Multi-agent cooperative control method and system

Abstract

The application discloses a multi-agent cooperative control method and system, and belongs to the technical field of equipment control. The method comprises the following steps: receiving a natural language task instruction, analyzing a task semantic and generating a task capability vector; based on the task capability vector, loading an ability enhancement sub-parameter on demand from a shared large language model parameter pool, dynamically instantiating a lightweight sub-agent, and activating only a parameter subset corresponding to the capability dimension of each sub-agent; continuously monitoring a task complexity, an ability gap and an agent load state through a life cycle manager, and dynamically triggering a splitting operation or an apoptosis operation of the agent; predicting a cognitive state of a target agent through a sending agent, and completing message coding and sending according to a prediction result; after the communication is completed, collecting a feedback signal, updating the cognitive state of the target agent, and online or offline optimizing a prediction model of the sending agent.

Description

Technical Field

[0001] This application belongs to the field of equipment control technology, and specifically relates to a multi-agent cooperative control method and system. Background Technology

[0002] Currently, existing multi-agent systems based on large language models typically pre-define a fixed number of agents with roles during the design phase. Each agent is bound to a large language model instance or its variant, and they collaborate through explicit message passing mechanisms, relying on pre-designed role divisions and interaction processes to complete complex tasks.

[0003] This technical approach has the following shortcomings: First, the number and capability structure of intelligent agents are fixed and cannot be dynamically adjusted according to the complexity of the task, which can easily lead to resource redundancy or capability shortage. Second, communication between intelligent agents does not take into account the differences in the cognitive state of the receiver, which can easily lead to information redundancy, semantic ambiguity, or even error propagation. Third, lifecycle management is crude, and intelligent agents continue to occupy resources after the task is completed, which is not conducive to resource scheduling in large-scale scenarios. Fourth, the hierarchical or centralized management scheme still relies on preset control logic and lacks the ability to adapt to changes in task requirements and environment.

[0004] In summary, existing systems still have room for improvement in areas such as task-driven dynamic construction of agents, fine-grained configuration of capabilities, cognitive adaptation of collaborative communication, and adaptive management of lifecycles. There is an urgent need for a new multi-agent collaborative control method and system to improve resource utilization efficiency, collaborative intelligence level, and overall operational performance. Summary of the Invention

[0005] To address the aforementioned problems, this application provides a multi-agent cooperative control method and system, which can improve the resource utilization efficiency, cooperative stability, and overall operational performance of multi-agent systems in complex dynamic task environments. The method includes: S1. Receive natural language task instructions, parse task semantics and generate task capability vector; based on task capability vector, load capability enhancement sub-parameters on demand from the shared large language model parameter pool, dynamically instantiate one or more lightweight sub-agents, and each sub-agent only activates the parameter subset corresponding to its capability dimension. S2. During task execution, the lifecycle manager continuously monitors task complexity, capability gaps and agent load status, dynamically triggers agent splitting operations to generate new sub-agents, or triggers agent apoptosis operations to reclaim idle or low-contribution agents. S3. Before the sending agent sends a message to the target agent, the cognitive state of the target agent is predicted by the mental simulation module built into the sending agent, and the expression form and information granularity of the message are adaptively selected according to the prediction result to complete the message encoding and sending. S4. After communication is completed, collect feedback signals, update the cognitive state of the target agent, and optimize the prediction model of the mental simulation module online or offline.

[0006] In this embodiment of the application, receiving natural language task instructions, parsing task semantics, and generating task capability vectors includes: Semantic parsing of natural language task instructions is performed to extract task objectives, subtasks, capability elements, and dependencies. The parsing results are mapped to a predefined set of capability dimensions to generate a weight vector; The weight vector is obtained through one or more of the following methods: large language model output, rule mapping, historical statistics, or online feedback, and then normalized.

[0007] In this embodiment, one or more lightweight sub-agents are dynamically instantiated, and each sub-agent activates only a subset of parameters corresponding to its capability dimension, including: Threshold filtering is performed on the task capability vectors to determine the set of capability dimensions that need to be activated; Retrieve the capability enhancement sub-parameter module corresponding to the capability dimension from the shared parameter pool; Load and bind the selected modules to the base large language model to generate lightweight sub-agents; The number of sub-agents is determined based on capability weights or task parallelism; each sub-agent is assigned an independent context cache and task state.

[0008] In the embodiments of this application, the conditions for dynamically triggering the splitting operation of an agent to generate a new sub-agent, or triggering the apoptosis operation of an agent to reclaim idle or low-contribution agents, include at least the following: the task complexity index exceeds a preset threshold, the capability coverage is less than 0.85, and the load rate of any agent is greater than 0.8 or its response latency exceeds a preset latency threshold.

[0009] In this embodiment of the application, the agent splitting operation specifically includes: Computational capability gap vector; Load capability modules that can fill the capability gaps from the shared parameter pool; The current execution context is cloned from the parent agent to generate a child agent instance; The sub-tasks corresponding to the capability gaps are delegated to the newly generated sub-agents, and the state matrix of the agent cluster is updated.

[0010] In this application embodiment, the conditions for triggering the agent apoptosis operation include at least the following: the subtask is completed and there are no unfinished downstream dependent tasks, the agent's contribution is lower than the lower threshold, the agent's continuous idle time exceeds the idle threshold, and the overall system memory occupancy rate exceeds 85%, when the agent is forcibly recycled in order of contribution from low to high.

[0011] In this embodiment of the application, the contribution is calculated using a comprehensive index:

[0012] in, This represents the incremental progress contribution of the agent. This indicates its resource utilization rate. This represents a very small constant that prevents division by zero. This represents the lower limit threshold for contribution.

[0013] In this embodiment of the application, the central intelligence simulation module predicts the cognitive state of the target intelligent agent in the following ways: Construct the cognitive state vector of the target intelligent agent. The cognitive state vector includes: ability level vector, context mastery level, historical interaction feature vector, load and attention state. The cognitive state vector is maintained and updated based on real-time observation data and historical communication records.

[0014] In this embodiment of the application, the specific process for adaptively selecting the message expression format and information granularity includes: Construct input features, including the target agent's state, the content features of the message to be sent, and the current task context; For each set of candidate representations, predict the probability of successful comprehension and the communication cost. The optimal expression form is selected through a comprehensive evaluation function; The message is re-encoded and sent according to the selected expression format and granularity level.

[0015] In this embodiment of the application, the formula for predicting the probability of success is:

[0016] in, This represents the embedded feature vector expressed in form k. This represents the Sigmoid function, and the parameters of the MLP are continuously fine-tuned based on historical communication feedback. The formula for predicting the communication cost is:

[0017] in, This indicates the token length of the message after encoding using the k-expression. Indicates the expected number of clarification rounds; Indicates the expected response delay time. , , This represents the weighting coefficient for different cost items.

[0018] In this embodiment of the application, the specific methods of feedback learning after communication include: The cognitive state vector of the target agent is updated using an exponential moving average. The predicted comprehension probability and the actual misunderstanding result are used to form training samples, and the cross-entropy loss function is used to update the prediction parameters of the mental simulator using gradient.

[0019] This application also provides a multi-agent cooperative control system for task capability adaptive generation, the system comprising: The parameter scheduling module is used to receive natural language task instructions, parse task semantics and generate task capability vectors; based on the task capability vectors, it loads capability enhancement sub-parameters on demand from the shared large language model parameter pool, dynamically instantiates one or more lightweight sub-agents, and each sub-agent only activates the parameter subset corresponding to its capability dimension. The cycle management module is used to continuously monitor task complexity, capability gaps and agent load status through the lifecycle manager during task execution, dynamically trigger agent splitting operations to generate new sub-agents, or trigger agent apoptosis operations to reclaim idle or low-contribution agents. The intelligent communication module is used to predict the cognitive state of the target intelligent agent by means of the mental simulation module built into the sending intelligent agent before the sending intelligent agent sends a message to the target intelligent agent, and adaptively selects the expression form and information granularity of the message according to the prediction result to complete the message encoding and sending. The state update module is used to collect feedback signals after communication is completed, update the cognitive state of the target agent, and optimize the prediction model of the mind simulation module online or offline.

[0020] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods provided in the above embodiments.

[0021] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method provided in the above embodiments.

[0022] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

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

[0024] Figure 1 This is a flowchart illustrating a multi-agent cooperative control method provided in an embodiment of this application.

[0025] Figure 2 This document provides a flowchart for instantiating a dynamic intelligent agent based on task capability vectors, as an embodiment of this application.

[0026] Figure 3 This application provides a flowchart for an adaptive intelligent agent cluster dynamic management process based on splitting and extinction conditions.

[0027] Figure 4 An implicit communication adaptive flowchart based on mental simulation is provided for an embodiment of this application.

[0028] Figure 5 This is a general flowchart of a multi-agent cooperative control method provided in an embodiment of this application. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] Multi-Agent Systems (MAS) are computerized systems composed of multiple autonomous agents that interact collaboratively. Their core objective is to accomplish complex tasks that are difficult for a single agent to perform independently through information exchange and task collaboration among agents.

[0031] In recent years, with the rapid development of large language model technology, multi-agent collaborative systems built on large language models have gradually become an important research direction in the field of artificial intelligence. These systems utilize the powerful natural language understanding, logical reasoning, and context learning capabilities of large language models to provide new technical approaches for autonomous cognition, task decision-making, and collaborative operations of intelligent agents.

[0032] For example, in existing technologies, the paper "[The Rise and Potential of Large Language ModelBased Agents]" proposes an agent framework based on a large language model. This framework divides the agent into three functional modules: brain, perception, and action. The large language model serves as the core decision-making unit, and the framework enhances the agent's task understanding and decision-making capabilities through input modality fusion, prompting engineering, retrieval enhancement, and tool invocation. The CAMEL framework ([CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society]) employs a role-playing mechanism, pre-setting the roles of the task assigner, AI user, and AI assistant through initial prompts, thus achieving multi-turn command-response collaboration. The CGMI framework ([CGMI: Configurable General Multi-Agent Interaction Framework]) constructs a tree-like role model based on the Big Five personality traits scale. Through coarse-grained node hierarchical allocation and depth-first traversal, it addresses the issues of agent role forgetting and insufficient stability. The AutoGen framework [AutoGen: EnablingNext-Gen LLM Applications via Multi-Agent Conversation] allows developers to customize agent behavior and communication patterns, enabling multiple large language model agents to collaborate through dialogue to decompose and execute complex tasks.

[0033] In terms of communication and collaboration topologies, existing methods mainly include centralized topologies (Mindstorms in Natural Language-Based Societies of Mind), decentralized topologies, distributed topologies (MetaGPT: MetaProgramming for A Multi-Agent Collaborative Framework), and hierarchical topologies (A DynamicLLM-Powered Agent Network for Task-Oriented Agent Collaboration). Centralized structures enable unified scheduling and planning but are susceptible to single points of failure. Decentralized and distributed structures allocate tasks through local communication between agents, exhibiting high robustness and fault tolerance. Hierarchical topologies combine the advantages of centralized and distributed approaches, enabling cross-level task coordination and decision optimization.

[0034] Furthermore, existing research has explored advanced mechanisms in multi-agent collaboration. MetaGPT utilizes metaprogramming techniques to implement structured workflows among agents, effectively decomposing complex tasks and reducing the risk of collaboration conflicts; Consensus-LLM (Multi-Agent Consensus Seeking via Large Language Models) introduces a consensus mechanism, enabling agents of large language models to negotiate and unify collaborative goals in dynamic environments; Agent Group Chat (AgentGroupChat-V2: Divide-and-Conquer Is What LLM-Based Multi-Agent System Need) and AutoAgents (AutoAgent: A Fully-Automated and Zero-Code Framework for LLM Agents) demonstrate dialogue-based task decomposition and dynamic agent generation methods, expanding the application capabilities of multi-agent systems in heterogeneous tasks and real-time collaborative scenarios.

[0035] However, while existing technologies have achieved certain results in task planning, role allocation, communication protocols, and collaboration mechanisms, they still have the following technical shortcomings: First, the mapping relationship between agent roles and task capabilities is mostly preset or statically configured, lacking a dynamic mechanism for on-demand generation, making it difficult to adaptively adjust the capability structure according to task complexity; Second, collaboration between agents relies on explicit message passing and fixed communication formats, resulting in information redundancy and efficiency loss; Third, the system lacks a closed-loop adaptive mechanism, making it difficult to achieve autonomous adjustment and evolution when task requirements change or abnormal situations occur.

[0036] To address the aforementioned technical deficiencies, this application aims to solve the following key technical problems: the problem of adaptive construction of task-oriented agent capability structure, the problem of high computational and storage costs in the instantiation and deployment of multi-agent system models, the problem of dynamic evolution of the number and capabilities of agents during execution and the problem of zombie agent resource occupation, and the problem of low collaboration efficiency caused by invalid communication and semantic ambiguity in multi-agent collaboration.

[0037] To address the aforementioned technical problems, this application proposes a multi-agent cooperative control method, which can improve the resource utilization efficiency, cooperative stability, and overall operational performance of multi-agent systems in complex dynamic task environments.

[0038] Figure 1 This is a flowchart illustrating a multi-agent cooperative control method provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes: S1. Receive natural language task instructions, parse task semantics and generate task capability vectors; based on task capability vectors, load capability enhancement sub-parameters on demand from the shared large language model parameter pool, dynamically instantiate one or more lightweight sub-agents, and each sub-agent only activates the parameter subset corresponding to its capability dimension.

[0039] S2. During task execution, the lifecycle manager continuously monitors task complexity, capability gaps, and agent load status, dynamically triggering agent splitting operations to generate new sub-agents, or triggering agent apoptosis operations to reclaim idle or low-contribution agents.

[0040] S3. Before the sending agent sends a message to the target agent, the cognitive state of the target agent is predicted by the mental simulation module built into the sending agent, and the expression form and information granularity of the message are adaptively selected according to the prediction result to complete the message encoding and sending.

[0041] S4. After communication is completed, collect feedback signals, update the cognitive state of the target agent, and optimize the prediction model of the mental simulation module online or offline.

[0042] This application receives natural language task instructions and extracts the distribution of the required capabilities for the task through semantic parsing and complexity evaluation, including capability types, strengths and evolution trends, without pre-setting any intelligent agent roles or numbers.

[0043] Based on the distribution of capability requirements, this application adaptively constructs the capability structure of intelligent agents. Through the shared base model and parameter pool reuse mechanism, intelligent agents with specific capability configurations are instantiated on demand in a lightweight manner, so as to realize the on-demand activation of capabilities rather than the repeated occupation of complete model instances.

[0044] Furthermore, this application organizes collaboration around capabilities, dynamically routing subtasks to agents with matching capabilities. During communication, the sending agent predicts the receiving agent's cognitive state (including capability boundaries, contextual understanding, and task progress) in advance, and adaptively adjusts the message expression and information granularity accordingly, forming implicit communication to reduce ineffective interactions.

[0045] In summary, this application can improve the resource utilization efficiency, collaborative stability, and overall operational performance of multi-agent systems in complex and dynamic task environments.

[0046] In this embodiment of the application, receiving a natural language task instruction, parsing the task semantics, and generating a task capability vector includes: performing semantic parsing on the natural language task instruction to extract the task objective, subtasks, capability elements, and dependencies; mapping the parsing results to a predefined set of capability dimensions to generate a weight vector.

[0047] The weight vector is obtained through one or more of the following methods: large language model output, rule mapping, historical statistics, or online feedback, and then normalized.

[0048] In this embodiment, one or more lightweight sub-agents are dynamically instantiated, and each sub-agent activates only a subset of parameters corresponding to its capability dimension. This includes: threshold filtering of the task capability vector to determine the set of capability dimensions to be activated; retrieving capability enhancement sub-parameter modules corresponding to the capability dimensions from a shared parameter pool; loading and binding the selected modules to the basic large language model to generate lightweight sub-agents; determining the number of sub-agents based on capability weights or task parallelism; and allocating independent context caches and task states to each sub-agent.

[0049] In the embodiments of this application, the conditions for dynamically triggering the splitting operation of an agent to generate a new sub-agent, or triggering the apoptosis operation of an agent to reclaim idle or low-contribution agents, include at least the following: the task complexity index exceeds a preset threshold, the capability coverage is less than 0.85, and the load rate of any agent is greater than 0.8 or its response latency exceeds a preset latency threshold.

[0050] In this embodiment of the application, the agent splitting operation specifically includes: a computational capability gap vector; Load capability modules that can fill the capability gap from the shared parameter pool; clone the current execution context from the parent agent to generate a child agent instance; delegate the subtasks corresponding to the capability gap to the newly generated child agent and update the agent cluster state matrix.

[0051] In this application embodiment, the conditions for triggering the agent apoptosis operation include at least the following: the subtask is completed and there are no unfinished downstream dependent tasks, the agent's contribution is lower than the lower threshold, the agent's continuous idle time exceeds the idle threshold, and the overall system memory occupancy rate exceeds 85%, when the agent is forcibly recycled in order of contribution from low to high.

[0052] In this embodiment of the application, the contribution is calculated using a comprehensive index:

[0053] in, This represents the incremental progress contribution of the agent. This indicates its resource utilization rate. This represents a very small constant that prevents division by zero. This represents the lower limit threshold for contribution.

[0054] In this embodiment of the application, the central intelligence simulation module predicts the cognitive state of the target intelligent agent in the following ways: S5. Construct the cognitive state vector of the target intelligent agent. The cognitive state vector includes: ability level vector, context mastery level, historical interaction feature vector, load and attention state.

[0055] S6. The cognitive state vector is maintained and updated based on real-time observation data and historical communication records.

[0056] In this embodiment of the application, the specific process for adaptively selecting the message expression format and information granularity includes: S7. Construct input features, including the target agent state, the content features of the message to be sent, and the current task context.

[0057] S8. For the set of candidate representations, predict the probability of successful understanding and the communication cost respectively.

[0058] S9. Select the optimal expression form through a comprehensive evaluation function.

[0059] S10. Encode the message again according to the selected expression form and granularity level and send it.

[0060] In this embodiment of the application, the formula for predicting the probability of success is:

[0061] in, This represents the embedding feature vector of the expression form k. This represents the Sigmoid function, and the parameters of the MLP are continuously fine-tuned based on historical communication feedback.

[0062] The formula for predicting communication costs is:

[0063] in, This indicates the token length of the message after encoding using the k-expression. Indicates the expected number of clarification rounds; Indicates the expected response delay time. , , This represents the weighting coefficient for different cost items.

[0064] In this embodiment of the application, the specific methods of feedback learning after communication include: S12. Update the cognitive state vector of the target agent using an exponential moving average.

[0065] S13. Combine the predicted comprehension probability with the actual misunderstanding result to form training samples, and use the cross-entropy loss function to update the prediction parameters of the mental simulator using gradient.

[0066] This application also provides a multi-agent cooperative control system, which includes: a parameter scheduling module, a period management module, an intelligent communication module, and a state update module.

[0067] The parameter scheduling module is used to receive natural language task instructions, parse task semantics and generate task capability vectors; based on the task capability vectors, it loads capability enhancement sub-parameters on demand from the shared large language model parameter pool, dynamically instantiates one or more lightweight sub-agents, and each sub-agent only activates the parameter subset corresponding to its capability dimension.

[0068] Furthermore, the parameter scheduling module includes a meta-controller, a shared large language model parameter pool, and a dynamic instantiation engine. The meta-controller receives natural language task instructions, parses the task semantics, and generates a task capability vector. The shared large language model parameter pool and the dynamic instantiation engine, based on the task capability vector, load capability enhancement sub-parameters on demand from the shared large language model parameter pool, dynamically instantiating one or more lightweight sub-agents. Each sub-agent only activates a subset of parameters corresponding to its capability dimension.

[0069] The lifecycle management module is used to continuously monitor task complexity, capability gaps, and agent load status during task execution through the lifecycle manager. It can dynamically trigger agent splitting operations to generate new sub-agents, or trigger agent apoptosis operations to reclaim idle or low-contribution agents.

[0070] Furthermore, the lifecycle management module includes: a dynamic lifecycle management module.

[0071] The lifecycle dynamic management module is used to continuously monitor task complexity, capability gaps and agent load status through the lifecycle manager during task execution, and dynamically trigger agent splitting operations to generate new sub-agents, or trigger agent apoptosis operations to reclaim idle or low-contribution agents.

[0072] Furthermore, the intelligent communication module includes a pre-communication cognitive adaptation module.

[0073] The pre-communication cognitive adaptation module is used to predict the cognitive state of the target intelligent agent before the sending intelligent agent sends a message to the target intelligent agent by means of the mental simulation module built into the sending intelligent agent, and adaptively selects the expression form and information granularity of the message according to the prediction result, and completes message encoding and sending.

[0074] The state update module is used to collect feedback signals after communication is completed, update the cognitive state of the target agent, and optimize the prediction model of the mind simulation module online or offline.

[0075] Taking the multi-agent cooperative control system provided in this application as an example, this application will be explained in detail. Figure 2 This application provides a flowchart for instantiating a dynamic intelligent agent based on task capability vectors, as an embodiment of the present application. Figure 3 This application provides a flowchart of an adaptive intelligent agent cluster dynamic management process based on splitting and extinction conditions, as shown in the embodiments of this application. Figure 4 This application provides an embodiment of an implicit communication adaptive flowchart based on mental simulation. Figure 5 A general flowchart of a multi-agent cooperative control method provided in this application embodiment is shown below. Figures 2 to 5 As shown, the multi-agent cooperative control system provided in this application sets a meta-controller at the task entry point, which is responsible for mapping any natural language task instruction into an n-dimensional task capability vector. Each dimension Quantify the required weight of corresponding abilities (such as coding, mathematics, finance, law, etc.) in the current task.

[0076] It should be noted that this system does not pre-define any fixed agent roles, but maintains a shared global parameter pool, which stores the frozen basic large language model and a large number of pre-trained lightweight modules (e.g., in the form of LoRA / Adapter).

[0077] Based on the above embodiments, the dynamic instantiation engine loads and combines corresponding capability modules from the parameter pool as needed according to the distribution of capability vector V, and "assembles" several lightweight sub-agents. Each sub-agent only activates a subset of parameters that are highly related to its own capability dimension, thereby achieving the generation of an agent structure with zero redundancy, zero cold start, and precise controllable capability ratio.

[0078] This system maintains only one basic agent instance at startup, which greatly reduces idle resource consumption. During task execution, the lifecycle manager continuously monitors task complexity, the number of subtask branches, and the load and contribution of each agent. When a capability gap or task differentiation is detected, agent splitting is automatically triggered: new capability modules are cloned from the parent agent and injected to form a sub-agent focused on a specific subtask. The sub-agent inherits the parent agent's context and some parameters, achieving seamless knowledge and state transfer.

[0079] When a subtask is completed, its contribution falls below a threshold, or it remains idle for an extended period, the agent is triggered to die, automatically releasing the video memory, context cache, and computing resources it occupies. This completely eliminates the "zombie agent" phenomenon, allowing the overall system workload to scale in a breathing manner.

[0080] Each agent embeds a lightweight mental simulator to infer the ability level of other agents, their current contextual understanding, and cognitive load in real time.

[0081] Before sending a message, the sender first uses a mental simulator to predict the granularity and expression of the information that the receiver can most easily understand (e.g., sending function implementations directly to code experts, sending annotated pseudocode to beginners, and sending minimalist keywords to high-load agents). This enables adaptive compression and style transfer of information expression, significantly reducing semantic ambiguity and redundant communication rounds. The entire communication process no longer relies on cumbersome explicit protocol negotiation, but achieves near-human intuitive natural collaboration through implicit modeling of the "other's cognitive state," significantly improving the collaborative efficiency between agents with heterogeneous capabilities.

[0082] Through this overall design, the system can start from an almost zero state when any task arrives, generate an intelligent agent cluster with the optimal size and capability ratio in seconds, and automatically return to zero after the task ends. This achieves the ultimate goal of maximizing computing resource utilization, maximizing collaboration efficiency, and maximizing system elasticity, completely surpassing the existing technical framework of fixed roles, fixed instances, and fixed communication paradigms.

[0083] This application proposes a multi-agent construction method that does not rely on a predefined set of roles. By structurally modeling the capabilities required for the task, the natural language task is mapped into a computable task capability vector. This capability vector is used as the sole driving signal to select and load capability enhancement sub-parameter modules from a shared model parameter pool as needed, and to dynamically instantiate one or more lightweight agent sub-instances, thereby forming an agent collaboration structure that is highly matched to the task requirements.

[0084] After receiving the natural language task instructions input by the user, the system first uses the meta-controller to invoke a large language model to perform semantic parsing on the task text. The parsing process includes, but is not limited to: task target identification, extraction of subtasks and implicit capability elements, and identification of task complexity and key dependency elements (e.g., whether retrieval is needed, whether tools need to be invoked, whether structured output is needed, etc.). The parsing results are not directly used to generate fixed-role intelligent agents, but are further mapped to a unified capability space to form a computable basis for subsequent instantiation.

[0085] This application predefines a set of scalable capability dimensions. Each capability dimension corresponds to a type of reusable cognitive or functional ability. Examples include: reasoning and planning ability, information retrieval and integration ability, code generation or tool usage ability, data analysis ability, structured expression ability, and domain knowledge ability.

[0086] It should be noted that the capability dimension set is an abstract coordinate axis of the model's capabilities, used to represent the combination relationship of capabilities, and is not the same as "roles". Furthermore, capability dimensions can be expanded or merged according to application scenarios.

[0087] Based on the semantic parsing results, the meta-controller generates a task capability vector:

[0088] Where n represents the number of capability dimensions; For corresponding capabilities The weighting coefficient in the current task reflects the importance, frequency of use, or computational intensity requirements of this capability; and Weight Capability weights can be obtained through one or more of the following methods: outputting capability weights by a large language model based on prompt templates; calculating weights based on rule mapping (such as a mapping table from keywords / intents to capability dimensions); correcting them based on historical task statistics or online feedback; or using normalization processing (such as Softmax or normalization ratio) to make the weights comparable.

[0089] For example, when a user enters the task instruction "Please analyze this financial statement and generate visualization code", the meta controller performs the following operations: Step 1: Semantic Feature Extraction: Use a lightweight model (or an LLM with specific prompts) to identify the intent of the instruction.

[0090] Step 2, Vector Generation: The system defines an n-dimensional space, for example: Code generation Mathematical calculations Financial knowledge :writing... Step 3: After softmax normalization, generate the capability weight vector. ; In summary, this task requires an agent with high financial knowledge (0.45) and coding ability (0.4).

[0091] This system maintains a shared pool of large language model parameters, which includes at least: base model parameters, preferably pre-loaded into the computing device's video memory to reduce instantiation latency and kept read-only during task execution; and a set of capability enhancement sub-parameter modules: multiple sets of independently loadable / unloadable capability enhancement parameter modules, each set of modules corresponding to one or more capability dimensions, used to enhance specific capabilities without copying the base model.

[0092] The aforementioned sub-modules can be implemented using efficient parameter fine-tuning methods, including but not limited to LoRA (Low-Rank Adaptation), Adapter Tuning, Prefix Tuning, Prompt Tuning, or expert modules. Taking LoRA as an example, each capability module can be treated as a set of low-rank incremental parameters, stored and indexed according to the capability dimension.

[0093] Specifically: Base Model: A shared, frozen, large base model; LoRa module library: Stores LoRA weight blocks corresponding to different capability dimensions, each weight can be tens of MB in size; Mapping relationships: For example... (Code generation capability dimension) (Financial knowledge and skills dimension).

[0094] When the task capability vector One of the ability dimensions And the weight exceeds the preset threshold At that time, the system triggers the instantiation process of the corresponding capability module. This threshold... It can be a fixed value (e.g., 0.1) or adaptively adjusted according to system resource usage and task priority.

[0095] In a specific example, the dynamic agent instantiation process includes the following steps S1 to S6: Step S1: For Perform threshold filtering to determine the set of capabilities that need to be activated. ; Step S2: Retrieve parameters from the parameter pool that match... The corresponding set of capability enhancement sub-parameter modules; when there are multiple candidate modules for a capability dimension, the optimal module can be selected according to module matching degree, historical effect or resource budget; Step S3: Load and bind the selected capability modules to the base model to generate a lightweight agent instance; Step S4: Different capability dimensions can correspond to one or more sub-agents, with the number of instantiations and corresponding weights. It shows a positive correlation, or is related to the degree of parallelism of the task and the number of subtasks; Step S5: Allocate an independent local context cache and task state description for each sub-agent to ensure that the inference sessions of each agent are isolated from each other.

[0096] Based on the above embodiments, each dynamically instantiated sub-agent includes at least: shared basic model parameters (read-only, not copied), independently activated capability enhancement sub-parameter modules (pluggable), independent context cache (e.g., dialogue history summary, tool call state), and independent task state (e.g., sub-task identifier, input / output constraints).

[0097] Through the above sharing mechanism, each agent does not copy the complete model, but only shares the basic weights and loads a small number of differentiated capability modules, thereby significantly reducing memory usage and instantiation overhead.

[0098] In this application embodiment, when the task span is large or the subtasks are naturally parallelizable, the system can split and instantiate multiple sub-agents according to the capability dimension, for example: Sub-agent A: Loads financial knowledge and ability modules (such as...) ), focusing on financial statement indicator analysis and data processing; Sub-agent B: loading code generation capability module (such as ), focusing on generating visualization code and chart configurations.

[0099] Through the above mechanism, the system can automatically form a set of agents that correspond one-to-one with the task's capability structure at the beginning of the task: the number of agents dynamically changes with the task requirements; the capabilities of agents are precisely constrained by capability vectors; and there are no redundant agents without task load. This collaborative structure does not rely on manually configured role templates, nor does it require fixing the agent size during the system initialization phase.

[0100] This application also proposes a dynamic role creation and lifecycle management mechanism to adaptively regulate the generation, evolution, and recycling of agent instances during multi-agent collaborative task execution, based on changes in task complexity, evolving capability requirements, and system resource status. Through this mechanism, the system can ensure execution efficiency while avoiding long-term redundant residency of agent instances, thereby achieving efficient utilization of computing resources.

[0101] This mechanism uses a single basic agent instance as the initial state of the system. Through two dual operations, agent splitting mechanism and agent apoptosis mechanism, the size of the agent cluster dynamically expands and shrinks as the task progresses, forming an agent lifecycle management system with adaptive characteristics.

[0102] During the initial system startup phase, only one base agent is instantiated, serving as the starting instance for subsequent agent evolution and scheduling. This base agent possesses the following design features: 1. Model configuration method: The basic agent loads the basic large language model parameters (BaseLLM) from the shared parameter pool, and only has general language understanding and reasoning capabilities. It does not preload any domain-specific or task-related LoRA-adapter or other capability modules.

[0103] This design ensures that the system's resource consumption is minimized when it is idle, while retaining the flexibility to load professional capability modules on demand.

[0104] 2. Capability Threshold Constraint: The basic intelligent agent only possesses basic capabilities such as general logical reasoning, simple task planning, and context management. In the capability vector representation, the upper limit of the weight of each capability dimension can be set as follows: This indicates that it lacks the ability to handle highly specialized tasks, but is sufficient to perform meta-control functions such as task analysis, capability requirement assessment, and split decision-making.

[0105] 3. Resource Usage Control Strategy: The memory usage of the basic intelligent agent is controlled between 20% and 30% of the system baseline, and the computational load is kept below 10%. This lightweight configuration allows the system to remain in standby mode with extremely low resource costs, reserving ample resource space for subsequent task-driven dynamic expansion.

[0106] The core responsibilities of a basic intelligent agent include: receiving and parsing natural language task instructions input by the user; calling the meta-controller to generate task capability vectors; assessing the degree of matching between its own capabilities and task requirements; and triggering a splitting mechanism to generate child intelligent agents when its capabilities are insufficient.

[0107] In summary, the agent splitting mechanism is the core means by which this application achieves dynamic expansion of the agent cluster. When the system detects that the current agent cluster's capacity coverage or processing load can no longer meet the task requirements, it automatically triggers the agent splitting mechanism to generate new sub-agent instances, thereby undertaking specific sub-tasks or supplementing capacity gaps.

[0108] During task execution, the system will automatically initiate the agent splitting process if any of the following triggering conditions are met: 1. Task complexity exceeds preset threshold: The system quantitatively evaluates the complexity of the current task, taking into account factors such as the number of subtask branches, the depth of logical nesting, and the estimated remaining execution time. Its complexity index can be expressed as:

[0109] in, This indicates the number of subtask branches after the current task is decomposed; It indicates the depth of task logic nesting, reflecting the complexity of reasoning and integration. Indicates the estimated remaining execution time of the task. This represents the weighting coefficient of each indicator, which can be configured according to the application scenario. Set a threshold (e.g., 0.7).

[0110] when Exceeding the preset threshold When the system determines that the complexity of the current task has exceeded the processing capacity of a single agent, it needs to split and generate more agents for parallel processing.

[0111] II. Insufficient capability vector coverage: The system continuously monitors the capability coverage of the current agent cluster. Capability Coverage The calculation method is as follows: compare the task requirement capability vector. With existing intelligent agent cluster capability matrix The degree of matching. A split is triggered when the capability coverage meets the following conditions:

[0112] This condition indicates that the existing intelligent agent cluster's capability combination cannot cover more than 85% of the capabilities required for the task, indicating a significant capability gap, and necessitates the development of new intelligent agents with the missing capabilities.

[0113] III. System load saturation state: The system monitors the resource load status of each agent in real time. When the load rate of any agent... If the response delay exceeds a preset latency threshold, the system determines that the agent is under load saturation and needs to share its workload by splitting it.

[0114] After any of the above triggering conditions are met, the system performs the agent splitting operation according to the following steps: (1) Identification and quantification of capability gaps: The system monitoring module first calculates the task requirement vector. With existing agent capability matrix The coverage differences between them determine the capability gap vector that needs to be filled:

[0115] in, This represents the capability vector of the i-th existing agent. This determines the weight of its contribution to the current task. Through this calculation, the system can accurately identify which capability dimensions have gaps and the extent of those gaps.

[0116] (2) Sub-agent initialization and capability injection: Based on the capability gaps identified in step one The system performs the following sub-agent initialization operations: First, retrieving and loading parameters from the shared parameter pool can supplement the data. The system prioritizes loading the Lora-adapter module that corresponds to the dimension with the largest capacity gap.

[0117] Secondly, the current execution context (Context Snapshot) is cloned from the parent agent, including dialogue history, intermediate inference results, and task state information, to ensure that the child agents can seamlessly connect to task execution.

[0118] Finally, a new sub-agent instance is generated, whose capability vector is set according to the following formula:

[0119] in, The capability vector representing the parent agent. This indicates the incremental capability provided by the newly loaded capability module.

[0120] (3) Task allocation and resource activation: The system delegates the sub-tasks corresponding to the capability gaps to the newly generated child agents and allocates the necessary computing resources to them. At the same time, it updates the state matrix of the agent cluster, recording the inheritance relationship between the child agents and the parent agents, so as to facilitate subsequent collaborative scheduling and lifecycle management.

[0121] Sub-agents generated through the splitting mechanism possess the following core characteristics: ① Focused Design: Sub-agents only load parameter modules highly relevant to their assigned sub-tasks, avoiding resource waste caused by loading irrelevant capabilities. Their memory usage is controlled between 40% and 60% of the baseline value for a single agent, achieving lightweight deployment.

[0122] ② Parameter inheritance mechanism: The child agent directly reuses the parent agent's call handle to the basic large language model, without having to repeatedly load the basic model parameters. It only incrementally loads its own capability extension modules, thereby achieving rapid instantiation.

[0123] ③ Context Inheritance Mechanism: Child agents inherit 80% to 90% of the parent agent's dialogue history and intermediate inference results, ensuring the continuity of task execution. To save token budget, the system uses digest compression technology, only transmitting refined key context summaries to child agents, rather than the complete historical record.

[0124] ④ Independent lifecycle: Each sub-agent has an independent contribution counter and death timer. Its lifecycle is determined independently by its own task completion status and resource contribution, and is not directly constrained by the state of the parent agent.

[0125] To prevent "zombie agents" from occupying system resources for extended periods, this application designs a robust agent apoptosis mechanism. This mechanism, through a comprehensive judgment of multiple triggering conditions, ensures that agents that have completed their mission or have been inefficient for a long time can be promptly reclaimed, releasing the computing and storage resources they occupy.

[0126] The system is configured with the following four types of apoptosis triggering conditions; the apoptosis process can be triggered if any one of these conditions is met: (1) Subtask completion determination: When a subtask that an agent is responsible for is marked as "completed" and there are no unfinished downstream dependent tasks for that subtask, the system determines that the agent has completed its mission and can trigger apoptosis.

[0127] (2) Contribution is below the threshold: The system continuously evaluates the contribution of each agent to the task progress. Contribution level. The basic calculation formula is:

[0128] in, This represents the incremental progress contribution of the agent. This indicates its resource utilization rate. This represents a very small constant that prevents division by zero. This represents the lower limit threshold for contribution (typically ranging from 0.1 to 0.2).

[0129] To more comprehensively evaluate the agent's contribution, the system also uses a comprehensive contribution index. :

[0130] The first metric measures the proportion of the agent's output tokens to the total input tokens of the system, reflecting its output efficiency; the second metric measures the improvement in task progress per unit of active time, reflecting its work efficiency; and the third metric measures the accuracy of its output, reflecting its work quality. , , Let be the weighting coefficient, satisfying .

[0131] (3) Long-term idle timeout: When an agent remains idle for a longer period of time than a preset idle threshold is reached... When the time (default value range is 30 to 60 seconds) is reached, the system determines that the agent is no longer needed and triggers its apoptosis.

[0132] (4) Resource recycling needs: When the overall system video memory usage When the utilization rate exceeds 85%, in order to ensure the normal operation of the system, the system will forcibly eliminate agents and reclaim resources in order of their contribution from low to high, until the memory utilization rate drops to a safe level.

[0133] Once the apoptosis condition is triggered, the system executes the agent's apoptosis operation according to the following procedure: (1) State snapshots are saved for knowledge transfer: Before an agent is destroyed, the system first extracts its valuable information, including intermediate results generated during task execution, compressed context summaries, and local knowledge or patterns learned during execution. This critical information is broadcast to the parent agent or other relevant collaborating agents to ensure that important knowledge is not lost with the agent's demise.

[0134] (2) Parameter unloading and resource release: The system calls the model engine to unload the LoRA-adapter capability module specific to this agent; clear its context cache, KV cache and tool session state; and release the GPU memory, CPU memory, process handle and related network connection resources it occupies.

[0135] (3) Cluster status update and gap detection: The system removes the apoptotic agent instances from the agent registry and synchronously updates the capability coverage matrix of the agent cluster. After the update is complete, the system automatically triggers a new round of capability gap detection. If the detection finds new capability gaps in the apoptotic cluster and there are still unfinished tasks, it may trigger a new round of agent splitting, forming a dynamic equilibrium closed loop of "apoptosis-detection-split".

[0136] To enable the system to adapt to the characteristics of different types of tasks, this invention designs a threshold adaptive adjustment mechanism. The system continuously records resource utilization and task completion results during historical task execution, and dynamically adjusts key threshold parameters, including the complexity threshold, using the exponential moving average (EMA) method. Idle timeout threshold wait.

[0137] For tasks marked as high priority, the system can temporarily lower the split trigger threshold, enabling the agent cluster to scale up more quickly and ensuring sufficient parallel processing capabilities; for regular tasks, the system adopts a relatively conservative threshold setting to balance task efficiency and resource consumption.

[0138] Based on the above embodiments, this application also proposes an implicit communication adaptive mechanism based on the Theory of Mind Simulation. This mechanism enables the sending agent to predict the cognitive state of the receiving agent during multi-agent collaboration, and accordingly adaptively adjusts the expression form and information granularity of the communication content. This mechanism is built into each agent as a lightweight module, without relying on predefined complex communication protocols or fixed message formats. Instead, it achieves intelligent selection of information expression methods and optimization of content details through dynamic modeling of the "other's cognitive state," thereby reducing communication ambiguity, minimizing invalid interaction rounds, and improving the collaboration efficiency between heterogeneous agents.

[0139] Each agent integrates a lightweight ToM Simulator module, which is primarily responsible for constructing the internal cognitive representations of other target agents in the system and predicting the probability of different message expressions being correctly understood by the target agents and their communication costs. To reduce additional overhead, the parameter size of this module is controlled to within 5% of the total number of parameters of the agent (preferably implemented as an MLP or a lightweight Transformer Encoder).

[0140] For any target agent j, its cognitive state Represented in four-dimensional vector form:

[0141] in, It is a capability vector, derived from the dynamic instantiation mechanism. , representing the relative strength of agent j in each capability dimension (such as code, mathematics, finance), with a value range of [0,1]. The number of dimensions is related to the task capability vector. Consistent; It is the context mastery state, used to quantify the degree to which agent j has mastered the current context (dialogue history, intermediate results), and takes the value [0,1]. The closer the value is to 1, the more fully the understanding of the context is. Historical interaction features are feature vectors obtained by statistically analyzing the agent's past communication records. These include statistical indicators such as average number of clarification rounds, historical misunderstanding rate, and response accuracy. Typically, these features have 3-5 dimensions and are used to characterize the agent's stability and comprehensibility in dialogue. It is the load / attention state, which reflects the current resource load and attention allocation level of the target agent. The value range is [0,1]. The closer the value is to 1, the higher the load or the execution is blocked.

[0142] The mental simulation module maintains a target agent state table to record the states of all active agents in the system. Value. For newly added agents to the task, its Initial values ​​can be obtained through capability vectors It is initialized with a preset default template and updated in real time during runtime.

[0143] The aforementioned mental simulation module includes a state acquisition unit, a state prediction unit, and a policy generation unit. The state acquisition unit collects real-time observation data related to the target agent from the system bus, lifecycle manager, and message logs, including capability configuration, current load metrics, and historical message statistics. This provides the basic data input for state estimation.

[0144] The state prediction unit is used to input the above observation data into the lightweight model (MLP / Transformer) and output the updated cognitive state vector. And predict the probability of understanding under different forms of expression.

[0145] The strategy generation unit evaluates the available expression forms and information granularities based on the output of the state prediction unit, and generates corresponding communication strategies to guide subsequent message encoding and transmission.

[0146] Specifically, when the sending agent i prepares to send message m to the target agent j, it will first perform a pre-communication cognitive simulation and policy selection through the mental simulation module, as follows (this process is designed to be lightweight, with computational latency controlled within 50ms): I. Input Feature Construction: Constructing mental simulation input features ; is the current state vector of the target intelligent agent; m represents the features of the content to be sent, including type (code snippet / control instruction / data description, etc.), complexity (number of tokens, depth of logical dependencies), and domain label; current task context information, including the current task stage (such as planning, execution, verification) and the urgency of the task (such as whether there are time constraints, etc.).

[0147] II. Understanding the Difficulty and Cost Prediction: For a predefined set of candidate expressions:

[0148] The state prediction unit calculates the probability that the target agent j will successfully understand the message for each expression form k. and the corresponding communication costs .

[0149] (1) Understanding success probability prediction:

[0150] in, This represents the embedding feature vector of the expression form k. The parameters of the MLP are continuously fine-tuned based on historical communication feedback, using the Sigmoid function.

[0151] (2) Communication cost forecast:

[0152] in, The length of the message token after being encoded using the expression form k; Is the expected number of clarification rounds (based on) (Estimation of indicators such as historical misunderstanding rate); It is the expected response delay time (based on) Load state estimation in (the process). , , These are the weighting coefficients for different cost items, which can be configured according to the application scenario.

[0153] III. Optimization and Selection of Expression Strategies: Based on the probability of understanding and the communication cost, the generation unit constructs a comprehensive evaluation function to optimize the selection of the expression form:

[0154] in, This represents the cost sensitivity parameter, with a default value of 0.3, which can be adaptively adjusted based on system operating conditions. The optimal expression form needs to be determined. In addition, you can also select the information granularity level. This is used to control the level of detail in the message content. For example, high granularity can include detailed comments and step-by-step instructions, while low granularity only retains key interface or parameter information.

[0155] IV. Adaptive Message Encoding and Transmission: The sending agent is based on the selected expression form. and particle size class The original message m is encoded a second time. After encoding, the message is sent to the target agent j via the system message bus without additional handshake or protocol negotiation. The following is a typical scenario example (for illustrative purposes only): (1) When the receiver is an intelligent agent with strong coding capabilities and low load, the "source code" format can be selected for low-granularity output, for example: def func(x): return x 2.

[0156] (2) When the receiver is a beginner and the current load is high, the “pseudocode” form can be selected for medium-granularity output, for example: function square(input x): returns the square of x.

[0157] (3) When the receiver lacks current context information, the "natural language interpretation" form can be selected for high-granular output, for example: Please calculate x², where x=5, the result should be 25. The function above returns the square of the input value.

[0158] To enable the mind simulation module to have self-learning and continuous optimization capabilities, this application also designs a post-communication feedback collection and status update mechanism.

[0159] I. Feedback Signal Acquisition: Within a preset time window (e.g., 10 seconds) after each communication, the system collects the following feedback metrics: number of clarification requests (the number of times the receiving agent sends clarification or restatement requests for this message), response latency (the time interval from when the message is sent to the receiver's first valid response), and task gain (the progress improvement of related subtasks after this communication). ), Decision accuracy (the correctness of the decision made or the result generated by the recipient based on the message).

[0160] II. Cognitive Status Update: Based on the above feedback signals, the cognitive state vector of the target agent... Update as needed. An exponential moving average (EMA) is preferred to balance historical stability with sensitivity to new feedback.

[0161] in, This represents the learning rate, typically ranging from 0.1 to 0.3. This represents the estimated value of the letter based on the feedback from this communication. For example, when there are many clarification requests, the value can be appropriately reduced. The value of ; when the delay increases significantly, it can be improved. The estimated value.

[0162] III. Online Optimization of Predictive Models: Furthermore, the predicted probability of understanding can be predicted. The training samples are composed of actual feedback results (e.g., whether a misunderstanding occurred), and the loss value is calculated using the cross-entropy loss function. And by applying gradient descent to the parameters of the MLP or Transformer model Perform online or offline updates:

[0163] in, This is the learning rate parameter. Through continuous training, the prediction accuracy of understanding probability and communication cost can be gradually improved for different agents and in different scenarios.

[0164] Compared to existing multi-agent systems, which generally suffer from static agent structures, low resource utilization, and insufficient communication and cooperation efficiency, this invention achieves the following beneficial technical effects by introducing a task-driven dynamic structure evolution mechanism, an agent lifecycle management mechanism, and an implicit communication adaptive mechanism based on mental simulation: I. Achieving dynamic self-adaptation of intelligent agent structure to improve the ability to handle complex tasks. This application does not pre-determine a fixed number of agents or roles. Instead, it continuously assesses changes in capability requirements during task execution based on task capability vectors and instantiates or reclaims sub-agents as needed, allowing the system structure to dynamically evolve as the task progresses.

[0165] Therefore, it avoids the problem of insufficient or redundant capabilities caused by the fixed roles of intelligent agents in existing technologies, and enables the system to maintain a capability configuration that matches the complexity of the task at different stages, thus significantly improving its adaptability to complex and multi-stage tasks.

[0166] II. Significantly improves system resource utilization and reduces invalid computation and video memory usage. This application uses a dynamic role creation and lifecycle management mechanism to generate sub-agents only when task complexity increases or capability gaps are detected, and automatically triggers apoptosis and recycling when sub-tasks are completed, contribution decreases, or resources are scarce.

[0167] This mechanism can effectively prevent long-term resident "zombie agents" from occupying computing and storage resources, and allows the number of agents to be dynamically adjusted according to the task load, thereby significantly improving the overall resource utilization efficiency of the system and reducing operating costs.

[0168] III. Reduce invalid communication and redundant clarification to improve the efficiency of multi-agent cooperation. This application introduces a mental simulation module within the intelligent agent, enabling the sending intelligent agent to predict the receiving intelligent agent's capability level, contextual understanding, and load status before communication, and adaptively select the information expression form and content granularity accordingly.

[0169] Compared to existing solutions that rely on fixed communication formats or explicit protocol negotiation, this mechanism can effectively reduce communication ambiguity and the number of repeated interactions, thereby reducing collaboration delays and improving overall task execution efficiency.

[0170] IV. Support natural collaboration among heterogeneous intelligent agents and enhance system scalability Since both communication and role adjustment are based on dynamic capability modeling and cognitive state prediction, this application can support agents with different capability levels and different load states to participate in collaboration without manual configuration.

[0171] Therefore, this application can still maintain stable and efficient operation in scenarios with expanded agent scale or heterogeneous capabilities, significantly improving the scalability and robustness of multi-agent systems.

[0172] In summary, this application, through the synergistic effect of dynamic structural evolution, lifecycle management, and implicit communication adaptation, enables multi-agent systems to outperform existing technologies in terms of task adaptability, resource utilization, and collaborative efficiency, demonstrating clear technological advancement and practical application value.

[0173] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods provided in the above embodiments.

[0174] Based on the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described in the above embodiments.

[0175] The above are merely specific embodiments 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 scope of the technology 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.

[0176] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A multi-agent cooperative control method, characterized in that, The method includes: Receive natural language task instructions, parse task semantics and generate task capability vector; based on the task capability vector, load capability enhancement sub-parameters on demand from the shared large language model parameter pool, dynamically instantiate one or more lightweight sub-agents, and each sub-agent only activates the parameter subset corresponding to its capability dimension; During task execution, the lifecycle manager continuously monitors task complexity, capability gaps, and agent load status, dynamically triggering agent splitting operations to generate new sub-agents, or triggering agent apoptosis operations to reclaim idle or low-contribution agents. Before the sending agent sends a message to the target agent, the cognitive state of the target agent is predicted by the mental simulation module built into the sending agent, and the expression form and information granularity of the message are adaptively selected according to the prediction result to complete the message encoding and sending. After communication is completed, feedback signals are collected, the cognitive state of the target agent is updated, and the prediction model of the mental simulation module is optimized online or offline.

2. The method according to claim 1, characterized in that, Receive natural language task instructions, parse the task semantics, and generate a task capability vector, including: Semantic parsing of natural language task instructions is performed to extract task objectives, subtasks, capability elements, and dependencies. The parsing results are mapped to a predefined set of capability dimensions to generate a weight vector; The weight vector is obtained through one or more of the following methods: large language model output, rule mapping, historical statistics, or online feedback, and then normalized.

3. The method according to claim 1, characterized in that, Dynamically instantiate one or more lightweight sub-agents, each of which activates only a subset of parameters corresponding to its capability dimension, including: Threshold filtering is performed on the task capability vectors to determine the set of capability dimensions that need to be activated; Retrieve the capability enhancement sub-parameter module corresponding to the capability dimension from the shared parameter pool; Load and bind the selected modules to the base large language model to generate lightweight sub-agents; The number of sub-agents is determined based on capability weights or task parallelism; each sub-agent is assigned an independent context cache and task state.

4. The method according to claim 1, characterized in that, The conditions for dynamically triggering the splitting operation of an agent to generate a new sub-agent, or triggering the apoptosis operation of an agent to reclaim idle or low-contribution agents, include at least the following: the task complexity index exceeds a preset threshold, the capability coverage is less than 0.85, and the load rate of any agent is greater than 0.8 or its response latency exceeds a preset latency threshold.

5. The method according to claim 4, characterized in that, The agent splitting operation specifically includes: Computational capability gap vector; Load capability modules that can fill the capability gaps from the shared parameter pool; The current execution context is cloned from the parent agent to generate a child agent instance; The sub-tasks corresponding to the capability gaps are delegated to the newly generated sub-agents, and the state matrix of the agent cluster is updated.

6. The method according to claim 4, characterized in that, The conditions for triggering the agent apoptosis operation include at least the following: the subtask is completed and there are no unfinished downstream dependent tasks, the agent's contribution is lower than the lower threshold, the agent's continuous idle time exceeds the idle threshold, and the overall system memory occupancy rate exceeds 85%. In these cases, the agent is forcibly reclaimed in order of contribution from low to high.

7. The method according to claim 6, characterized in that, The contribution was calculated using a comprehensive indicator: in, This represents the incremental progress contribution of the intelligent agent. Indicates resource utilization rate. This represents a very small constant that prevents division by zero. This represents the lower limit threshold for contribution.

8. The method according to claim 1, characterized in that, The mental simulation module predicts the cognitive state of the target agent, including: Construct a cognitive state vector for the target intelligent agent, which includes: ability level vector, context mastery level, historical interaction feature vector, load and attention state; The cognitive state vector is maintained and updated based on real-time observation data and historical communication records.

9. The method according to claim 1, characterized in that, The specific process for adaptively selecting message expression format and information granularity includes: Construct input features, including the target agent's state, the content features of the message to be sent, and the current task context; For each set of candidate representations, predict the probability of successful comprehension and the communication cost. The optimal expression form is selected through a comprehensive evaluation function; The message is re-encoded and sent according to the selected expression format and granularity level.

10. The method according to claim 9, characterized in that, The formula for predicting the probability of successful understanding is as follows: in, This represents the embedded feature vector expressed in form k. This represents the Sigmoid function, and the parameters of the MLP are continuously fine-tuned based on historical communication feedback. The formula for predicting the communication cost is: in, This indicates the token length of the message after encoding using the k-expression. Indicates the expected number of clarification rounds; Indicates the expected response delay time. , , This represents the weighting coefficient for different cost items.

11. The method according to claim 1, characterized in that, Specific methods of feedback learning after communication include: The cognitive state vector of the target agent is updated using an exponential moving average. The predicted comprehension probability and the actual misunderstanding result are used to form training samples, and the cross-entropy loss function is used to update the prediction parameters of the mental simulator using gradient.

12. A multi-agent cooperative control system for task-capability adaptive generation, characterized in that, include: The parameter scheduling module is used to receive natural language task instructions, parse task semantics and generate task capability vectors; based on the task capability vectors, it loads capability enhancement sub-parameters on demand from the shared large language model parameter pool, dynamically instantiates one or more lightweight sub-agents, and each sub-agent only activates the parameter subset corresponding to its capability dimension. The cycle management module is used to continuously monitor task complexity, capability gaps and agent load status through the lifecycle manager during task execution, dynamically trigger agent splitting operations to generate new sub-agents, or trigger agent apoptosis operations to reclaim idle or low-contribution agents. The intelligent communication module is used to predict the cognitive state of the target intelligent agent by means of the mental simulation module built into the sending intelligent agent before the sending intelligent agent sends a message to the target intelligent agent, and adaptively selects the expression form and information granularity of the message according to the prediction result to complete the message encoding and sending. The state update module is used to collect feedback signals after communication is completed, update the cognitive state of the target agent, and optimize the prediction model of the mind simulation module online or offline.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method of any one of claims 1 to 11.

14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method of any one of claims 1 to 11.

Images