects, environmental indicator fluctuation patterns, and task execution success rates, using a gradient descent algorithm to iteratively optimize the threshold, ensuring that the threshold adapts to dynamic environmental changes.

[0034] Forward-Looking Planning-Execution (FLPE) Paradigm: Under stable or low-entropy conditions, the system automatically switches to and adopts the FLPE paradigm. This paradigm is efficiency-maximizing and long-term goal-achieving, leveraging predictive models of future states and the chain-reasoning capabilities of large-scale intelligent agents to drive process-driven planning and execution. For example, in intelligent contract review with known business processes, the agent can pre-plan each review step, which external legal knowledge base to invoke, and under what conditions risk assessment will be conducted, ensuring maximum efficiency and accuracy of the review process through algorithm optimization. This paradigm fully leverages the advantages of large-scale intelligent agents in logical reasoning and complex planning.

[0035] Contextual Reactive (CR) Paradigm: When the environmental situation awareness module detects high environmental entropy, a sharp increase in task complexity, or sudden dynamics (such as sudden user interruption, external system crash, or a sharp drop in the inference confidence of a large model agent), the system will quickly and automatically switch to the CR paradigm. This paradigm prioritizes immediate response and local behavior correction to ensure the agility and security of the agent system. For example, when a large model agent is generating complex code, if a compilation error or an urgent modification instruction from the user is detected, the CR paradigm will immediately drive the agent to pause the current generation, perform error checking, or make local code corrections based on the new instruction, without waiting for the completion of the original complex planning chain. This immediate response mechanism can effectively avoid potential dangers, reduce the impact of "illusions," and maintain the robustness of the agent system in the face of uncertainty.

[0036] In the above embodiments, a stable decision-making and execution framework for large-scale intelligent agents is constructed through vertical hierarchical decoupling. Furthermore, a dynamic balance is established between system efficiency, global consistency, local agility, and security through dual-paradigm fusion and adaptive switching. This enables the large-scale intelligent agent system to intelligently adjust its behavioral strategies according to real-time changes in the environment, thereby exhibiting superior adaptability, robustness, and a higher task success rate in high-uncertainty scenarios.

[0037] The agent collaborative decision-making and system based on hierarchical decoupling and adaptive switching provided in this application's embodiments is fundamentally about decoupling the complex large-scale agent collaborative decision-making problem into tasks at different levels of abstraction. By dynamically adjusting the decision-making paradigm of the large-scale agent through real-time perception of the environmental state, it aims to achieve an optimal balance between efficiency, robustness, and agility. It possesses at least the following characteristics: 1. Hierarchical structure: Vertical hierarchical decoupling mechanism The hierarchical structure proposed in this invention, namely the vertical hierarchical decoupling mechanism, is the foundation for achieving efficient collaboration and robust decision-making in large-scale intelligent agent systems. This mechanism divides the decision-making and execution process of the entire large-scale intelligent agent system into a macro-planner layer and a micro-execution layer. The two layers interact and transmit instructions through a clear interface, thereby achieving vertical decoupling of decision-making tasks in terms of structure.

[0038] 1.1 Macro-Planning Layer The macro-planning layer functions as the "brain" of the large-scale model agent system. Its main responsibilities include long-term task decomposition, global resource pre-allocation, and high-level strategy formulation. This layer does not directly participate in specific fine-grained tool calls or physical operations, but focuses on abstract task management and global optimization. For example, in a complex scientific experiment planning process involving multiple large-scale model agents, the macro-planning layer receives the research objective (such as "exploring the synthetic route of novel catalysts") and decomposes it into a series of executable sub-tasks (such as "literature review," "molecular simulation," and "experimental design"). Simultaneously, it pre-allocates resources to the large-scale model agents based on global objectives (such as maximizing discovery efficiency and minimizing experimental costs) and plans the logical sequence and time windows for these sub-tasks. The macro-planning layer operates through an abstract set of task instructions. Commands are passed to the micro-execution layer, and each instruction contains high-level semantic information such as task ID, target, constraints, and expected output format.

[0039] The key advantage of this layer lies in its guarantee of global logical coherence. By planning at a higher level of abstraction, the macro-planning layer can effectively guide the actions of large model agents in the micro-execution layer, preventing the "logical drift" problem (i.e., deviating from the initial goal or generating incoherent sub-goals during task execution) that may occur in long-chain reasoning, ensuring that the final task result is highly consistent with the initial intention. By providing a stable, high-level task framework and constraints, the macro-planning layer can effectively constrain the behavioral space of micro-execution layer agents, preventing them from losing sight of the global goal in local optimization. In addition, the macro-planning layer is also responsible for handling resource conflicts and coordination issues across large model agents. Through global optimization algorithms, it ensures optimal resource allocation and task distribution, avoiding inefficiency caused by competition among local agents.

[0040] 1.2 Micro-Execution Layer The micro-execution layer consists of a series of atomic execution units, which act as the "hands and feet" of the large-scale intelligent agent system. They are responsible for receiving instructions from the macro-planning layer and translating them into specific, fine-grained physical operations or tool calls. These execution units can be: Specialized large model agents: such as an agent that focuses on specific API calls, or an agent that is good at executing specific algorithms.

[0041] Function Call Tools: External functions or API interfaces encapsulated by an LLM (Library Management System) to perform specific operations, such as database queries, code execution, and image processing.

[0042] Traditional AI models or algorithm modules, such as computer vision modules, speech recognition modules, and path planning algorithms, are encapsulated into atomic execution units for large model agents to call.

[0043] These micro-execution units place great emphasis on real-time environmental awareness and the accuracy and agility of local operations. For example, when the macro-planning layer issues the instruction to "execute a database query to obtain user order information," the "database query agent" or corresponding function call tool in the micro-execution layer will be responsible for constructing the SQL statement, connecting to the database, executing the query, and parsing the results. It will utilize its own perception capabilities (such as parsing the status codes returned by the database and handling query timeouts) to perceive the execution results in real time and adjust its subsequent operations based on the perception results to ensure the accuracy and security of the operation.

[0044] The core advantage of this layer lies in its vertical decoupling of decision-making and specific operations within a large model agent. By decomposing complex tasks into independently executable atomic operations, with the micro-execution layer responsible for their implementation, the macro-planning layer can focus on strategic questions of "what to do" and "when to do it," while the micro-execution layer focuses on tactical questions of "how to do it." This decoupling prevents the macro-planning layer from being bogged down in tedious execution details, thereby improving the efficiency and maintainability of the entire large model agent system. Simultaneously, upon receiving macro-instructions, the micro-execution layer possesses a degree of autonomy to make decisions and adjustments within a local scope, adapting to real-time environmental changes or the results of tool calls. This design enables the large model agent system to maintain the consistency of global tasks while demonstrating rapid adaptability to local changes when facing complex, dynamic, and uncertain environments.

[0045] 2. Core Mechanism: Dual Paradigm Fusion and Adaptive Switching The second core innovation of this invention lies in its dual-paradigm fusion and adaptive switching mechanism. This mechanism enables large-scale model intelligent agent systems to dynamically switch between two decision-making paradigms: "Forward Planning-Execution (FLPE)" and "Contextual Response (CR)," to cope with environments of varying complexity and uncertainty. These two paradigms are functionally complementary. The FLPE paradigm primarily focuses on maximizing efficiency and achieving long-term goals, while the CR paradigm emphasizes robustness and safety. The combination of the two provides large-scale model intelligent agent systems with comprehensive adaptability.

[0046] 2.1 Forward-looking Planning-Execution (FLPE) Paradigm The FLPE paradigm is the primary operating mode for large-scale intelligent agent systems in relatively stable, predictable, or low-entropy environments. Under this paradigm, large-scale intelligent agent systems fully leverage their powerful logical reasoning capabilities and predictive models for process-driven planning and execution. This means that the agent, based on current known information and predictions of future states, formulates detailed, step-by-step action plans (e.g., through CoT, ToT, and other thought chain technologies) and strictly executes these plans. For example, in a pre-defined intelligent question-answering process, when a user's question is explicit and within the knowledge base, the agent under the FLPE paradigm will pre-plan the steps of retrieving the knowledge base, analyzing relevant information, and synthesizing and generating an answer, ensuring the highest efficiency and accuracy of the entire question-answering process through chained reasoning.

[0047] The advantage of the FLPE paradigm lies in its ability to optimize resource allocation and efficiently complete tasks. Through advance planning, large model agents can avoid ad-hoc decision-making overhead, reduce resource waste (such as unnecessary LLM calls), and utilize global information for better decision-making. It provides large model agents with a clear and explicit action blueprint, ensuring the system exhibits superior performance and efficiency in goal-oriented tasks. However, its limitation lies in its low sensitivity to environmental changes. If unexpected and drastic environmental fluctuations occur, or if the large model agent experiences "illusions" during inference, the pure FLPE paradigm may lead to delayed or erroneous decisions due to plan failure.

[0048] 2.2 Contextual Response (CR) Paradigm The CR paradigm is an emergency response mode for large model intelligent agent systems when facing environments with high uncertainty, high environmental entropy, or sudden events. Under this paradigm, the large model intelligent agent system will pause or partially adjust the long-term planning of the FLPE paradigm, prioritizing immediate response and local behavioral correction. The CR paradigm emphasizes rapid perception of the current environmental situation and immediate decision-making based on the current context to ensure the agility and security of the intelligent agent system. For example, in a real-time code writing scenario assisted by a large model intelligent agent, when the code editor suddenly reports an error or the user enters an emergency repair command, the CR paradigm will immediately trigger the agent to perform error diagnosis or perform local code correction based on the new command, without waiting for the macro-planning layer to re-complete the entire code generation process.

[0049] The advantages of the CR paradigm lie in its rapid response capability and robustness to unexpected events. It ensures that large model agents prioritize their own safety and task continuity in complex, dangerous, or unpredictable environments, effectively reducing the negative impact of LLM "illusions." By focusing on local, short-term behavioral corrections, the CR paradigm effectively addresses the challenges of highly dynamic environments, preventing agent systems from deadlocking or crashing. While the CR paradigm may not reach global optimum, it plays an irreplaceable role in ensuring the safety of agent systems and the continuity of tasks, especially when rapid correction of LLM generation errors or adaptation to sudden external changes is required.

[0050] By dynamically fusing and switching between the FLPE and CR paradigms, this invention achieves a delicate balance between efficiency and robustness. In stable environments, large model agents prioritize efficiency and long-term planning; in turbulent environments, they prioritize safety and agility, thus enabling large model agent systems to maintain high performance and high reliability in various complex dynamic scenarios.

[0051] 3. Switching criteria: Environmental situation awareness module and quantified entropy index In this invention, the switching between the large model agent decision-making paradigm (FLPE and CR) is not random or based on a fixed time, but is dynamically driven by a dedicated environmental situation awareness module based on a series of objective and quantifiable environmental indicators. This data- and model-based switching mechanism significantly improves the objectivity, accuracy, and adaptability of the switching decision.

[0052] The environmental situation awareness module monitors and comprehensively analyzes various key indicators in real time to fully assess the complexity and uncertainty of the current environment. These indicators include, but are not limited to: Environmental entropy: This is a core indicator for measuring environmental uncertainty and randomness. A higher environmental entropy value indicates a greater amount of information in the environment, more drastic changes, and greater difficulty in prediction. For example, environmental entropy can be quantified by calculating the Shannon entropy of external data streams (such as sensor data and user input streams), the number of dynamic targets in the environment and their rate of change in speed, the packet loss rate or latency fluctuations of communication channels, and the degree of conflict in the perception of the external environment by large-scale intelligent agents. When environmental entropy exceeds a preset threshold, it indicates that the environment has entered a state of high uncertainty, and the intelligent agent system tends to switch to the CR paradigm.

[0053] Task complexity: Assess the inherent difficulty of the current subtask and the capabilities required of large model agents. For example, task complexity can be quantified by evaluating the amount of input information that needs to be processed, the number of large model agents that need to coordinate, the size of the decision space, the complexity of the required tool calls, and the urgency of task completion. When task complexity increases rapidly and exceeds the processing capacity of the FLPE paradigm, it may trigger a switch to the CR paradigm.

[0054] Outbreak Indicators: These indicators are used to detect and identify unexpected anomalies or emergencies in the environment. For example, anomaly detection algorithms can identify significant outliers returned by external systems, detect urgent user commands, unexpected interruptions in the inference process within a large model agent, or discover local execution results that are severely inconsistent with the macro-plan. Once an outbreak is identified, the agent system will immediately switch to the CR paradigm for response.

[0055] Large Model Agent Inference Confidence: In the FLPE paradigm, large model agents rely on their internal inference chains and predictive capabilities for planning. The environmental situation awareness module continuously evaluates the confidence of these inference processes. For example, this can be assessed by analyzing the probability distribution of the LLM output, the consistency of multiple inference paths, or the degree of matching with known factual knowledge. When the confidence of the large model agent's inference output drops below a preset threshold, it indicates that the LLM's ability to predict or understand future states has weakened, posing a risk of "illusion," and the environment may become unpredictable. In this case, switching to the CR paradigm should be considered.

[0056] Dynamic rate of change: The rate at which key parameters in the monitoring environment (such as key task variables, external data indicators, and the internal state of large model agents) change over time. A high dynamic rate of change indicates an unstable environment that requires a faster response.

[0057] These metrics are comprehensively evaluated using a fusion algorithm to generate a dynamic "environmental state vector" as follows: The intelligent agent system pre-defines a set of decision rules and thresholds. When the environmental state vector meets specific conditions, the environmental situational awareness module issues a switching command, driving the large-scale intelligent agent system to switch from the FLPE paradigm to the CR paradigm, and vice versa. The switching decision function can be a piecewise function: in, This is the comprehensive environmental status assessment value mentioned above. and It is a preset switching threshold, and is usually set. By introducing a hysteresis effect, frequent switching is prevented. This quantitative and adaptive switching strategy enables large-scale intelligent agent systems to accurately adjust their decision-making behavior according to the real needs of the environment, thereby significantly improving the system's adaptability and robustness in high-uncertainty scenarios while ensuring efficiency.

[0058] 4. Closed-loop control: Feedback-driven execution correction mechanism The feedback-driven execution correction mechanism introduced in this invention, particularly for the behavior of large model agents under the CR (Contextual Response) paradigm, forms a local, dynamic closed-loop control system. This mechanism allows the local results and experiences generated by the large model agent during immediate reactions to effectively influence macro-planning, achieving small-scale adaptive optimization and further enhancing the system's robustness and adaptability. Under the CR paradigm, the large model agent primarily performs immediate responses and local behavior corrections. The results of these immediate responses include: successfully avoiding external API call failures; timely correction of syntax errors in LLM-generated code; adjusting dialogue content based on urgent user instructions; local clarification of ambiguous user intentions; and the success / failure status returned by the tool after executing a specific tool call, along with detailed output information.

[0059] These results are considered valuable feedback. This feedback is no longer a simple status report, but rather contains data that provides "fine-tuning suggestions" or "validation" for the current macro-planning. Specifically, when a large model agent is in CR mode and performs a series of local corrective actions, the results are fed back to the macro-planning layer through the micro-execution layer. For example, when a large model agent is executing a complex data processing flow set by the macro-planning layer, it switches to CR mode due to an abnormal input data format and successfully calls a data cleaning tool to repair local data. At this time, its new local data processing path information, the additional time or computing resources consumed during data cleaning, and the data quality report after cleaning are all uploaded as feedback data. The macro-planning layer does not completely overturn the original large-scale plan, but uses this local feedback to make fine-tuning adjustments to the current macro-planning without affecting the global objectives. For example, it can update the parameters of a sub-module in the data processing flow, adjust the processing strategy of subsequent agents for this type of data, or revise the expected task completion time.

[0060] The advantage of this feedback-driven execution correction mechanism is that: The fusion of local optimization and global coordination allows large-scale agents in the micro-execution layer to make local corrections to macro-planning without compromising global consistency. This "small-scale dynamic closed loop" avoids the costly and time-consuming large-scale replanning or global reasoning of large-scale agents every time a local problem occurs, thereby improving the response efficiency and resource utilization of the agent system.

[0061] Enhancing the adaptability of macro-planning: By continuously receiving actual execution feedback from the CR model, the macro-planning layer can gradually learn and adapt to more complex environmental dynamics. For example, if a particular API call fails frequently, the macro-planning layer may adjust the API calling strategy based on feedback or try using alternative tools.

[0062] Enhancing the robustness of intelligent agent systems: Even if macro-planning fails to fully anticipate all potential complexities, the immediate reactions and feedback correction mechanisms of large model agents under CR mode can effectively compensate for this deficiency, enabling intelligent agent systems to maintain functionality and task continuity when facing unknown and sudden situations, effectively reducing the risks brought about by the LLM "illusion".

[0063] Data-driven decision-making iteration: This feedback data can also be stored and analyzed for subsequent offline training and optimization of macro-planning algorithms and large model agent inference strategies, forming a longer-term learning loop and further improving the overall intelligence level of the large model agent system.

[0064] In summary, by constructing a stable structure through a vertical hierarchical decoupling mechanism, achieving flexible strategy adjustment through dual-paradigm fusion and adaptive switching mechanisms, and supplementing it with switching criteria based on quantitative indicators and a feedback-driven closed-loop correction mechanism, this invention provides a comprehensive, in-depth, robust, and efficient solution for collaborative decision-making control of large-scale intelligent agents, enabling it to exhibit outstanding performance in complex, dynamic, and highly uncertain application scenarios.

[0065] Compared to existing technologies, the agent collaborative decision-making method and system based on hierarchical decoupling and adaptive switching provided in this application have at least the following differences: First, this invention pioneers a hierarchical decoupling architecture for large-scale model intelligent agents and a dual-paradigm adaptive switching mechanism: overcoming the limitations of traditional single-paradigm intelligent agents, such as insufficient robustness and difficulty in balancing global consistency and local agility. By introducing vertical hierarchical decoupling, this invention separates macro-planning from micro-execution and creatively designs a dual-paradigm adaptive switching mechanism based on environmental situation awareness: "Forward-Looking Planning-Execution (FLPE)" and "Contextual Response (CR)." This mechanism can dynamically adjust decision-making strategies based on quantitative indicators such as environmental entropy, task complexity, unexpected events, and LLM confidence, achieving an optimal balance of efficiency, robustness, and agility, fundamentally improving the adaptability of large-scale model intelligent agent systems in complex dynamic environments.

[0066] Second, this invention proposes a method for quantifying environmental situational awareness and agent reasoning confidence assessment: For high-uncertainty scenarios, this invention designs an environmental situational awareness module that constructs a comprehensive environmental state assessment value by fusing multi-source information (such as Shannon entropy, task complexity models, and anomaly detection indicators). Simultaneously, it innovatively combines LLM output probability distribution or negative entropy to quantify the reasoning confidence of the large model agent itself. These quantitative indicators provide objective and accurate basis for the adaptive switching of decision-making paradigms, significantly enhancing the system's response speed to unpredictable events and its ability to avoid the "illusion" risk of LLM.

[0067] Third, a feedback-driven closed-loop control mechanism for local execution correction is constructed: Under the CR paradigm, this invention establishes a closed-loop mechanism that feeds back local correction results from the micro-execution layer to the macro-planning layer. This mechanism allows the macro-planning layer to finely adjust and optimize its own parameters using empirical data from local, real-time responses (such as actual resource consumption and error correction results) without interrupting the global task. This small-scale dynamic adaptive capability effectively compensates for the shortcomings of macro-planning, improves the system's robustness and recovery capability in the face of local unexpected situations, and provides a data foundation for long-term learning and iterative optimization.

[0068] Fourth, it significantly improves the success rate and reliability of large model agents in high-uncertainty scenarios: Combining the stability of layered decoupling, the flexibility of dual-paradigm switching, and the accuracy of quantified perception, this invention comprehensively enhances the fault tolerance and adaptability of large model agent systems from the perspectives of architecture and operating mechanism. It enables large model agent systems to efficiently complete complex tasks, while maintaining functionality, stability, and a high task success rate even when facing high-uncertainty challenges such as unexpected events, incomplete information, or potential "illusions" of LLM (Local Model Modules). This greatly expands the deployment scope and reliability of large model agents in complex real-world applications.

[0069] In another aspect, this application also provides an agent collaborative decision-making system based on hierarchical decoupling and adaptive switching, for executing the agent collaborative decision-making method based on hierarchical decoupling and adaptive switching described in the foregoing embodiments.

[0070] The implementation process of the agent collaborative decision-making method based on hierarchical decoupling and adaptive switching can be as described in any of the previous embodiments. The embodiments can be combined arbitrarily without contradiction, and will not be repeated here.

[0071] In another aspect, this application also provides a computer-readable storage medium, such as a memory including an executable program, which is executed by a processor to complete the steps of the agent cooperative decision-making method based on hierarchical decoupling and adaptive switching described in any embodiment of this application, and can achieve the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may include a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0072] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0073] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a computer, server, network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0074] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A collaborative decision-making method for intelligent agents based on hierarchical decoupling and adaptive switching, characterized in that, include: The system receives a global objective, decomposes the global objective into a set of executable subtasks, and sends subtask execution instructions to the micro-execution layer based on the set of executable subtasks; wherein, the subtask execution instructions include: task ID, objective, constraints, and expected output format; The system receives the subtask execution instruction, converts the subtask execution instruction into an execution event, and assigns the execution event to the corresponding micro-level execution unit for execution. The micro-level execution unit includes: an agent based on API calls, a function or API interface based on LLM encapsulation, and an algorithm model that executes a preset function.

2. The agent collaborative decision-making method based on hierarchical decoupling and adaptive switching according to claim 1, characterized in that, Also includes: Monitor key environmental indicators and switch between a forward-looking planning-implementation paradigm and a situational response paradigm based on these key environmental indicators; The key environmental indicators include at least one of the following: environmental entropy, task complexity, sudden event indicators, and large model agent inference confidence.

3. The agent cooperative decision-making method based on hierarchical decoupling and adaptive switching according to claim 2, characterized in that, The switching between the forward-looking planning-implementation paradigm and the situational response paradigm based on the key environmental indicators includes: The environmental status assessment value is calculated by weighting and aggregating the key environmental indicators; the weighted aggregation formula is expressed as follows: ； in, For environmental entropy, For task complexity, For emergency indicators, For large model agents, inference confidence, These are the corresponding weighting coefficients; Based on environmental state assessment values, the intelligent agent collaborative decision-making system switches between a forward-looking planning-execution paradigm and a situational response paradigm by switching decision functions; the switching decision functions are as follows: in, Environmental status assessment value, and It is a preset switching threshold, and .

4. The agent collaborative decision-making method based on hierarchical decoupling and adaptive switching according to claim 1, characterized in that, The global objectives include data-driven global objectives, operational global objectives, and strategic global objectives; the method further includes: Corresponding inspection strategies are adopted for the data-type global objectives, the operational global objectives, and the strategic global objectives to ensure that each of the global objectives contains preset components. The components of the data-driven global objective include: quantitative indicators, benchmark values, and time nodes; the components of the operational global objective include: operational scope, execution boundaries, and acceptance criteria; and the components of the strategic global objective include: strategic direction, phased objectives, and evaluation dimensions.

5. An intelligent agent collaborative decision-making system based on hierarchical decoupling and adaptive switching, characterized in that, include: The macro-planning layer receives the global objective, decomposes the global objective into a set of executable subtasks, and sends subtask execution instructions to the micro-execution layer based on the set of executable subtasks; wherein, the subtask execution instructions include: task ID, objective, constraints, and expected output format; The micro-execution layer includes multiple micro-execution units, which receive the subtask execution instructions, convert the subtask execution instructions into execution events, and allocate the execution events to the corresponding micro-execution units for execution. The micro-execution units include: an agent based on API calls, a function or API interface based on LLM encapsulation, and an algorithm model that executes a preset function.

6. The agent-based collaborative decision-making system based on hierarchical decoupling and adaptive switching according to claim 5, characterized in that, Also includes: An environmental situation awareness module is used to monitor key environmental indicators and switch between a forward-looking planning-execution paradigm and a situational response paradigm based on the key environmental indicators; wherein, the key environmental indicators include at least one of the following: environmental entropy, task complexity, contingency indicators, and large model agent inference confidence.

7. The agent cooperative decision-making method based on hierarchical decoupling and adaptive switching according to claim 6, characterized in that, The environmental situation awareness module is also specifically used to perform weighted aggregation based on the key environmental indicators to calculate the environmental status assessment value. The weighted aggregation formula is expressed as follows: ； in, For environmental entropy, For task complexity, For emergency indicators, For large model agents, inference confidence, These are the corresponding weighting coefficients; Based on environmental state assessment values, the intelligent agent collaborative decision-making system switches between a forward-looking planning-execution paradigm and a situational response paradigm by switching decision functions; the switching decision functions are as follows: in, Environmental status assessment value, and It is a preset switching threshold, and .

Images