Construction method and system of universal intelligent agent framework

Abstract

The invention relates to the technical field of intelligent agent systems, in particular to a construction method and system of a universal intelligent agent framework, and the method comprises an agent operation engine and a meta-capability management platform. And plug and play of the capability, flexible arrangement of tasks and self-adaptive response of the environment are realized. The adaptive capacity of the intelligent agent in multi-scene migration, cross-system collaboration and dynamic environments can be remarkably improved, and the development and maintenance cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent system technology, and more specifically, to a method and system for constructing a general intelligent agent framework. Background Technology

[0002] With the continuous development of artificial intelligence, automated decision-making, and software autonomy technologies, intelligent agents, as a key technology capable of sensing the environment, making autonomous decisions, and executing corresponding behaviors, have been widely applied in fields such as intelligent manufacturing, smart operation and maintenance, intelligent scheduling, human-machine collaboration, and complex system management. Existing intelligent agent implementations are typically designed around specific business needs, completing specific tasks through predefined rules, models, or processes, thus improving the automation and intelligence level of the system to a certain extent. However, with the increasing complexity of application scenarios and the rapid evolution of task types, the limitations of single-scenario driven intelligent agent technology have gradually become apparent. On the one hand, existing agent systems mostly adopt a tightly coupled architecture design, with perception, decision-making, and execution logic deeply bound to specific business logic. This lack of unified functional abstraction and interface specifications makes it difficult to migrate or reuse agent capabilities. When application scenarios change, large-scale reconstruction is often required, resulting in high development and maintenance costs. On the other hand, intelligent agents built by different vendors or in different systems differ significantly in model structure, data interfaces, and collaboration mechanisms, lacking a universal framework to support them and hindering cross-system and cross-task collaborative operation and unified management. Furthermore, in dynamic and complex environments, existing intelligent agents have limited adaptability to changes in task objectives, environmental uncertainties, and resource status fluctuations. They typically rely on static strategies or offline configurations for responses, making it difficult to achieve real-time adjustment and continuous evolution of agent behavior, thus affecting the overall stability and intelligence level of the system. Therefore, how to provide a general, scalable, and reusable intelligent agent framework construction method to support the rapid deployment, flexible expansion, and adaptive operation of intelligent agents in different application scenarios has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0003] In view of this, the present invention proposes a method and system for constructing a general intelligent agent framework, aiming to solve the problems in the current intelligent agent technology, which are difficult to reuse and extend due to tight architecture, lack of unified abstraction and general framework support, insufficient cross-system collaboration capability, and limited adaptability and continuous evolution capability to task and environment changes in dynamic and complex environments. This invention proposes a method for constructing a general intelligent agent framework, comprising: Obtain atomic capability description information submitted by external systems, perform structured parsing on the atomic capability description information, generate standardized capability metadata, and store the standardized capability metadata in the capability registry. Based on the task objective description information, retrieve the capability registration library, match the candidate capability set corresponding to the task objective description information, and combine and sort the candidate capability set according to the preset strategy rules to generate an execution capability sequence. Real-time acquisition of multi-source context data in the operating environment; feature extraction and fusion processing of multi-source context data; generation of dynamic context vectors. The dynamic context vector is correlated with the executable capability sequence to obtain the fit score of each capability in the current context. Based on the fit score, the executable capability sequence is reordered or replaced, and the optimized capability execution sequence is output. The optimized capability execution sequence is encapsulated into a standard instruction package and sent to the agent runtime engine based on a standardized message bus. The agent runtime engine then completes the perception, decision-making, and execution operations, and synchronizes the state transition logs to the meta-capability management platform. Furthermore, when performing structured parsing of the atomic capability description information to generate standardized capability metadata, the following is included: Extract the input parameter type, output parameter type, preconditions, post-effects, and resource consumption indicators from the atomic capability description information; The input parameter types are semantically matched with a predefined parameter ontology library to obtain standardized input identifiers; The output parameter types are semantically matched with the parameter ontology library to obtain standardized output identifiers; Convert the preconditions into logical expressions and the post-effects into state change rules; Resource consumption metrics are quantified into a numerical vector of three dimensions: CPU utilization, peak memory usage, and network bandwidth requirements. Standardized input identifiers, standardized output identifiers, logical expressions, state change rules, and numerical vectors are combined into standardized capability metadata. Furthermore, when retrieving the capability registration library based on the task objective description information and matching the candidate capability set, it includes: Natural language parsing is performed on the task objective description information to extract target keywords and constraints; Map the target keywords to the target nodes in the capability ontology tree; Using the target node as the root node, perform a breadth-first traversal in the capability registration library to filter out all capability entries that meet the constraints and whose output type covers the target keyword, and form a candidate capability set based on several capability entries. Furthermore, when performing feature extraction and fusion processing on multi-source context data to generate dynamic context vectors, the following steps are included: Data from the device status monitor, user interaction logs, and external API feedback are input into the corresponding feature encoders to obtain device status feature vectors, user intent feature vectors, and environmental disturbance feature vectors. The device status feature vector, user intent feature vector, and environmental disturbance feature vector are concatenated into a joint feature matrix; An attention weighting mechanism is applied to the joint feature matrix to obtain the importance weights of each feature dimension; The importance weights are multiplied by the joint feature matrix and then summed to output the dynamic context vector. Furthermore, when obtaining the suitability score of each capability in the current context, it includes: Extract historically similar context fragments from the state transition log; Obtain the cosine similarity between the dynamic context vector and each historical similar context segment; Select capability execution records corresponding to historical context segments with a cosine similarity greater than a preset threshold, and count the number of successful executions and failures of each capability in the capability execution records; The base fit score is obtained by dividing the number of successful executions by the total number of executions, and then multiplying the base fit score by the current resource availability. Furthermore, when reordering or replacing the sequence of executable capabilities, this includes: When the fit score of a certain ability is lower than the first fit threshold, the alternative ability with the same output type and the highest fit score is selected from the candidate ability set. When there is a resource conflict between two adjacent capabilities, insert resources to release the intermediate capability; When the overall sequence execution time exceeds the task time limit, remove the non-critical capabilities with the lowest adaptability score. Retrieve the execution order dependencies of each capability in the sequence, generate an acyclic directed graph, and output the optimized capability execution sequence by topological sorting. Furthermore, the construction of the state transition log includes: Before each action is executed, record the current set of system state variables; After the action is completed, record the new set of system state variables; Package the action identifier, execution timestamp, current system state variable set, new system state variable set, and execution result code into a single log entry; Log entries are written to a distributed log storage system, and an inverted index is built based on the log entries written to the distributed log storage system, with the context vector hash value as the index key and the task target identifier as the inverted index. Furthermore, the standardized message bus adopts a Protobuf-based message format; Define a uniform message header that includes message type, source address, destination address, and sequence number; The message body contains capability instruction payload or status feedback payload; The agent runtime engine and the meta-capability management platform maintain a session via a long TCP connection and detect connection activity based on a heartbeat mechanism. The connection reconstruction process is triggered after three consecutive heartbeat timeouts. Compared with existing technologies, the beneficial effects of this invention are as follows: By performing structured parsing of atomic capabilities submitted by external systems and generating standardized capability metadata, a unified description and centralized management of heterogeneous capability resources is achieved. This enables capabilities from different sources and with different implementation methods to be standardized and reused in the same capability registry, thereby effectively reducing the complexity of capability integration and improving the versatility and scalability of the intelligent agent framework. Secondly, by retrieving the capability registry based on task target description information and combining and sorting candidate capabilities according to preset strategy rules, an execution capability sequence that meets task requirements can be dynamically generated without relying on specific business implementation details. This avoids the problem of strong coupling between capability orchestration and business logic in traditional agent systems, and improves the agent's ability to quickly deploy and flexibly reconfigure in multi-task and multi-scenario environments. Furthermore, by collecting multi-source context data from the operating environment in real time and constructing dynamic context vectors, the agent can perceive the real-time characteristics of environmental states, resource changes, and task constraints. This context information is then correlated with executable capability sequences, and the capability sequences are reordered or replaced based on adaptability scores. This enables dynamic optimization of capability selection and execution strategies, significantly enhancing the agent's adaptive decision-making capabilities in complex and uncertain environments. Finally, by encapsulating the optimized capability execution sequences into standard instruction packages and sending them to the agent's runtime engine via a standardized message bus, not only is the perception, decision-making, and execution processes decoupled and coordinated, but also, by synchronously recording state transition logs to the meta-capability management platform, traceable management and continuous optimization of the agent's operation process are supported. This improves the overall stability, maintainability, and intelligent evolution capabilities of the system. On the other hand, this application also provides a system for building a general intelligent agent framework, including: The proxy runtime engine and meta-capability management platform include: The agent operation engine includes a perception unit, a decision-making unit, an execution unit, a state maintenance unit, and a communication adaptation unit; The meta-capability management platform and the agent runtime engine are connected through a standardized message bus. The meta-capability management platform includes a capability registration module, a policy mapping module, a context parsing module, an evolution scheduling module, and a collaborative control module. Furthermore, the meta-capability management platform includes a capability registration module, a policy mapping module, a context parsing module, an evolution scheduling module, and a collaborative control module, which are specifically as follows: The capability registration module is configured to receive atomic capability description information submitted by external systems, perform structured parsing of the atomic capability description information, generate standardized capability metadata, and store the standardized capability metadata in the capability registration library. The strategy mapping module is configured to retrieve the capability registration library based on the task target description information, match the candidate capability set corresponding to the task target description information, and combine and sort the candidate capability set based on preset strategy rules to generate an executable capability sequence. The context parsing module is configured to collect multi-source context data in the runtime environment in real time, perform feature extraction and fusion processing on the multi-source context data, and generate dynamic context vectors. The evolution scheduling module is configured to perform correlation analysis between the dynamic context vector and the sequence of executable capabilities, calculate the fitness score of each capability in the current context, reorder or replace the sequence of executable capabilities based on the fitness score, and output the optimized capability execution sequence. The collaborative control module is configured to encapsulate the optimized capability execution sequence into standard instruction packages and send them to the agent runtime engine via a standardized message bus. Furthermore, the agent runtime engine includes: The sensing unit is configured to subscribe to sensing instructions in the standard instruction package through the communication adapter unit, call the corresponding sensor interface or data access plug-in, acquire raw environmental data, perform format normalization processing on the raw environmental data, and output structured sensing results to the decision unit. The decision-making unit is configured to receive structured perception results and optimized capability execution sequences, determine the next execution action based on the built-in state transition diagram model, and generate action instructions; The execution unit is configured to parse action instructions, call the corresponding capability implementation module in the capability registration library, and complete the specific behavior output; The state maintenance unit is configured to record state snapshots before and after each action is executed, build a state transition log, and synchronize the state transition log to the meta-capability management platform; The communication adapter unit is configured to provide interfaces for three protocols: RESTful API, MQTT, and gRPC. It automatically negotiates the communication protocol based on connection requests from external systems and establishes a bidirectional data channel. It is understood that the construction method and system of a general intelligent agent framework in the above embodiments of the present invention have the same beneficial effects, and will not be described again. Attached Figure Description

[0004] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a method for constructing a general intelligent agent framework according to an embodiment of the present invention; Figure 2 A flowchart illustrating a method for constructing a general intelligent agent framework according to an embodiment of the present invention; Figure 3 This is a functional block diagram of a system for constructing a general intelligent agent framework, as provided in an embodiment of the present invention. Detailed Implementation

[0005] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments. Atomic Capability Description Information (ACDI) refers to the basic descriptive information that provides a structured characterization of the smallest functional unit in a system across multiple dimensions, including functional boundaries, input / output constraints, execution conditions, resource dependencies, and behavioral characteristics, through a unified capability modeling framework and standardized description rules. It is used to support the standardized registration, retrieval, combination, and scheduling of capabilities. Structured parsing refers to the process of syntactically decomposing, semantically recognizing, and mapping unstructured or semi-structured information according to a pre-defined data model and semantic rules, transforming it into a structured data representation with clear hierarchical relationships, field meanings, and constraint rules to support subsequent storage, retrieval, analysis, and automated processing. Standardized Capability Metadata refers to the meta-layer data formed by structurally encapsulating information such as functional attributes, interface characteristics, constraints, dependencies, and performance indicators of atomic capabilities through a unified capability description model and a standardized field system. It is used to achieve unified capability registration, cross-system identification, rapid matching, and reusable management. A Capability Registry Repository is an infrastructure used to centrally store and manage standardized capability metadata. Through a unified indexing mechanism and retrieval rules, it registers, maintains, and queries available atomic capabilities and their combinations in the system to support task-driven capability discovery, dynamic orchestration, and collaborative invocation. Multi-source contextual data refers to comprehensive data collected in real time from multiple heterogeneous sources, such as the environment perception module, system operating status, task execution feedback, and external information sources, during the operation of intelligent agents. This data describes the current operating environment, resource status, and task constraints, and is used to characterize the agent's situation and support adaptive decision-making and dynamic capability scheduling. Dynamic Context Vector (RCV) refers to a time-updated vectorized representation formed by feature extraction, encoding, and fusion of multi-source context data. It is used to comprehensively characterize the environmental state, task conditions, and resource constraints of an intelligent agent at the current moment, thereby providing a unified context input for capability adaptation assessment, decision optimization, and behavior adjustment. A standardized message bus is a message transmission mechanism built using a unified communication protocol, message format, and interaction specification. It is used to achieve decoupled communication between various functional modules, capability components, and the runtime engine in an intelligent agent framework, supporting reliable transmission of instructions, status feedback, and event notifications, thereby improving the system's scalability, interoperability, and operational stability. The Agent Runtime Engine is the core execution component in the general intelligent agent framework used to host and schedule agent instances. It coordinates perception processing, decision logic execution, and behavior output by parsing and scheduling standard instruction packages, and manages and provides feedback on the agent's running status to support the stable operation and dynamic evolution of the intelligent agent. A state transition log is a log information continuously recorded during the operation of an intelligent agent, including the agent's state, capability execution phase, and the sequence of key events. It describes the evolution of the agent from one operating state to another, supporting operation traceability, behavior analysis, and system optimization. The Meta-capability Management Platform is a management system used to uniformly manage and schedule various capabilities and their operational status in the intelligent agent framework. Through centralized maintenance of capability metadata, capability lifecycle and status change logs, it enables capability registration governance, version evolution, operation monitoring and policy optimization support. Semantic matching refers to the analysis and alignment of the semantic relationship between task objective description and capability description based on a unified semantic model and concept mapping rules. It identifies the degree of consistency between the two in terms of functional intent, constraints and contextual meaning, thereby achieving accurate retrieval, association and adaptation of capabilities. A logical expression is a standardized representation of conditions, constraints, decision rules, or relational judgments using formal logical symbols and operational rules. It is used to provide a computable and reasonable description of task conditions, capability applicability, or strategy rules to support automated judgment and decision execution. Post-effect transformation refers to the processing mechanism that updates and maps the system state, environmental parameters, or context information based on the execution result after a certain capability or action is completed. It is used to characterize the state changes caused by capability execution and their impact on subsequent decision-making and capability selection. State transition rules are a set of rules based on predefined logical conditions and constraints, used to specify when a system or agent transitions from its current state to a target state when certain triggering conditions are met, in order to ensure the controllability, consistency and reasonability of the state evolution process. Breadth-First Traversal (BFT) refers to a traversal strategy in graph structure or state space search that visits nodes sequentially in hierarchical order. That is, it first visits all adjacent nodes of the starting node and then expands outward layer by layer. It is used to find feasible solutions with the fewest steps or the shallowest level in state deduction, capability combination or path search. A device state monitor is a functional module used to collect and monitor the operating status of equipment in real time. By acquiring the equipment's working parameters, operating conditions, and abnormal indication information, it forms a continuous perception of the equipment's health status and availability, so as to support the intelligent agent's status judgment and decision adjustment. User interaction logs are data sets that continuously record user actions, input commands, and interactive feedback within a system. They are used to characterize user behavior patterns, operational preferences, and interaction contexts, providing a basis for intent recognition and strategy optimization. External API feedback refers to the response information received by the intelligent agent after calling the interface of an external system. This includes the execution result, status code, and returned data, which are used to reflect the reachability, execution effect, and abnormal situations of the external service, thereby affecting subsequent capacity scheduling and decision-making processes. A User Intent Feature Vector (UIM) is a vectorized representation formed by extracting and encoding features from user input, interaction behavior, and historical preferences. It is used to quantitatively describe the user's current task goals and intent tendencies, and serves as an important input for intelligent agents to perform semantic understanding and capability matching. The Environmental Disturbance Feature Vector (EDV) is a vectorized feature representation formed after quantitatively modeling the uncertain factors affecting agent behavior in the operating environment. It is used to describe the intensity and direction of environmental changes, abnormal disturbances, or external interferences to support the agent's robust decision-making and adaptive adjustment. A joint feature matrix is ​​a matrix-based data representation formed by aligning and fusing feature vectors from multiple sources according to a unified dimensional rule. It is used to comprehensively characterize multi-dimensional information such as user intent, device status, and environmental disturbances, providing a unified data foundation for subsequent correlation analysis, matching calculation, and decision optimization. Cosine similarity is a similarity metric that measures the degree of directional consistency between two vectors by calculating the cosine of the angle between them in a vector space. It is often used to evaluate the semantic or feature matching between task target features and capability features, or between context vectors. Capability Execution Records (VARs) are data that comprehensively records the time, execution order, execution results, and abnormal states of a capability during its actual operation. They are used to analyze the historical performance, stability, and execution effectiveness of a capability, providing a basis for subsequent capability evaluation and optimization. The Base Adaptation Score is an initial adaptation score calculated based on the degree of matching between task requirements and inherent capability attributes without considering real-time dynamic factors. It is used to reflect the applicability of a capability to the task objective under static conditions. Resource availability ratio refers to the ratio of computing, communication, or physical resources available for capability execution within a given time window to its total resource capacity. It is used to measure the executability and load-bearing capacity of a capability in the current operating environment. Non-critical capabilities refer to capability modules that do not constitute a critical dependency on the overall task completion during the execution of intelligent agents or system tasks. That is, their abnormality or absence will not cause the core functions of the system to fail, but can be used to optimize efficiency, improve auxiliary performance, or enhance system flexibility. A Directed Acyclic Graph (DAG) is a graph structure consisting of nodes and directed edges, where all edges are directional and there is no path from a node back to itself along an edge. It is used to represent task dependencies, capability call order, or data flow processing, ensuring the acyclicity and schedulability of execution logic. A distributed log storage system is a system that stores log data on multiple physical nodes using sharding and replication mechanisms. Through a distributed architecture, it achieves high availability, scalability, and persistent access. It is used to record agent execution status, event transitions, and system operation history, and supports auditing, backtracking, and analysis. The Protocol Buffers-based Message Format refers to a data serialization standard defined by Google Protocol Buffers (Protobuf), which encodes structured data into a compact, cross-language, and efficient binary message format. This format is used to achieve standardized and efficient communication and data transmission between modules and systems within an intelligent agent framework. A TCP persistent connection (TCP Persistent Connection) refers to a mechanism under the Transmission Control Protocol (TCP) that allows a client and a server to maintain continuous communication for a long period of time after establishing a connection once. By reusing the same connection to send and receive data multiple times, the overhead of establishing and releasing connections is reduced, and communication efficiency and real-time performance are improved. A heartbeat mechanism is a mechanism in distributed systems or network communication that uses periodic sending of lightweight messages to detect connection status, confirm node liveness, or maintain session activity. It is used to promptly detect disconnected or abnormal connections, ensuring the stability and reliability of the system. Connection liveness is an indicator that measures whether a network connection or communication session is in a normal and available state. It is usually judged based on heartbeat response, data exchange frequency, or timeout detection, and is used to decide whether to maintain, rebuild, or close the connection. RESTful API (Representational State Transfer API) refers to application programming interfaces designed based on the REST architectural style. Through unified resource identifiers (URIs), standard HTTP methods (GET, POST, PUT, DELETE, etc.), and stateless interaction, it enables data access and function calls between systems and features simplicity, scalability, and cross-platform compatibility. MQTT (Message Queuing Telemetry Transport) is a lightweight publish / subscribe messaging protocol suitable for low-bandwidth, high-latency, or unreliable network environments. It manages message distribution through topics, enabling real-time communication and event-driven data exchange between devices and systems. gRPC (Google Remote Procedure Call) is a high-performance remote procedure call (RPC) framework based on HTTP / 2 and Protobuf. It supports multi-language development, enables efficient synchronous or asynchronous service calls in distributed systems, and provides strict interface definitions, streaming transmission, and built-in load balancing capabilities. like Figures 1-2 As shown in some embodiments of this application, this embodiment provides a method for constructing a general intelligent agent framework, including: Step S100: Obtain atomic capability description information submitted by external systems, perform structured parsing on the atomic capability description information, generate standardized capability metadata, and store the standardized capability metadata in the capability registration library. Specifically, the process of structured parsing of atomic capability description information to generate standardized capability metadata includes: extracting input parameter types, output parameter types, preconditions, post-effects, and resource consumption indicators from the atomic capability description information; semantically matching the input parameter types with a predefined parameter ontology to obtain standardized input identifiers; semantically matching the output parameter types with the parameter ontology to obtain standardized output identifiers; converting preconditions into logical expressions and post-effects into state change rules; quantifying resource consumption indicators into numerical vectors of three dimensions: CPU utilization, peak memory usage, and network bandwidth requirements; and combining the standardized input identifiers, standardized output identifiers, logical expressions, state change rules, and numerical vectors into standardized capability metadata. Understandably, a standardized representation of capabilities is achieved through systematic analysis and structured processing of atomic capability description information. This involves: extracting capability input / output parameters, preconditions, post-capabilities, and resource consumption indicators through feature extraction; mapping input / output parameters to a unified parameter ontology using semantic matching methods to achieve consistent recognition across systems and semantics; transforming preconditions and post-capabilities into logical expressions and state change rules, respectively, enabling formal computation of capability constraints and effects; quantifying resource consumption into numerical vectors to provide measurable data for subsequent scheduling and optimization; and finally, fusing various types of information to generate standardized capability metadata, providing a unified data foundation for capability registration, retrieval, and dynamic combination. For example, an external system submits a "file compression" capability. Its atomic capability description includes the input parameter being the path to the file to be compressed, the output parameter being the path to the compressed file, the precondition being the existence of the file, the post-effect being the generation of a compressed file, and resource consumption of approximately 30% CPU usage, approximately 100MB of peak memory, and 0 network bandwidth requirement. After structured parsing, the system matches the input and output parameters with a parameter ontology to obtain standardized identifiers, transforms "file existence" into a logical expression, converts the effect of generating a compressed file into state change rules, and combines CPU, memory, and bandwidth metrics into a numerical vector. Finally, it generates unified, standardized capability metadata, which is stored in a capability registry for subsequent task scheduling and capability combination. Step S200: Retrieve the capability registration library based on the task target description information, match the candidate capability set corresponding to the task target description information, combine and sort the candidate capability set based on preset strategy rules, and generate an execution capability sequence. Specifically, when retrieving the capability registration library based on the task target description information and matching the candidate capability set, the process includes: performing natural language parsing on the task target description information to extract target keywords and constraints; mapping the target keywords to target nodes in the capability ontology tree; using the target node as the root node, performing a breadth-first traversal in the capability registration library to select all capability entries that meet the constraints and whose output type covers the target keywords, and constructing a candidate capability set based on several capability entries. Understandably, semantic parsing and structured retrieval achieve precise matching between task objectives and capability resources. This involves parsing the task objective description using natural language to extract keywords and constraints, transforming unstructured target information into computable semantic features. Then, target keywords are mapped to corresponding nodes in the capability ontology tree to establish semantic associations between objectives and capabilities. A breadth-first traversal is performed based on the capability registry to select capability entries that satisfy both constraints and output coverage of target requirements, forming a candidate capability set. Finally, candidate capabilities are combined and sorted using preset strategy rules to generate an execution capability sequence, providing schedulable capability paths for subsequent task execution. For example, if the task objective is described as "generating a compressed file and uploading it to the cloud," the system first extracts the keywords "compressed file" and "upload to the cloud" through natural language parsing, and obtains constraints such as file size limits and upload bandwidth limits. Then, it maps the keywords to the corresponding "file compression" and "file upload" nodes in the capability ontology tree. Using these nodes as roots, it performs a breadth-first traversal in the capability registration library to filter out capability entries that meet the constraints and match the output type, such as "ZIP compression capability," "GZIP compression capability," "FTP upload capability," and "HTTP upload capability," forming a candidate capability set. Finally, it combines and sorts the candidate capabilities according to preset strategy rules (such as prioritizing low resource consumption and compatibility) to generate the execution capability sequence of "ZIP compression → HTTP upload," which is then executed by the agent runtime engine. Step S300: Collect multi-source context data in the operating environment in real time, perform feature extraction and fusion processing on the multi-source context data, and generate dynamic context vectors. Specifically, when performing feature extraction and fusion processing on multi-source context data to generate a dynamic context vector, the process includes: inputting data from the device status monitor, user interaction logs, and external API feedback into the corresponding feature encoders to obtain device status feature vectors, user intent feature vectors, and environmental disturbance feature vectors; concatenating the device status feature vectors, user intent feature vectors, and environmental disturbance feature vectors into a joint feature matrix; applying an attention weighting mechanism to the joint feature matrix to obtain the importance weights of each feature dimension; and multiplying the importance weights by the joint feature matrix and summing the results to output the dynamic context vector. Understandably, a multi-source data feature encoding and fusion mechanism enables dynamic context awareness of the operating environment. This involves inputting multi-source context data (such as device status monitors, user interaction logs, and external API feedback) from different sources into the corresponding feature encoder, converting heterogeneous data into vectorized representations (device status feature vectors, user intent feature vectors, and environmental disturbance feature vectors) to achieve a unified representation of different types of information. Then, the feature vectors are concatenated to form a joint feature matrix, and an attention mechanism is used to calculate the importance weights of each feature dimension, highlighting the information most relevant to task and environmental decisions. Finally, the weights are weighted and summed with the joint feature matrix to generate a dynamic context vector, providing quantitative input for the intelligent agent's capability selection, decision optimization, and execution scheduling in the current environment. For example, in a smart home scenario, a device status monitor might detect an air conditioner temperature of 26°C and humidity of 60%, while user interaction logs show a recent user command to "lower the temperature." An external API feedback indicates the weather forecast predicts the temperature will rise to 30°C in the afternoon. The system encodes device status, user intent, and environmental disturbances into feature vectors, concatenating them into a joint feature matrix. An attention mechanism determines that the "user intent" and "ambient temperature" dimensions have high importance, while the device status dimension has a moderate weight. Finally, the weighted joint matrix is ​​summed to obtain a dynamic context vector, which guides the smart home agent to adjust air conditioner settings, achieving a personalized and dynamically adaptive temperature control strategy. Step S400: Perform correlation analysis between the dynamic context vector and the executable capability sequence to obtain the fit score of each capability in the current context. Reorder or replace the executable capability sequence based on the fit score and output the optimized capability execution sequence. Specifically, when obtaining the adaptability score of each capability in the current context, the process includes: extracting historical similar context fragments from the state transition log; obtaining the cosine similarity between the dynamic context vector and each historical similar context fragment; selecting capability execution records corresponding to historical context fragments with a cosine similarity greater than a preset threshold, and counting the number of successful executions and failures of each capability in the capability execution records; dividing the number of successful executions by the total number of executions as the basic adaptability score, and multiplying the basic adaptability score by the current resource availability rate to obtain the final adaptability score. Specifically, when reordering or replacing the executable capability sequence, the following steps are taken: when the fit score of a certain capability is lower than the first fit threshold, select a replacement capability with the same output type and the highest fit score from the candidate capability set; when there is a resource conflict between two adjacent capabilities, insert an intermediate capability to release resources; when the overall sequence execution time exceeds the task time limit, remove the non-critical capability with the lowest fit score; re-acquire the execution order dependencies of each capability in the sequence, generate an acyclic directed graph, and output the optimized capability execution sequence according to the topological sort. Understandably, when combining historical experience data with the current dynamic context to achieve adaptive optimization and dynamic adjustment of the intelligent agent's capability sequence, the cosine similarity is calculated between the dynamic context vector and the context fragments stored in the historical state transition log. This identifies historical scenarios highly similar to the current environment, and then obtains the capability execution records corresponding to these scenarios. The number of successful and failed executions of each capability is counted in the capability execution records. The number of successful executions is divided by the total number of executions to obtain the basic fit score, which is then combined with the current resource availability to generate the final fit score, thereby quantifying the execution reliability and adaptability of each capability in the current context. Subsequently, the executable capability sequence is reordered, replaced, inserted, or removed based on the fit score, resource constraints, task time limits, and dependencies between capabilities. For example, when a capability's fit score is below a preset threshold, a replacement capability with the highest score and matching output type can be selected from the candidate capability set; when adjacent capabilities have resource conflicts, intermediate capabilities can be inserted to release resources and avoid conflicts; if the overall execution time of the sequence exceeds the task time limit, non-critical capabilities with the lowest scores can be removed to shorten the execution time. Finally, by generating an acyclic directed graph (DAG) and performing topological sorting, the reasonable dependencies of capability execution order are ensured, and the optimized capability execution sequence is finally output, providing a reliable and efficient scheduling scheme for intelligent agents to perform tasks in dynamic and complex environments. Taking a file processing task as an example, the initial capability execution sequence is "ZIP compression → HTTP upload → data encryption". The system first converts the current context information (such as network bandwidth, CPU usage, and user urgency) into a dynamic context vector and calculates its cosine similarity with context fragments in the historical state transition log. It finds that "HTTP upload" has a low success rate in low-bandwidth environments, with a basic fit score of 0.6. Combined with the current resource availability of 0.5, the final fit score is only 0.3, below the set threshold of 0.5. The system selects "FTP upload" with the highest fit score from the candidate capability set to replace "HTTP upload". Simultaneously, due to bandwidth resource conflicts between "ZIP compression" and "FTP upload", the system inserts a "bandwidth release capability" between them to ensure resource availability. Furthermore, if the total task execution time exceeds the time limit, the system removes the non-critical capability "data encryption" to ensure the core task is completed on time. Finally, the system generates an acyclic directed graph based on capability dependencies and performs topological sorting to obtain an optimized capability execution sequence: "ZIP compression → bandwidth release → FTP upload". This sequence takes into account capability adaptability, resource constraints, and task time limits, achieving efficient and reliable execution of tasks in dynamic environments, while providing scheduling experience that can be referenced for subsequent similar tasks. Step S500: The optimized capability execution sequence is encapsulated into a standard instruction package and sent to the agent running engine based on the standardized message bus. The agent running engine completes the perception, decision-making and execution operations, and synchronizes the state transition log to the meta-capability management platform. Specifically, the construction of the state transition log includes: recording the current set of system state variables before each action is executed; recording the new set of system state variables after the action is executed; packaging the action identifier, execution timestamp, current set of system state variables, new set of system state variables, and execution result code into a log entry; writing the log entry into a distributed log storage system, and establishing an inverted index with the context vector hash value and the task target identifier as the index key based on the log entry written to the distributed log storage system. Specifically, the standardized message bus adopts a Protobuf-based message format; it defines a unified message header that includes message type, source address, destination address, and sequence number; the message body includes capability instruction payload or status feedback payload; the agent running engine and the meta-capability management platform maintain a session through a TCP long connection and detect connection activity based on a heartbeat mechanism, wherein a connection reconstruction process is triggered after three consecutive heartbeat timeouts. Understandably, by achieving efficient scheduling and traceability of intelligent agent capability execution through standardized instruction encapsulation, reliable message transmission, and distributed log management, the optimized capability execution sequence is encapsulated into standardized instruction packages. These packages are then sent to the agent runtime engine via a standardized message bus using a Protobuf-based message format and unified message header definition. This achieves structured instructions, cross-module compatibility, and efficient transmission. The agent runtime engine performs perception, decision-making, and execution operations based on the instructions. Simultaneously, it records a set of system state variables before and after each action execution, packaging the action identifier, execution timestamp, state change, and execution result into log entries. These entries are written to a distributed log storage system, and an inverted index of context vector hash values ​​and task target identifiers is established to achieve traceability of capability execution and historical data management. Furthermore, the runtime engine and the meta-capability management platform maintain a session through a long TCP connection, using a heartbeat mechanism to monitor connection activity and trigger connection reconstruction in abnormal situations, ensuring the reliability and continuity of system communication. For example, in file processing task execution, the optimized capability execution sequence is "ZIP compression → bandwidth release → FTP upload". The system encapsulates this sequence into a standard instruction package. The message header contains the message type "capability instruction", the source address is the agent scheduling module, the destination address is the runtime engine, and the sequence number is used for unique identification. The message body contains the specific capability instruction payload. After receiving the instruction, the agent runtime engine executes the compression, bandwidth release, and upload operations sequentially. Before each action is executed, the current CPU usage, memory usage, and network bandwidth status are recorded. After execution, the new status is recorded, and the action identifier, execution timestamp, status change, and execution result code are formed into a log entry and written to the distributed log storage system. At the same time, an inverted index is established with the context vector hash value and the task target identifier as the key. The runtime engine and the meta-capability management platform maintain a session through a TCP long connection, sending heartbeat messages at fixed intervals to check the connection status. When three consecutive heartbeats fail to respond, connection reconstruction is automatically triggered to ensure log synchronization and reliable instruction transmission. Through the above mechanism, the entire capability execution process achieves standardized scheduling, dynamic monitoring, and traceable management. In the above embodiments, by performing structured parsing of atomic capabilities submitted by external systems and generating standardized capability metadata, a unified description and centralized management of heterogeneous capability resources is achieved. This enables capabilities from different sources and with different implementation methods to be standardized and reused in the same capability registry, effectively reducing the complexity of capability integration and improving the versatility and scalability of the intelligent agent framework. Secondly, by retrieving the capability registry based on task objective description information and combining and sorting candidate capabilities according to preset strategy rules, a sequence of execution capabilities that meets task requirements can be dynamically generated without relying on specific business implementation details. This avoids the problem of strong coupling between capability orchestration and business logic in traditional agent systems and improves the agent's ability to quickly deploy and flexibly reconfigure in multi-task and multi-scenario environments. Furthermore, by collecting multi-source context data from the runtime environment in real time and constructing dynamic context vectors, the agent can perceive the real-time characteristics of environmental status, resource changes, and task constraints. This context information is then correlated with the executable capability sequence, and the capability sequence is reordered or replaced based on the adaptability score. This achieves dynamic optimization of capability selection and execution strategies, significantly enhancing the agent's adaptive decision-making ability in complex and uncertain environments. Finally, by encapsulating the optimized capability execution sequence into a standard instruction package and sending it to the agent runtime engine via a standardized message bus, not only is the decoupling and collaboration of the perception, decision-making and execution processes achieved, but also by synchronously recording state transition logs to the meta-capability management platform, the traceable management and continuous optimization of the agent runtime process are supported, which is conducive to improving the overall stability, maintainability and intelligent evolution capabilities of the system. In another preferred embodiment based on the above embodiments, this implementation provides a system for constructing a general intelligent agent framework, including: an agent operation engine and a meta-capability management platform, wherein: the agent operation engine includes a perception unit, a decision-making unit, an execution unit, a state maintenance unit, and a communication adaptation unit; the meta-capability management platform is connected to the agent operation engine through a standardized message bus, and the meta-capability management platform includes a capability registration module, a policy mapping module, a context parsing module, an evolution scheduling module, and a collaborative control module. Specifically, the meta-capability management platform includes a capability registration module, a strategy mapping module, a context parsing module, an evolution scheduling module, and a collaborative control module. Specifically: the capability registration module is configured to receive atomic capability description information submitted by external systems, perform structured parsing of the atomic capability description information, generate standardized capability metadata, and store the standardized capability metadata in the capability registration library; the strategy mapping module is configured to retrieve the capability registration library based on task target description information, match a set of candidate capabilities corresponding to the task target description information, and combine and sort the candidate capability sets based on preset strategy rules to generate an executable capability sequence; the context parsing module is configured to collect multi-source context data in the runtime environment in real time, perform feature extraction and fusion processing on the multi-source context data, and generate a dynamic context vector; the evolution scheduling module is configured to perform correlation analysis between the dynamic context vector and the executable capability sequence, calculate the adaptability score of each capability in the current context, reorder or replace the executable capability sequence based on the adaptability score, and output an optimized capability execution sequence; the collaborative control module is configured to encapsulate the optimized capability execution sequence into a standard instruction package and send it to the agent runtime engine through a standardized message bus. Specifically, the agent operation engine includes: a perception unit configured to subscribe to perception instructions in the standard instruction package through the communication adaptation unit, call the corresponding sensor interface or data access plugin to obtain raw environmental data, perform format normalization processing on the raw environmental data, and output structured perception results to the decision unit; a decision unit configured to receive the structured perception results and the optimized capability execution sequence, determine the next execution action based on the built-in state transition graph model, and generate action instructions; an execution unit configured to parse the action instructions, call the corresponding capability implementation module in the capability registration library, and complete the specific behavior output; a state maintenance unit configured to record state snapshots before and after each action execution, build a state transition log, and synchronize the state transition log to the meta-capability management platform; and a communication adaptation unit configured to provide RESTful API, MQTT, and gRPC protocol interfaces, automatically negotiate communication protocols based on connection requests from external systems, and establish a bidirectional data channel. It is understood that the construction method and system of a general intelligent agent framework in the above embodiments of the present invention have the same beneficial effects, and will not be described again. Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for constructing a general intelligent agent framework, characterized in that, include: Obtain atomic capability description information submitted by external systems, perform structured parsing on the atomic capability description information, generate standardized capability metadata, and store the standardized capability metadata in the capability registry. Based on the task objective description information, retrieve the capability registration library, match the candidate capability set corresponding to the task objective description information, and combine and sort the candidate capability set according to the preset strategy rules to generate an execution capability sequence. Real-time acquisition of multi-source context data in the operating environment; feature extraction and fusion processing of multi-source context data; generation of dynamic context vectors. The dynamic context vector is correlated with the executable capability sequence to obtain the fit score of each capability in the current context. Based on the fit score, the executable capability sequence is reordered or replaced, and the optimized capability execution sequence is output. The optimized capability execution sequence is encapsulated into a standard instruction package and sent to the agent runtime engine based on a standardized message bus. The agent runtime engine then completes the perception, decision-making, and execution operations, and synchronizes the state transition logs to the meta-capability management platform.

2. The method for constructing the general intelligent agent framework as described in claim 1, characterized in that, When performing structured parsing of atomic capability description information to generate standardized capability metadata, the following is included: Extract the input parameter type, output parameter type, preconditions, post-effects, and resource consumption indicators from the atomic capability description information; The input parameter types are semantically matched with a predefined parameter ontology library to obtain standardized input identifiers; The output parameter types are semantically matched with the parameter ontology library to obtain standardized output identifiers; Convert the preconditions into logical expressions and the post-effects into state change rules; Resource consumption metrics are quantified into a numerical vector of three dimensions: CPU utilization, peak memory usage, and network bandwidth requirements. Standardized input identifiers, standardized output identifiers, logical expressions, state change rules, and numerical vectors are combined into standardized capability metadata.

3. The method for constructing the general intelligent agent framework as described in claim 2, characterized in that, When retrieving the capability registration library based on the task objective description information and matching the candidate capability set, it includes: Natural language parsing is performed on the task objective description information to extract target keywords and constraints; Map the target keywords to the target nodes in the capability ontology tree; Using the target node as the root node, perform a breadth-first traversal in the capability registration library to filter out all capability entries that meet the constraints and whose output type covers the target keyword, and form a candidate capability set based on several capability entries.

4. The method for constructing the general intelligent agent framework as described in claim 3, characterized in that, When performing feature extraction and fusion processing on multi-source context data to generate dynamic context vectors, the following steps are included: Data from the device status monitor, user interaction logs, and external API feedback are input into the corresponding feature encoders to obtain device status feature vectors, user intent feature vectors, and environmental disturbance feature vectors. The device status feature vector, user intent feature vector, and environmental disturbance feature vector are concatenated into a joint feature matrix; An attention weighting mechanism is applied to the joint feature matrix to obtain the importance weights of each feature dimension; The importance weights are multiplied by the joint feature matrix and then summed to output the dynamic context vector.

5. The method for constructing the general intelligent agent framework as described in claim 4, characterized in that, When obtaining the suitability score of each capability in the current context, the following are included: Extract historically similar context fragments from the state transition log; Obtain the cosine similarity between the dynamic context vector and each historical similar context segment; Select capability execution records corresponding to historical context segments with a cosine similarity greater than a preset threshold, and count the number of successful executions and failures of each capability in the capability execution records; The base fit score is obtained by dividing the number of successful executions by the total number of executions, and then multiplying the base fit score by the current resource availability.

6. The method for constructing the general intelligent agent framework as described in claim 5, characterized in that, When reordering or replacing the sequence of executable capabilities, the following applies: When the fit score of a certain ability is lower than the first fit threshold, the alternative ability with the same output type and the highest fit score is selected from the candidate ability set. When there is a resource conflict between two adjacent capabilities, insert resources to release the intermediate capability; When the overall sequence execution time exceeds the task time limit, remove the non-critical capabilities with the lowest adaptability score. Retrieve the execution order dependencies of each capability in the sequence, generate an acyclic directed graph, and output the optimized capability execution sequence by topological sorting.

7. The method for constructing the general intelligent agent framework as described in claim 1, characterized in that, The construction of the state transition log includes: Before each action is executed, record the current set of system state variables; After the action is completed, record the new set of system state variables; Package the action identifier, execution timestamp, current system state variable set, new system state variable set, and execution result code into a single log entry; Log entries are written to a distributed log storage system, and an inverted index is built based on the log entries written to the distributed log storage system, with the context vector hash value as the index key and the task target identifier as the inverted index.

8. The method for constructing the general intelligent agent framework as described in claim 2, characterized in that, The standardized message bus uses a Protobuf-based message format; Define a uniform message header that includes message type, source address, destination address, and sequence number; The message body contains capability instruction payload or status feedback payload; The agent runtime engine and the meta-capability management platform maintain a session via a long TCP connection and detect connection activity based on a heartbeat mechanism. The connection reconstruction process is triggered after three consecutive heartbeat timeouts.

9. A system for constructing a general intelligent agent framework, applicable to the method for constructing a general intelligent agent framework as described in any one of claims 1-8, characterized in that, include: The proxy runtime engine and meta-capability management platform include: The agent operation engine includes a perception unit, a decision-making unit, an execution unit, a state maintenance unit, and a communication adaptation unit; The meta-capability management platform and the agent runtime engine are connected via a standardized message bus. The meta-capability management platform includes a capability registration module, a policy mapping module, a context parsing module, an evolution scheduling module, and a collaborative control module; among which: The capability registration module is configured to receive atomic capability description information submitted by external systems, perform structured parsing of the atomic capability description information, generate standardized capability metadata, and store the standardized capability metadata in the capability registration library. The strategy mapping module is configured to retrieve the capability registration library based on the task target description information, match the candidate capability set corresponding to the task target description information, and combine and sort the candidate capability set based on preset strategy rules to generate an executable capability sequence. The context parsing module is configured to collect multi-source context data in the runtime environment in real time, perform feature extraction and fusion processing on the multi-source context data, and generate dynamic context vectors. The evolution scheduling module is configured to perform correlation analysis between the dynamic context vector and the sequence of executable capabilities, calculate the fitness score of each capability in the current context, reorder or replace the sequence of executable capabilities based on the fitness score, and output the optimized capability execution sequence. The collaborative control module is configured to encapsulate the optimized capability execution sequence into standard instruction packages and send them to the agent runtime engine via a standardized message bus.

10. The system for constructing the general intelligent agent framework as described in claim 9, characterized in that, The proxy runtime engine includes: The sensing unit is configured to subscribe to sensing instructions in the standard instruction package through the communication adapter unit, call the corresponding sensor interface or data access plug-in, acquire raw environmental data, perform format normalization processing on the raw environmental data, and output structured sensing results to the decision unit. The decision-making unit is configured to receive structured perception results and optimized capability execution sequences, determine the next execution action based on the built-in state transition diagram model, and generate action instructions; The execution unit is configured to parse action instructions, call the corresponding capability implementation module in the capability registration library, and complete the specific behavior output; The state maintenance unit is configured to record state snapshots before and after each action is executed, build a state transition log, and synchronize the state transition log to the meta-capability management platform; The communication adapter unit is configured to provide interfaces for three protocols: RESTful API, MQTT, and gRPC. It automatically negotiates the communication protocol based on connection requests from external systems and establishes a bidirectional data channel.

Images