An AI autonomous agent execution framework architecture system based on hierarchical decoupling

Abstract

This invention belongs to the field of artificial intelligence communication technology and discloses an AI autonomous intelligent agent execution framework architecture system based on layered decoupling. The invention integrates the core capabilities of ReAct execution loop, intelligent context compression, and autonomous error correction in the decision layer to achieve accurate parsing of instruction intent, dynamic iteration of decision logic, and autonomous error correction. The transmission layer establishes a multi-platform message intelligent routing and protocol unification mechanism to achieve efficient data interaction and instruction transmission across platforms and protocols. The execution layer integrates hybrid memory management and sandbox-isolated practical execution capabilities to achieve multi-dimensional memory of execution data, secure isolation of execution behavior, and accurate implementation of practical tasks. The system's layers are clearly defined, decoupled, independent, and collaborative, possessing high scalability, cross-platform adaptability, execution security, and decision autonomy.

Description

Technical Field

[0001] This invention belongs to the field of information technology, specifically relating to an AI autonomous intelligent agent execution framework architecture system based on layered decoupling. Background Technology

[0002] As a core carrier for the practical application of artificial intelligence technology, AI autonomous agents, with their autonomous decision-making and execution capabilities, have become an important direction for the intelligent upgrading of various industries. Traditional AI autonomous agent execution frameworks often adopt an integrated, coupled design, combining decision-making reasoning, instruction transmission, and task execution into a single module without clear layering and functional boundary delineation. This leads to numerous specific and non-macro-level technical problems exposed in practical applications, making it difficult to meet the autonomous execution needs of complex scenarios. Specific problems include: Problem 1: The decision-making level has a simple execution logic, low efficiency in context processing, and no self-correction capability, resulting in poor decision accuracy and iterative nature. Traditional AI autonomous agents often employ a fixed execution logic of "one-time reasoning - single action" in their decision-making layers, lacking a dynamic iterative decision-making mechanism similar to the human "thinking-action-feedback-rethinking" process. This makes them prone to decision-making biases when faced with complex instructions. Furthermore, the decision-making process requires processing massive amounts of contextual data. Traditional technologies often use full loading without effective lightweight compression, resulting in high latency in context processing, high system resource consumption, and reduced decision-making efficiency. In addition, the decision-making layer lacks autonomous error correction capabilities. When errors occur in the reasoning results or intermediate actions, it cannot autonomously identify and correct them, requiring manual intervention and significantly reducing the accuracy and autonomy of decision-making.

[0003] Question 2: The transport layer lacks dedicated protocol adaptation and message routing mechanisms, resulting in weak cross-platform adaptability and low transmission efficiency and stability. Traditional AI autonomous agents treat transmission functions as an adjunct to decision-making or execution modules, lacking an independent transmission layer and dedicated adaptation mechanisms for multiple platforms and protocols. When facing different platforms such as Windows, Linux, IoT devices, and mobile terminals, as well as different communication protocols such as TCP / IP, MQTT, and HTTP, they cannot achieve unified protocol adaptation and message format standardization, leading to problems such as format incompatibility and communication rule conflicts in cross-platform command transmission. At the same time, the lack of intelligent message routing mechanisms and the use of "broadcast" or "fixed path" command distribution, which cannot select the optimal transmission path based on the execution end status and network environment, resulting in high transmission latency, high command distribution error rate, and difficulty in guaranteeing transmission efficiency and stability.

[0004] Question 3: The execution layer memory management model is simplistic, lacking a secure and isolated execution environment, resulting in poor execution security and scenario adaptability. Traditional AI autonomous agents often employ a single memory management model for their execution layer, storing only basic execution data. This lack of multi-dimensional classification based on data type, usage frequency, and scenario characteristics leads to low data retrieval efficiency and weak scenario adaptability during execution. Furthermore, the execution layer performs tasks directly within the host system without establishing a secure and isolated execution environment. Errors in execution instructions or malicious operations can easily damage the host system, causing data leaks, system crashes, and other security issues. In addition, the execution layer lacks an effective feedback mechanism with the decision-making layer. Execution status and results cannot be accurately and promptly fed back to the decision-making layer, hindering the formation of a closed-loop execution logic and impacting the overall effectiveness of task execution.

[0005] Existing technologies for improving AI autonomous agents mostly focus on localized optimizations of single functions, such as simply increasing the accuracy of decision-making and reasoning or simply adding transmission encryption methods. They fail to address the root causes of the three major technical problems by adopting a layered, decoupled, and systematic design from the perspective of native architecture design. They also fail to clearly define the functional boundaries and collaborative logic of decision-making, transmission, and execution. Therefore, there is an urgent need to develop an AI autonomous agent execution framework architecture system based on layered decoupling. This system should employ a native layered design, similar to a "brain-nerve-limb" structure, to build core capabilities at each level, achieving decoupling and independent operation as well as collaborative linkage between levels. This would enhance the AI ​​autonomous agent's decision-making autonomy, transmission adaptability, and execution security. Summary of the Invention

[0006] The purpose of this invention is to overcome the technical defects of the integrated coupling in traditional AI autonomous agent execution frameworks and provide an AI autonomous agent execution framework architecture system based on layered decoupling. It innovatively designs a "brain-nerve-limb" layered decoupling native architecture, clearly defining the native layered logic and functional boundaries of the decision-making layer (brain), transmission layer (nerve), and execution layer (limb). Addressing the three specific technical problems mentioned in the background technology—"single execution logic in the decision-making layer, low efficiency in context processing, and lack of autonomous error correction capability," "weak cross-platform adaptability of the transmission layer, and lack of dedicated routing and protocol adaptation mechanisms," and "single memory management in the execution layer, lack of a secure isolated execution environment, and missing feedback mechanisms"—this invention designs six core technologies. Through architectural innovation and algorithm modeling, it builds core capabilities at each level, achieving full-process autonomous closed-loop execution of "decision-transmission-execution-feedback." The specific invention content is as follows: Regarding question 1: Innovation of core capabilities at the decision-making level to achieve dynamic decision-making, efficient context processing, and autonomous error correction. 1. Integrate the ReAct execution loop mechanism to create a dynamic iterative decision-making logic of "reasoning-action". Breaking away from the fixed logic of traditional decision-making layers—"one-time reasoning - single action"—this approach uses the ReAct execution loop mechanism as the core execution logic of the decision-making layer. Simulating human thinking and action patterns, it constructs a dynamic iterative decision-making process of "reasoning → action → feedback → re-reasoning." This mechanism generates initial reasoning results based on external input instructions, selects corresponding execution actions based on these results, and iteratively updates the reasoning logic through feedback from the execution actions until the optimal executable task instruction is generated. This solves the problems of traditional decision-making layers having a single execution logic and significant decision-making biases with complex instructions, greatly improving the autonomous decision-making capability and adaptability to complex scenarios of the decision-making layer.

[0007] 2: Design a context-based intelligent compression algorithm to achieve lightweight and efficient processing of massive context data. To address the inefficiencies and resource consumption caused by traditional full-load context loading in decision-making processes, an innovative intelligent context compression algorithm is designed. This algorithm constructs a context processing logic of "semantic encoding - attention weighting - redundancy pruning - feature fusion," achieving lightweight processing of massive context data through a dedicated modeling formula. The algorithm extracts global context features through semantic encoding, strengthens core key features using an attention mechanism, identifies and prunes redundant and ineffective features, and ultimately generates standardized lightweight context feature vectors. While preserving core context information, it significantly reduces the volume of context data, improves context processing efficiency, reduces system resource consumption, and provides efficient context support for the ReAct execution loop.

[0008] 3. Construct an autonomous error correction model to enable autonomous identification and correction of errors in the decision-making process. Breaking through the limitations of traditional decision-making levels lacking autonomous error correction capabilities, this paper constructs an autonomous error correction model based on error propagation and semantic matching, thereby enhancing the autonomous error correction capability of decision-making levels. This model pre-builds an error feature library and a correction rule library for the decision-making process, monitors the inference results and intermediate actions in the ReAct execution loop in real time, and performs semantic matching between the monitoring results and the error feature library to accurately identify the error type, location, and cause. Then, based on the correction rule library, it autonomously corrects the decision results and reselects the execution actions. This achieves closed-loop processing of decision-making errors without human intervention, solving the problem of low decision-making accuracy in traditional decision-making levels and improving the reliability of decisions.

[0009] Regarding question 2: Innovation in core transport layer capabilities to achieve unified routing and protocols across multiple platforms and improve cross-platform transmission efficiency. 4. Develop a multi-platform intelligent message routing algorithm to achieve accurate and efficient distribution of instructions. To address the issues of low efficiency and high error rate in traditional transport layer command distribution, a multi-platform intelligent message routing algorithm was developed, constructing a message routing logic of "platform attribute parsing - routing rule matching - optimal path selection - precise distribution". This algorithm pre-builds a multi-platform device identifier and routing rule library, which can parse the platform attributes of task commands issued by the decision-making layer, accurately match the target execution end with the corresponding routing rules, and dynamically select the optimal transmission path based on the real-time status of the transmission network (bandwidth, latency, packet loss rate). This achieves precise and efficient distribution of task commands, solving the drawbacks of traditional "broadcast" and "fixed path" distribution, and improving the efficiency and accuracy of command transmission.

[0010] 5. Establish a unified protocol smoothing mechanism to achieve seamless data interaction across platforms and protocols. To address the weak cross-platform and cross-protocol adaptability of traditional transport layers, a unified protocol smoothing mechanism is developed, constructing a protocol adaptation logic of "protocol type identification - format standardization conversion - rule adaptation mapping - seamless interaction". This mechanism pre-builds a format conversion rule and communication rule mapping library for mainstream communication protocols (TCP / IP, MQTT, HTTP, WebSocket, etc.). It can identify the protocol type of message data transmitted across platforms, standardize and unify the message format according to the conversion rules, and adapt and convert communication rules according to the mapping library. This completely solves the problems of format incompatibility and rule conflicts between different platforms and different protocols, achieving seamless data interaction and significantly improving the cross-platform adaptability of the transport layer.

[0011] Regarding question 3: Innovation of core capabilities at the execution layer to achieve multi-dimensional memory, secure isolated execution, and real-time feedback. 6: Integrating hybrid memory management with sandbox-based isolated execution to create secure, efficient, and feedback-enabled execution capabilities. To address the issues of traditional execution-layer memory management being simplistic, lacking execution security, and devoid of feedback mechanisms, a hybrid memory management model is integrated with sandbox-based isolated practical execution to create core execution capabilities at the execution layer: Hybrid memory management model: It integrates three dimensions: short-term memory, long-term memory and contextual memory. It classifies and stores execution data according to data type, usage frequency and scenario characteristics, and builds a multi-dimensional memory index library to realize fast data retrieval, dynamic updates and timeliness management. It solves the problems of single memory management and low retrieval efficiency in traditional memory management and improves the scenario adaptability of the execution layer. Sandbox isolated execution environment: Construct a virtualized sandbox environment that is completely isolated from the host system, configure minimal execution permissions and resources, and complete all practical tasks within the sandbox. Execution behavior is monitored in real time and unauthorized operations are prohibited. Temporary data is cleaned up after execution, which fundamentally ensures the security of the host system and the execution process, and solves the problem of traditional execution layers lacking security isolation and easily causing system damage. Real-time status feedback: The execution layer collects the status, progress and result data of task execution in real time, and after standardization by the transmission layer, it is fed back to the decision layer. This provides accurate feedback data for the decision layer's ReAct execution loop and autonomous error correction, forming an autonomous closed-loop execution logic of "decision-transmission-execution-feedback". This solves the problem of the lack of feedback mechanism in the traditional execution layer and realizes closed-loop control of the entire process.

[0012] Beneficial effects: The core advantage of this invention, a layered decoupled AI autonomous agent execution framework architecture system, lies in its native layered decoupling design and the synergistic linkage of core capabilities at each layer: the decision-making layer, transmission layer, and execution layer have clear functional boundaries and are decoupled and independent, supporting independent upgrades, expansions, and replacements of single layers, significantly improving the system's scalability; each layer interacts with data through standardized interfaces, creating an autonomous closed-loop execution logic of "decision-transmission-execution-feedback," realizing the full-process autonomous execution of the AI ​​autonomous agent; simultaneously, each layer is equipped with a status monitoring and anomaly handling module, which automatically triggers a fault-tolerance mechanism when an anomaly occurs at a certain layer, ensuring the stable operation of the entire system. This invention breaks through the design limitations of traditional integrated coupling in AI autonomous agents, significantly improving the decision-making autonomy, transmission adaptability, execution security, and scenario adaptability of AI autonomous agents, making it suitable for various complex AI agent practical application scenarios. Attached Figure Description Figure 1 Flowchart illustrating the working principle of the invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the system architecture and execution method of this invention will be described in detail below with reference to four specific embodiments. Each embodiment details the steps and qualitatively explains the principles of architectural innovation and algorithm efficiency enhancement, compares the optimization effect with existing technologies, and highlights the advantages of the layered decoupling architecture and the algorithm modeling innovation at each level, even without data support.

[0014] Example 1: Implementation of core capabilities of the decision-making layer's "brain" (ReAct loop + context compression + autonomous error correction) This embodiment addresses the three core aspects of the decision-making layer, solving the problems of traditional decision-making layers having simple execution logic, low context processing efficiency, and no autonomous error correction capability. It includes the core formula of the intelligent context compression algorithm, and the specific steps are as follows: External instruction reception and initial context loading The decision-making layer receives natural language instructions from external input (such as "Configure water pump operating parameters in the industrial PLC system, synchronize with the office system, and generate operating reports"), and simultaneously loads the massive amount of raw context data related to the instruction. This includes PLC system parameter templates, office system report formats, historical configuration records, equipment operation rules, etc., providing basic data support for decision-making and reasoning.

[0015] Context-aware intelligent compression processing generates lightweight feature vectors. The context compression unit is activated to perform lightweight processing of the original massive context data using a context-intelligent compression algorithm: ① To Perform semantic encoding Extract global semantic features from the context; ② Identify key features of the context (Such as PLC parameter configuration rules, core report fields), through attention mechanisms Calculate its weights and perform feature enhancement; ③ Identify and extract contextual redundancy features (Such as irrelevant historical running data, duplicate configuration instructions), through redundancy pruning operations. Cut; ④ Set the redundancy pruning weight coefficient Through the core formula Perform weighted feature fusion to generate a compressed, lightweight context feature vector. It enables lightweight processing of massive amounts of context.

[0016] The ReAct execution loop enables dynamic iterative decision-making based on "reasoning-action". Initiate the ReAct execution loop unit to lightweight the context feature vector. To support this, dynamic iterative decision-making is carried out: ① Initial reasoning: based on external instructions and The initial reasoning result was that "the PLC water pump parameters need to be configured first, then the data needs to be synchronized to the office system, and finally a report needs to be generated." ② Action selection: Based on the initial reasoning results, select "PLC parameter configuration" as the first action to execute; ③ Feedback reception: Receive action execution feedback from the execution layer through the transmission layer (e.g., "PLC parameter configuration interface is abnormal and cannot be configured directly"). ④ Iterative reasoning: Update the reasoning logic based on the feedback results, and correct it to "first repair the PLC parameter configuration interface, then configure the parameters, and the subsequent steps remain unchanged"; ⑤ Action Reselection: Based on the updated inference results, select "Repair PLC Configuration Interface" as the new action to execute; Repeat the process of "action selection - feedback reception - iterative reasoning" until the reasoning result can be implemented and the action is executed without deviation. Finally, standardized executable task instructions are generated, which include the execution sequence of "repairing PLC interface → configuring water pump parameters → synchronizing office system → generating operation report" and the execution constraints of each step.

[0017] Autonomous error correction and monitoring ensure the accuracy of decision-making results. The autonomous error correction unit is activated to perform error monitoring and autonomous correction throughout the entire ReAct execution loop: ① Pre-construct a database of decision-making error features (such as logical contradictions, mismatch between actions and instructions, and missing execution conditions) and a database of correction rules; ② Monitor the inference results and execution actions in real time. If it is detected that "the inference results require direct configuration of PLC parameters, but the interface status is not considered, and there is an error of missing execution conditions"; ③ Perform semantic matching between the error and the error feature library to accurately identify the error type as "missing execution condition"; ④ Based on the rule base, automatically add the pre-processing action of "detect and repair PLC interface" to complete the autonomous correction of the decision result without manual intervention.

[0018] Instruction standardization generates task instructions that can be recognized by the transport layer. The instruction standardization unit at the decision-making level transforms the final decision result into standardized task instructions, which include information such as execution end identifier, execution action sequence, execution constraints, and execution priority, providing standardized data for instruction distribution at the transport layer.

[0019] Principles of architecture and algorithm efficiency enhancement: Architectural innovation: The decision-making layer is set up as the core "brain" and decoupled from the transmission and execution layers. This focuses on the core capabilities of decision-making reasoning, avoiding the problem of decision-making functions being interfered with by other functions in the integrated coupling design, and improving the focus and efficiency of decision-making. At the same time, the three major capabilities of ReAct execution loop, context compression, and autonomous error correction are integrated into the decision-making layer to form an integrated decision-making system of "reasoning-processing-error correction", thus creating the autonomous decision-making capability of the decision-making layer.

[0020] Algorithm efficiency enhancements: Compared to the traditional full-load context processing method, the intelligent context compression algorithm significantly reduces the size of the context while retaining core information through semantic encoding, attention weighting, and redundancy pruning, thereby improving context processing efficiency and reducing system resource consumption; the ReAct execution loop mechanism realizes dynamic iteration of "reasoning-action," solving the problem of poor adaptability of traditional fixed decision logic to complex instructions; the autonomous error correction model realizes autonomous identification and correction of decision errors without human intervention, greatly improving the accuracy and autonomy of decision-making, and significantly reducing the decision deviation rate in complex scenarios compared to the traditional decision layer.

[0021] Example 2: Implementation of Core Capabilities of the Transport Layer "Neuron" (Multi-Platform Routing + Protocol Smoothing) This embodiment addresses the two core aspects of the transport layer, resolving the issues of weak cross-platform adaptability and lack of dedicated routing and protocol adaptation mechanisms in traditional transport layers. The specific steps are as follows: Standardized task instruction reception and platform attribute parsing The transport layer receives standardized task instructions from the decision-making layer through a standardized interface, activates the multi-platform message routing unit, parses the platform attributes of the task instructions, and extracts the core attributes of the target execution end: including two execution ends, namely the industrial PLC system (IoT device, using MQTT protocol) and the office computer (Windows system, using TCP / IP protocol), and clarifies that the instructions need to be distributed to two execution ends with different platforms and different protocols.

[0022] Intelligent message routing across multiple platforms, matching routing rules with the optimal path. Command routing is performed based on a multi-platform intelligent message routing algorithm. ① Call the pre-built multi-platform device identifier and routing rule library, and match the unique device identifier and corresponding routing rule of the PLC system and Windows office computer according to the parsed execution end attributes; ② Collect real-time status data of the transmission network, including bandwidth, latency, and packet loss rate of each transmission path, and select the optimal transmission path of "edge gateway-IoT private network" for the PLC system and the optimal transmission path of "LAN-fiber optic" for the Windows office computer. ③ Based on execution priority, the instruction sequence of "Repair PLC interface → Configure water pump parameters" is distributed to the PLC system, and the instruction sequence of "Synchronize data → Generate operation report" is distributed to the Windows office computer, so as to achieve accurate and efficient distribution of instructions.

[0023] Unified protocol smoothing, completing cross-protocol format and rule conversion. Initiate the protocol smoothing unit to complete cross-protocol adaptation and conversion based on the unified protocol smoothing mechanism: ① The protocol type of the instructions distributed to the PLC system is identified, and the target protocol is determined to be MQTT. The protocol type of the instructions distributed to the Windows office computer is identified, and the target protocol is determined to be TCP / IP. ② Call the pre-built protocol format conversion rule base to convert the standardized task instructions into the PUBLISH message format of the MQTT protocol and the data packet format of the TCP / IP protocol respectively, thus completing the standardized adaptation of message formats; ③ Call the communication rule mapping library to map the system's general communication rules to the "subscribe-publish" rules of the MQTT protocol and the "three-way handshake-data transmission" rules of the TCP / IP protocol, respectively, to complete the adaptation and conversion of communication rules; ④ Add a checksum and timestamp to the converted instruction data to ensure data integrity and timeliness.

[0024] Command transmission and status monitoring ensure transmission stability. The transport layer transmits task instructions to the corresponding execution end according to the matched optimal path and the converted protocol format. At the same time, the transmission status monitoring unit monitors the status of the transmission link and the progress of instruction transmission in real time. If anomalies such as link interruption or packet loss occur, the retransmission mechanism and backup path switching are immediately triggered to ensure the stability of instruction transmission. After the transmission is completed, the transmission log is recorded to provide a basis for subsequent anomaly investigation.

[0025] Execution status data reception and standardized feedback The transport layer receives execution status and result data (such as "PLC interface repair completed", "parameter configuration successful", "data synchronization completed", etc.) uploaded by the execution layer, identifies the protocol type and standardizes the format of the data, and feeds it back to the decision layer according to the reverse routing path, providing accurate feedback data for the decision layer's ReAct execution loop.

[0026] Principles of architecture and algorithm efficiency enhancement: Architectural innovation: The transport layer is set up as an independent "nerve" core, decoupled from the decision-making and execution layers, creating dedicated cross-platform transport and protocol adaptation capabilities. This changes the traditional design pattern of transport functions as auxiliary functions, realizing the specialization and efficiency of transport capabilities. At the same time, the transport layer adopts an integrated design of "routing-protocol-monitoring", forming a full-process transport system of "parsing-routing-conversion-transmission-feedback", ensuring the efficient and stable transmission of instructions and data.

[0027] Algorithm Enhancement: The multi-platform message intelligent routing algorithm achieves accurate instruction distribution through platform attribute parsing and optimal path selection, solving the problems of resource waste and high error rate of traditional "broadcast" distribution, and improving the efficiency and accuracy of instruction distribution; The protocol unification and smoothing mechanism achieves seamless data interaction across platforms and protocols through protocol type identification and format and rule conversion, completely solving the protocol compatibility problem of traditional transport layer and greatly improving the system's cross-platform adaptability; Compared with traditional transmission methods, the latency of cross-platform instruction transmission is significantly reduced and the transmission success rate is greatly improved.

[0028] Example 3: Implementation of Core Capabilities of the Execution Layer "Limbs" (Hybrid Memory Management + Sandbox Isolation Execution) This embodiment corresponds to the core of the execution layer, solving the problems of traditional execution layers such as single memory management, lack of secure isolation execution environment, and missing feedback mechanism. The specific steps are as follows: Standardized instruction receiving and execution environment initialization The execution layer receives standardized task instructions through the transport layer and completes environment initialization on both the PLC system and the Windows office computer: starting the practical execution unit and the status feedback unit, loading the hybrid memory management model, and building a sandbox isolated execution environment to prepare for task execution.

[0029] Hybrid memory management model loading enables multi-dimensional data memory support. The hybrid memory management unit is activated, integrating short-term memory, long-term memory, and episodic memory to provide data support for the execution process: ① Long-term memory area: Pre-stores frequently reused data such as parameter configuration rules of PLC system, report generation templates of office system, and general rules for equipment operation, and builds a long-term memory index to achieve fast retrieval; ② Short-term memory area: Stores temporary data for this execution, such as repair instructions for the PLC interface, target operating parameters of the water pump, and synchronization data from the office system, ensuring real-time data requirements during the execution process; ③ Context memory area: Stores the context features and historical execution records of the collaborative execution of the industrial PLC and office system, providing a basis for context adaptation in this execution; ④ During execution, multi-dimensional memory data is retrieved in real time according to execution requirements. After execution, the effective rules and scenario characteristics of this execution are updated to the long-term memory area, and temporary data is cleaned up from the short-term memory area to achieve dynamic management of memory data.

[0030] Sandbox-based execution environment construction to achieve secure isolated execution. Start the sandbox execution environment unit to build independent sandbox isolated execution environments for the two execution ends: ① PLC system side: Build a virtualized sandbox isolated from the PLC main system, configure it with minimal execution permissions that only include "interface repair and parameter configuration", and prohibit operations within the sandbox from accessing the core data and functions of the PLC main system; ② Windows Office PC Client: Build a sandbox environment isolated from the host system of the office PC, configure it with minimal execution permissions that only include "data synchronization and report generation", and prohibit operations within the sandbox from accessing the computer's local files and core programs; ③ After the sandbox environment is built, standardized task instructions are converted into executable operation sequences within the sandbox, ensuring that the operation sequences can only be executed within the sandbox.

[0031] The process was conducted in a sandbox environment, and execution status data was collected in real time. The execution layer's practical execution units drive the execution end to complete practical tasks within a sandbox isolation environment: ① PLC system side: Execute the "interface repair → parameter configuration" operation sequence in the sandbox, and complete the precise configuration of the water pump operating parameters according to the parameter configuration rules in the hybrid memory management model; ② Windows office computer: Execute the "Data Synchronization → Generate Run Report" operation sequence in the sandbox to synchronize the configured parameter data from the PLC system and generate standardized run reports according to the report template; ③ During execution, the status feedback unit collects execution status data in real time, including execution progress, operation results, and abnormal information, such as "interface repair successful (90%)", "parameter configuration completed (100%)", "report generated successfully", etc., to provide accurate data for subsequent feedback.

[0032] Execution result cleanup and status data feedback After the task is completed, the sandbox execution environment unit immediately cleans up all temporary data and operation records in the sandbox, releases the resources occupied by the sandbox, and ensures the cleanliness and security of the host system. At the same time, the status feedback unit standardizes and encapsulates the collected execution status and result data, and feeds it back to the decision layer through the transmission layer to complete the core execution process of the execution layer.

[0033] Principles of architecture and algorithm efficiency enhancement: Architectural Innovation: The execution layer is made an independent "limb" core, decoupled from the decision-making and transmission layers, focusing on the core capabilities of practical execution, and creating an integrated execution system of "memory-sandbox-execution-feedback". This changes the traditional design mode of the execution layer with single function and no security protection. At the same time, the execution layer adopts a two-layer design of "hybrid memory + sandbox isolation", which not only provides comprehensive memory support for the execution process, but also ensures the security of the execution process, achieving a dual improvement in execution efficiency and security.

[0034] Algorithm Enhancement: The hybrid memory management model integrates short-term, long-term, and contextual memory dimensions, enabling the categorized storage and dynamic management of execution data. This solves the problems of low retrieval efficiency and weak scenario adaptability in traditional single-memory management models, significantly improving the scenario adaptability and data retrieval efficiency of the execution layer. The sandbox-isolated execution environment, through complete isolation from the host system and minimal permission configuration, fundamentally avoids damage to the host system caused by execution errors or malicious operations, resolving the security risks of traditional execution layers and ensuring the security of the execution process. The real-time status feedback mechanism provides accurate feedback data for the dynamic iteration and autonomous error correction of the decision-making layer, realizing effective linkage between the execution layer and the decision-making layer. Compared with traditional execution layers, the accuracy and reliability of task execution are greatly improved.

[0035] Example 4: Implementation of Autonomous Closed-Loop Execution of the Entire System (Decision-Transmission-Execution-Feedback) This embodiment corresponds to the overall layered decoupled architecture and autonomous closed-loop execution logic of the system, demonstrating the collaborative linkage process of the decision-making layer, transmission layer, and execution layer, and verifying the system's full-process autonomous execution capability. The specific steps are as follows: System initialization, with each level decoupled and started independently. Based on a layered and decoupled architecture, the AI ​​autonomous intelligent agent execution framework architecture system is launched. The decision layer, transmission layer, and execution layer start independently and complete the environment initialization: each layer loads its own core algorithms and models, completes the connection of standardized interfaces, the transmission layer establishes communication links between each layer, the execution layer completes the hybrid memory management and sandbox environment initialization, the decision layer is in the instruction receiving ready state, and the system as a whole enters the execution state.

[0036] Command reception and autonomous reasoning at the decision-making level generate standardized task instructions. The decision layer receives complex cross-domain instructions from external inputs and, following the process in Example 1, completes context intelligent compression, ReAct execution loop dynamic decision-making and autonomous error correction to generate standardized task instructions containing execution end identifiers, execution sequences and execution constraints. These instructions are then sent to the transport layer through a standardized interface to complete the "decision-making" process.

[0037] Precise routing and protocol adaptation at the transport layer enable cross-platform distribution of commands. The transport layer receives standardized task instructions from the decision layer. Following the process in Example 2, it completes the parsing of the execution platform attributes, intelligent routing of messages across multiple platforms, and unified protocol smoothing. The instructions are then accurately distributed to the corresponding execution terminals. At the same time, the transport status is monitored in real time to ensure the stability of instruction transmission, thus completing the "transmission" stage.

[0038] Secure execution and status acquisition at the execution layer enable practical task implementation. The execution layer receives instructions distributed by the transport layer and, following the process in Example 3, completes hybrid memory management, sandbox isolation execution environment construction, and sandbox practical execution, collecting execution status and result data in real time to complete the "execution" stage.

[0039] State feedback and iterative optimization at the decision-making level form an autonomous closed loop. The execution layer collects execution status and result data, which is then standardized by the transmission layer and fed back to the decision layer. The decision layer then performs iterative optimization of the ReAct execution loop based on the feedback data: if all tasks are completed and the results meet the instruction requirements, the decision layer determines that the task execution is successful, generates the task completion result, and outputs it; if an exception occurs during the task execution and cannot be resolved at the execution layer, the decision layer performs autonomous error correction based on the feedback exception information, corrects the decision result, generates new standardized task instructions, and redistributes them to the execution layer through the transmission layer until the task is successfully executed, forming an autonomous closed-loop execution logic of "decision-transmission-execution-feedback-re-decision".

[0040] Task completion and system resource reclamation After the task is successfully executed, the system performs resource recycling at each level: the decision layer cleans up the context data and temporary decision results of this decision, the transmission layer cleans up the transmission logs and temporary routing data, the execution layer cleans up the sandbox environment and short-term memory data, and only updates the effective rules and scenario features to the long-term memory area. The system is then restored to the pending execution state to prepare for the next task execution.

[0041] Principles of architecture and algorithm efficiency enhancement: Architectural Innovation: The core architectural advantage of this invention is its native layered decoupling design, which resembles a "brain-nerve-limb". The decision-making layer, transmission layer, and execution layer have clear functional boundaries and are decoupled and independent, supporting independent upgrades, expansions, and replacements at each level, which greatly improves the system's scalability and maintainability. Each level achieves collaborative linkage through standardized interfaces, creating a fully autonomous closed-loop execution logic. This enables the AI ​​autonomous agent to perform autonomous execution from instruction reception to task completion, breaking the execution limitations of traditional integrated coupled architectures.

[0042] System Efficiency Enhancement: Compared to traditional integrated and coupled AI autonomous agent execution frameworks, this system achieves four core improvements through layered decoupling and core capability innovation at each level: ① Significantly enhanced autonomous decision-making capability at the decision-making layer, capable of handling complex cross-domain instructions, with significantly improved decision accuracy and iterative capabilities; ② Significantly enhanced cross-platform adaptability at the transmission layer, enabling seamless data interaction across multiple platforms and protocols, with significantly improved transmission efficiency and stability; ③ Significantly enhanced execution security and scenario adaptability at the execution layer, enabling secure isolated execution and multi-dimensional memory support, with significantly improved execution accuracy and reliability; ④ Significantly enhanced overall system scalability and maintainability, with each layer capable of independent upgrades without requiring overall system modification, reducing upgrade and maintenance costs. Simultaneously, the system's autonomous closed-loop execution logic achieves fully autonomous execution without human intervention, significantly enhancing the intelligence level and practical application capabilities of the AI ​​autonomous agent.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A hierarchical decoupling-based AI autonomous agent execution framework architecture system, characterized in that, The system adopts a layered, decoupled native architecture resembling "brain-nerve-limbs," comprising three core layers: a decision-making layer, a transmission layer, and an execution layer. Each layer has independent functional boundaries and collaborates through standardized interfaces. The system's execution methods include: a. The decision-making layer acts as the "brain," receiving external input instructions and completing dynamic iterative decision-making of "reasoning-action" based on the ReAct execution loop mechanism. It performs lightweight processing of massive context data through a context intelligent compression algorithm, and at the same time relies on an autonomous error correction model to realize error identification and autonomous correction of the decision-making process and intermediate results, generating standardized and executable task instructions. b. The transport layer acts as the "nerve," receiving task instructions from the decision-making layer. It achieves accurate distribution of instructions through multi-platform message intelligent routing algorithms and completes the standardized conversion of message formats and communication rules across platforms and protocols by relying on the protocol unification and smoothing mechanism, thereby achieving efficient and stable transmission of instructions and execution status data. c. The execution layer, acting as a "limb," receives standardized task instructions forwarded by the transport layer. Through a hybrid memory management model, it completes multi-dimensional storage, retrieval, and updating of various types of data during the execution process. It executes practical tasks in a sandbox-isolated execution environment and collects execution status and result data in real time. The data is then fed back to the decision layer via the transport layer, forming an autonomous closed-loop execution logic of "decision-transmission-execution-feedback". The modeling formula for the context-intelligent compression algorithm is: Ccomp = SencodeCraw⊙AattCkey + μ PpruneCred, where Ccomp is the compressed lightweight context feature vector, Sencode For semantic encoding operations, Craw represents the raw, massive contextual data, ⊙ represents the feature weighting and fusion operation, and Aatt represents the semantic encoding operation. This is the attention weight calculation operation, where Ckey is the core key feature of the context, μ is the redundancy pruning weight coefficient, and Pprune... This is a redundant feature pruning operation, where Cred represents the contextual redundant feature.

2. The system according to claim 1, characterized in that, The decision-making layer includes a ReAct execution loop unit, a context compression unit, an autonomous error correction unit, and an instruction standardization unit. The implementation logic of the ReAct execution loop unit is as follows: generate an initial inference result based on external input instructions → select an action to be executed based on the inference result → provide feedback on the action execution result → iteratively update the inference result → until the optimal executable task instruction is generated, thereby realizing a dynamic loop decision-making process of "inference-action".

3. The system according to claim 2, characterized in that, The context compression unit achieves lightweight processing of massive context data through a context intelligent compression algorithm. The steps include: semantically encoding the original context data to extract global features → calculating and strengthening the weights of core key features through an attention mechanism → identifying and pruning redundant context features → weighted fusion of encoded features, key features, and pruned features → generating a standardized lightweight context feature vector to provide efficient context support for the ReAct execution loop.

4. The system according to claim 2, characterized in that, The autonomous error correction model of the autonomous error correction unit is based on error propagation and semantic matching. The steps include: constructing a decision-making process error feature library and a correction rule library → real-time monitoring of the inference results and intermediate actions of the ReAct execution loop → semantic matching of the monitoring results with the error feature library → identifying error types and error locations → completing the autonomous correction of decision results and reselection of actions according to the correction rule library to ensure the accuracy of decision-making.

5. The system according to claim 1, characterized in that, The transport layer includes a multi-platform message routing unit, a protocol smoothing unit, a transmission status monitoring unit, and a data standardization unit. The implementation steps of the multi-platform message intelligent routing algorithm include: constructing a multi-platform device identifier and routing rule base → parsing the platform attributes of the task instructions issued by the decision layer → matching the corresponding routing rules with the target execution end → dynamically selecting the optimal transmission path according to the transmission network status → achieving accurate and efficient distribution of task instructions.

6. The system according to claim 5, characterized in that, The protocol smoothing unit's protocol unification smoothing mechanism includes: constructing a format conversion rule and communication rule mapping library for mainstream communication protocols (TCP / IP, MQTT, HTTP, WebSocket, etc.) → identifying the protocol type of message data transmitted across platforms → standardizing and unifying the message format according to the conversion rules → adapting and converting the communication rules according to the mapping library → achieving seamless data interaction between different platforms and different protocols.

7. The system according to claim 1, characterized in that, The execution layer includes a hybrid memory management unit, a sandbox execution environment unit, a practical execution unit, and a status feedback unit. The hybrid memory management model integrates short-term memory, long-term memory, and contextual memory. The implementation steps include: storing temporary data of real-time execution in the short-term memory area, storing frequently reused execution rules and templates in the long-term memory area, and storing the execution process and results of specific scenarios in the contextual memory area → constructing a multi-dimensional memory index library to achieve rapid retrieval of memory data → completing the dynamic updating, migration, and cleaning of memory data according to execution needs and memory timeliness, providing comprehensive memory support for practical execution.

8. The system according to claim 7, characterized in that, The sandbox isolated execution environment unit is an independent virtualized execution space. The implementation steps include: constructing a sandbox environment isolated from the host system and configuring minimal execution permissions and resources → converting standardized task instructions into executable operation sequences within the sandbox → driving the actual execution unit to complete task execution within the sandbox → monitoring the execution behavior within the sandbox in real time and prohibiting unauthorized operations and malicious behavior → cleaning up temporary data within the sandbox after execution to ensure the security of the host system and the execution process.

9. The system according to claim 1, characterized in that, The decision-making layer, transmission layer, and execution layer adopt a decoupled and independent deployment mode. Data interaction between each layer is achieved through standardized RESTful APIs and message queue interfaces. It supports independent upgrades, expansions, and replacements of single layers. At the same time, each layer is configured with a status monitoring and exception handling module. When an exception occurs in a certain layer, the fault tolerance mechanism can be automatically triggered to ensure the stable operation of the overall system.

10. An electronic device, characterized in that, The system includes a processor and a memory for storing processor-executable instructions, the processor being configured to invoke the instructions stored in the memory to implement the execution method of the AI ​​autonomous agent execution framework architecture system based on hierarchical decoupling as described in any one of claims 1-9, driving the system to complete autonomous closed-loop execution of "decision-transmission-execution-feedback".

Images