An Agent Construction Method and System Based on Agentic Workflow

Abstract

The present invention relates to the technical field of agent construction, and discloses an agent construction method and system based on an agentic workflow. The method includes the following steps: defining the core function modules of the agent and assigning unique identifiers, establishing a communication protocol standard between the modules, clarifying the message format and priority rules, allocating computing resources for each module and setting an elastic scaling policy, generating a module dependency relationship graph and an architecture metadata file, detecting the loss and delay of multi-modal input data, generating compensation features based on historical context, calculating the quality confidence of each modality and performing dynamic weighted fusion. The present invention proposes a three-level alignment architecture to achieve cross-modal semantic consistency and ensure that the generated content conforms to the logic of the real scenario; designs a two-layer decision-making mechanism to drive the dynamic optimization of the workflow, significantly enhancing the system agility under complex tasks; constructs a privacy-efficiency balanced federated architecture to support the generation of differentiated content while protecting the enterprise data sovereignty.

Description

Technical Field

[0001] The present invention relates to the field of agent construction, and more specifically, it relates to a method and system for constructing an agent based on an agentic workflow. Background Art

[0002] With the rapid development of AGI (Artificial General Intelligence) technology, multi-modal generation and cross-domain collaboration capabilities have become the core challenges in constructing enterprise-level agents.

[0003] Traditional static workflows are difficult to adapt to dynamic business requirements. Agent collaboration is rigid, resulting in response delays and inefficient resource utilization. At the same time, there are semantic fragmentation problems in cross-modal content generation. Due to the lack of a fine-grained alignment mechanism for different modal information, the generated content has chaotic logic and scene mismatches. Summary of the Invention

[0004] The present invention provides a method and system for constructing an agent based on an agentic workflow, which solves the technical problems of cross-modal semantic inconsistency and low efficiency of static workflows in related technologies.

[0005] The present invention provides a method for constructing an agent based on an agentic workflow, including the following steps:

[0006] S100, define the core function modules of the agent and assign unique identifiers, establish the communication protocol standard between modules, clarify the message format and priority rules, allocate computing resources for each module and set the elastic scaling policy, and generate a module dependency relationship graph and an architecture metadata file;

[0007] S200, detect the loss and delay of multi-modal input data, generate compensation features based on historical context, calculate the quality confidence of each modality and perform dynamic weighted fusion, filter low-confidence features according to a preset quality threshold, and record the time consumption and quality metrics;

[0008] S300, quantify the historical task contribution of each module, verify the credibility of nodes through a consensus protocol, construct a decision weight topology graph between modules, and dynamically adjust the resource quota according to the weight to balance the load;

[0009] S400, inject differential noise into the model parameters at the central node, send lightweight parameters to the edge nodes and fuse local domain knowledge, and generate multi-modal content using an encryption algorithm and bind a digital watermark;

[0010] S500, construct a three-dimensional evaluation model of quality degradation, recovery efficiency, and evolution cost, dynamically adjust the module decision weight according to the evaluation result, directionally search the model parameter space and perform incremental hot updates;

[0011] S600 injects a high-intensity load to test the system limit in an isolated environment, dynamically replaces module components to verify output consistency, checks whether all-link metrics comply with the service level agreement, and generates an encrypted audit report with an attached integrity check hash value.

[0012] Furthermore, in S100, the specific steps are as follows:

[0013] S110, Component instantiation and registration: Define a set of intelligent agent function modules, and assign a unique identifier and version number to each module;

[0014] S120, Communication protocol modeling: Define the communication message format between modules, and establish priority and time-to-live rules;

[0015] S130, Resource quota allocation: Allocate computing resources to each module and set an elastic scaling policy;

[0016] S140, Architecture metadata generation: Output a standardized architecture description file and generate a runtime dependency graph;

[0017] The calculation formula for architecture metadata generation is as follows:

[0018] ;

[0019] ;

[0020] Where represents the architecture metadata set, including modules, message types, and dependencies, represents the set of intelligent agent function modules, represents the message type definition, represents the directed graph of module dependencies, represents the set of nodes, represents the set of edges.

[0021] Furthermore, in S200, the specific steps are as follows:

[0022] S210, Spatiotemporal feature compensation: Detect sensor data loss and latency, and generate compensation features based on historical context;

[0023] S220, Credibility weighted fusion: Calculate the confidence weights of each modality feature and perform dynamic weighted feature fusion;

[0024] S230, Quality threshold filtering: Evaluate the credibility of the fused features and trigger the feature resampling mechanism;

[0025] S240, Governance feature output: Generate a standardized feature description and record the governance process metadata;

[0026] The calculation formula for the governance feature output is as follows:

[0027] ;

[0028] ;

[0029] where represents the governance feature set, including features, weights, quality scores, and processing times, represents the feature output after governance, represents the set of confidence weights for each modality, represents the set of quality scores for each modality, represents the total governance delay, and the constraint for the delay is one-third of the total SLA delay, represents the maximum delay specified by the service level agreement, represents the processing time of the modality,

[0030] Furthermore, in S300, it specifically includes the following steps:

[0031] S310, Agent Capability Evaluation: Calculate the Shapley contribution value of each agent module to quantify the effectiveness of the module in historical tasks;

[0032] S320, BFT Consensus Protocol: Execute Byzantine Fault Tolerant Consensus at the decision-making layer to verify the credibility and weight of nodes;

[0033] S330, Dynamic Decision Topology Generation: Construct a weighted decision relationship graph and allocate decision weights between modules;

[0034] The calculation formula for dynamic decision topology generation is as follows:

[0035] ;

[0036] where represents the directed graph of decision weights, represents the set of modules, represents the decision-dependent edge, represents the weight matrix, and the element is , represents from to the decision weight, represents the module , capability values, represents the indicator function, which is 1 when the condition is met and 0 otherwise, represents the consensus achievement flag;

[0037] S340, Load Balancing Optimization: Adjust resource quotas according to decision weights to prevent single-point overload.

[0038] Furthermore, in S400, it specifically includes the following steps:

[0039] S410, Central Privacy Parameter Processing: Inject differential noise into the Carrot AI base model to generate privacy-protected parameters against reverse attacks;

[0040] ;

[0041] where represents the central model parameters after injecting differential privacy noise, represents the original central model parameters, represents Gaussian noise with a mean of 0 and a variance of ; represents the standard deviation of differential privacy noise, represents the differential privacy budget parameter, represents the differential privacy failure probability, represents the parameter cascade noise operator;

[0042] S420, Federated Edge Adaptation Training: Inject domain knowledge at edge nodes and perform privacy-preserving transfer learning;

[0043] S430, Secure Content Generation: Perform multimodal generation using encrypted parameters and bind digital watermarks with access control;

[0044] S440, Cross-Domain Content Verification: Verify the compliance and privacy of the generated content and generate an auditable evidence chain.

[0045] Furthermore, in S430, the calculation formula for secure content generation is as follows:

[0046] ;

[0047] where represents the output of encrypted secure multimodal content, represents the encryption operation using the key ; represents the generation model function, represents the central parameters after differential privacy processing, represents the local adapter parameter increment, represents the governed feature output, represents the encryption key generated based on zero-knowledge proof, represents the zero-knowledge proof - public key generation algorithm, represents the actuator drive module instance.

[0048] Further, in S500, it specifically includes the following steps:

[0049] S510, Three-dimensional fault tolerance index analysis: Quantify quality degradation, recovery time limit, and evolution cost;

[0050] ;

[0051] where represents the three-dimensional fault tolerance index vector, represents the decision weight aggregation value of S330 aggregation value, represents the modal quality score, represents the maximum value of the quality score, represents the fault recovery time, represents the maximum recovery time limit specified by the service level indicator, represents the evolution cost, including in S420, represents the baseline evolution cost, represents the total number of modes;

[0052] S520, Adaptive adjustment strategy: Dynamically update the module weight according to the fault tolerance index;

[0053] S530, Incremental model evolution: Perform parameter space directional search and hot update;

[0054] S540, Compliance closed-loop verification: Check the legal and ethical compliance of the evolution process.

[0055] Further, the calculation formula of the adaptive adjustment strategy is as follows:

[0056] ;

[0057] where represents the decision weight, represents the learning rate, represents the current three-dimensional fault tolerance index vector, represents the enterprise preset target vector, represents the square of the two-norm.

[0058] Further, in S600, it specifically includes the following steps:

[0059] S610, Extreme pressure test: Inject chaos engineering events to simulate extreme loads and record the system degradation and self-healing performance;

[0060] S620, Thermal replacement consistency verification: Dynamically replace the agent module and verify the output consistency;

[0061] S630, SLA Compliance Verification: Verify that all-link metrics meet the enterprise service level agreement;

[0062] S640, Verification Audit Report Generation: Generate a verifiable encryption performance certificate;

[0063] The calculation formula for the verification audit report is as follows:

[0064] ;

[0065] Where represents the encryption performance verification certificate, represents the encryption operation using the report key, represents the report key derived from the S430 step, represents the SLA compliance verification result, represents the hot replacement consistency verification result, represents the SHA3 hash value of

[0066] The present invention also proposes an agent construction system based on an agentic workflow, which executes the steps in the foregoing agent construction method based on an agentic workflow, including:

[0067] Architecture Management Module: Define the core function modules of the agent and assign unique identifiers, establish communication protocol standards between modules, clarify message formats and priority rules, allocate computing resources for each module and set elastic scaling policies, generate module dependency relationship graphs and architecture metadata files;

[0068] Feature Governance Module: Detect the loss and delay of multi-modal input data, generate compensatory features based on historical context, calculate the quality confidence of each modality and perform dynamic weighted fusion, filter low-trust features according to preset quality thresholds, and record time consumption and quality metrics;

[0069] Decision Coordination Module: Quantify the historical task contribution of each module, verify the credibility of nodes through a consensus protocol, construct a decision weight topology graph between modules, and dynamically adjust resource quotas according to weights to balance the load;

[0070] Federated Generation Module: Inject differential noise into model parameters at the central node, send lightweight parameters to edge nodes and fuse local domain knowledge, and generate multi-modal content using encryption algorithms and bind digital watermarks;

[0071] Fault Tolerance and Evolution Module: Construct a three-dimensional evaluation model of quality degradation, recovery efficiency and evolution cost, dynamically adjust the decision weights of modules according to the evaluation results, directionally search the model parameter space and perform incremental hot updates;

[0072] Performance Verification Module: Inject a high-intensity load in an isolated environment to test the system's limits, dynamically replace module components to verify output consistency, check whether all-link metrics meet the service level agreement, and generate an encrypted audit report with an attached integrity check hash value.

[0073] The beneficial effects of the present invention are as follows:

[0074] The present invention proposes a three-level alignment architecture to achieve cross-modal semantic consistency, ensuring that the generated content conforms to the logic of the real scenario; designs a two-layer decision-making mechanism to drive the dynamic optimization of the workflow, significantly enhancing the system agility under complex tasks; constructs a privacy-efficiency balanced federated architecture to support differential content generation while protecting the enterprise data sovereignty. Description of the Drawings

[0075] Figure 1 is a flowchart of an agent construction method based on an agentic workflow proposed by the present invention;

[0076] Figure 2 is a structural block diagram of an agent construction system based on an agentic workflow proposed by the present invention.

[0077] In the figure: 101, Architecture Management Module; 102, Feature Governance Module; 103, Decision Coordination Module; 104, Federated Generation Module; 105, Fault Tolerance Evolution Module; 106, Performance Verification Module. Detailed Embodiments

[0078] Now, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the protection scope of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described relative to some examples can also be combined in other examples.

[0079] As Figure 1 shown, an agent construction method based on an agentic workflow includes the following steps:

[0080] S100, Agent Architecture Registration: Define the core function modules of the agent and assign unique identifiers, establish the communication protocol standard between modules, clarify the message format and priority rules, allocate computing resources for each module and set the elastic scaling policy, and generate a module dependency relationship graph and an architecture metadata file;

[0081] In an embodiment of the present invention, the specific steps are as follows:

[0082] S110, Component Instantiation and Registration: Define the set of intelligent agent function modules, and assign a unique identifier and version number to each module;

[0083] The expression of the set of intelligent agent function modules is as follows:

[0084] ;

[0085] where represents the set of intelligent agent function modules, represents the perception module instance with version number , represents the cross-modal alignment module instance, represents the dynamic decision-making module instance, represents the actuator drive module instance, represents the module version number;

[0086] S120, Communication Protocol Modeling: Define the communication message format between modules, and establish priority and survival time rules;

[0087] Its calculation formula is as follows:

[0088] ;

[0089] where represents the message type definition, including various attributes of the message, represents the message source module identifier (taking values from ), represents the set of target module identifiers, represents the message processing priority (positive integer), represents the maximum survival period, represents the system preset maximum survival time threshold, represents the set of positive integers;

[0090] S130, Resource Quota Allocation: Allocate computing resources to each module and set the elastic scaling policy;

[0091] Its calculation formula is as follows:

[0092] ;

[0093] where represents the resource quota triple of module , represents the CPU core quota of module , represents the memory quota, represents the network bandwidth quota, represents the system's total computing resource constraint, represents the total number of modules;

[0094] S140, Architecture Metadata Generation: Output a standardized architecture description file and generate a runtime dependency graph;

[0095] Its calculation formula is as follows:

[0096] ;

[0097] ;

[0098] Where represents the architecture metadata set, including modules, message types, and dependencies, represents the set of agent function modules, represents the message type definition, represents the directed graph of module dependencies, represents the set of nodes, represents the set of edges, represents dependency output;

[0099] S200, Multimodal Feature Governance: Detect the loss and delay of multimodal input data, generate compensation features based on historical context, calculate the quality confidence of each modality and perform dynamic weighted fusion, filter low-confidence features according to a preset quality threshold, and record the time consumption and quality metrics of the governance process;

[0100] In an embodiment of the present invention, the specific steps are as follows:

[0101] S210, Spatiotemporal Feature Compensation: Detect the loss and delay of sensor data and generate compensation features based on historical context;

[0102] Its calculation formula is as follows:

[0103] ;

[0104] ;

[0105] Where represents the compensation feature of the th modality (such as vision , speech ), represents a long short-term memory network with parameter for feature prediction, represents the historical feature window, represents the th modality's data loss rate, , represents the Hadamard (element-wise) product, represents the Gaussian noise injection term, represents the noise variance;

[0106] S220, credibility weighted fusion: Calculate the confidence weights of each modality feature and perform dynamic weighted feature fusion;

[0107] Its calculation formula is as follows:

[0108] ;

[0109] ;

[0110] where represents the confidence weight of the th modality, represents the th modality quality score, which is obtained by calculation, represents the temperature parameter, represents the th modality's original feature, represents the th modality's compensation feature, represents the feature concatenation operator, represents the fused multi-modal feature, represents the total number of modalities;

[0111] S230, quality threshold filtering: Evaluate the credibility of the fused features and trigger the feature resampling mechanism;

[0112] Its calculation formula is as follows:

[0113] ;

[0114] ;

[0115] where represents the th modality quality score, represents the quality threshold, represents the output of the processed feature, represents the feature resampling function represents the fused feature, represents the maximum number of retries, represents the indicator function, which is 1 if the condition holds and 0 otherwise;

[0116] S240, processed feature output: Generate a standardized feature description and record the metadata of the processing process;

[0117] Its calculation formula is as follows:

[0118] ;

[0119] ;

[0120] where represents the governance feature set, including features, weights, quality scores, and elapsed times, represents the feature output after governance, represents the set of confidence weights for each modality, represents the set of quality scores for each modality, represents the total governance latency, and the constraint on the latency is one-third of the total SLA latency, represents the maximum latency specified by the service level agreement, represents the elapsed time for modality processing, represents the total number of modalities;

[0121] S300, Collaborative decision-making power allocation: Quantify the historical task contribution of each module, verify the credibility of nodes through the consensus protocol, construct a decision weight topology graph among modules, and dynamically adjust resource quotas according to weights to balance the load, ensuring that there is no risk of single-point overload in the decision network;

[0122] In one embodiment of the present invention, the specific steps are as follows:

[0123] S310, Agent ability evaluation: Calculate the Shapley contribution value of each agent module to quantify the effectiveness of the module in historical tasks;

[0124] Its calculation formula is as follows:

[0125] ;

[0126] where the Shapley ability value of module , represents a subset of , represents the subset the number of elements in represents the total number of modules, represents the subset the task completion score of calculated and obtained from S240, represents the task completion score after adding , represents the factorial operation;

[0127] S320, BFT consensus protocol: Execute Byzantine fault-tolerant consensus at the decision-making layer to verify the credibility and weights of nodes;

[0128] Its calculation formula is as follows:

[0129] ;

[0130] ;

[0131] Where represents the consensus achievement flag (1 for achieved, 0 for not achieved), represents the Byzantine fault-tolerant consensus algorithm, represents the set of voting results of each node, represents the maximum number of fault-tolerant nodes, represents the module 's ability value, represents the ability threshold, represents the floor function symbol;

[0132] S330, Dynamic Decision Topology Generation: Construct a weighted decision relationship graph and allocate decision weights between modules;

[0133] Its calculation formula is as follows:

[0134] ;

[0135] Where represents the directed graph of decision weights, represents the set of modules, represents the decision dependence edge, represents the weight matrix, and the element is , represents from to 's decision weight, represents the module , 's ability value, represents the indicator function, which is 1 when the condition holds and 0 otherwise, represents the consensus achievement flag;

[0136] S340, Load Balancing Optimization: Adjust resource quotas according to decision weights to prevent single-point overload;

[0137] Its calculation formula is as follows:

[0138] ;

[0139] Where represents the CPU quota of the module , represents the current decision topology weight, represents the total number of modules, represents to The decision-making weight, denotes the sum of weights for all modules;

[0140] Load balancing rate:

[0141] ;

[0142] S400, Federal resilience generation: Inject differential noise into the model parameters at the central node to protect privacy, distribute lightweight parameters to the edge nodes and fuse local domain knowledge, generate multimodal content using encryption algorithms and bind digital watermarks, and verify the privacy compliance and quality baseline compliance of the generated results;

[0143] In an embodiment of the present invention, the specific steps are as follows:

[0144] S410, Central privacy parameter processing: Inject differential noise into the Carrot AI basic model to generate privacy-protected parameters resistant to reverse attacks;

[0145] Its calculation formula is as follows:

[0146] ;

[0147] Where denotes the central model parameters after injecting differential privacy noise, denotes the original central model parameters, denotes Gaussian noise with a mean of 0 and a variance of , denotes the standard deviation of differential privacy noise, denotes the differential privacy budget parameter (privacy strength), denotes the differential privacy failure probability, denotes the parameter cascade noise operator;

[0148] S420, Federal edge adaptation training: Inject domain knowledge at the edge nodes and perform privacy-protected transfer learning;

[0149] Its calculation formula is as follows:

[0150] ;

[0151] Where denotes the local adapter parameter increment, denotes the governed feature output, denotes the generation model function, denotes the central parameters after differential privacy processing, denotes the parameter fusion / adaptation operator, denotes the private data of the edge node, Denotes the second norm, Denotes the Frobenius norm, Denotes the regularization coefficient;

[0152] S430, Secure Content Generation: Perform multimodal generation using encryption parameters, bind digital watermarking and access control;

[0153] Its calculation formula is as follows:

[0154] ;

[0155] Where Denotes the output of encrypted secure multimodal content, Denotes the encryption operation using the secret key Such as AES-256-CTR, Denotes the generation model function, Denotes the central parameter after differential privacy processing, Denotes the local adapter parameter increment, Denotes the output of the processed features, Denotes the encryption key generated based on zero-knowledge proof, Denotes the zero-knowledge proof - public key generation algorithm, Denotes the actuator drive module instance;

[0156] S440, Cross-Domain Content Verification: Verify the compliance and privacy of the generated content, generate an auditable evidence chain;

[0157] Its calculation formula is as follows:

[0158] ;

[0159] Where Denotes the verification result of content compliance and privacy (1 for compliant, 0 for non-compliant), Denotes the content quality score after decryption, Denotes the minimum quality score threshold, Denotes the parameter leakage risk coefficient, Denotes the norm of the vector / parameter, Denotes the indicator function, using the constraint in S130 Training time-consuming, inheriting the in S330 to determine the cooperation priority of edge nodes;

[0160] Output verification:

[0161] Privacy protection: (Central parameter leakage rate ≤ 1%);

[0162] Generation aging: (2 / 3 of the total remaining time delay);

[0163] S500, Fault-tolerant evolution evaluation: Construct a three-dimensional evaluation model for quality degradation, recovery efficiency, and evolution cost. Dynamically adjust the module decision weights according to the evaluation results, directionally search the model parameter space, and perform incremental hot updates to verify the legal and ethical compliance of the evolution process;

[0164] In an embodiment of the present invention, the specific steps are as follows:

[0165] S510, Three-dimensional fault-tolerant index analysis: Quantify quality degradation, recovery aging, and evolution cost;

[0166] Its calculation formula is as follows:

[0167] ;

[0168] Where represents the three-dimensional fault-tolerant index vector, represents the decision weight of S330 aggregation value, represents the modal quality score, represents the maximum quality score, represents the fault recovery time, represents the maximum recovery time delay specified by the service level indicator, represents the evolution cost, including in S420, represents the baseline evolution cost, represents the total number of modes;

[0169] S520, Adaptive adjustment strategy: Dynamically update the module weights according to the fault-tolerant index;

[0170] Its calculation formula is as follows:

[0171] ;

[0172] Where represents the decision weight, represents the learning rate, represents the current three-dimensional fault-tolerant index vector, represents the enterprise preset target vector (such as [1, 0.5, 0.8]), represents the square of the two-norm;

[0173] S530, Incremental model evolution: Perform directional search and hot update of the parameter space;

[0174] Its calculation formula is as follows:

[0175] ;

[0176] ;

[0177] ;

[0178] where represents the new model parameters after evolution, represents the central parameters after differential privacy processing, represents the dimensions before taking the gradient direction, represents the gradient of the parameter , represents the mathematical expectation, represents the quality score of the secure content, represents the maximum parameter offset, represents the norm of the vector / parameter, represents the parameter update operator;

[0179] S540, Compliance Closed-Loop Verification: Check the legal and ethical compliance of the evolution process;

[0180] Its calculation formula is as follows:

[0181] ;

[0182] where represents the compliance verification result (1 for compliance, 0 for non-compliance), represents the total number of compliance indicators, represents the th compliance indicator value, represents the th compliance threshold, represents the parameter leakage risk coefficient, represents the fairness deviation caused by weight adjustment, represents the indicator function;

[0183] The input in S440 is used to calculate , and inherit the constraint on the weight update amplitude ;

[0184] Output Verification:

[0185] Quality Evolution Gain:

[0186] ;

[0187] represents the first dimension (quality) of the three-dimensional fault tolerance index vector;

[0188] Recovery efficiency constraint:

[0189] ;

[0190] where represents the second dimension (recovery efficiency) of the three-dimensional fault tolerance index vector;

[0191] Full compliance: ;

[0192] S600, Sandbox effectiveness verification: Inject a high-intensity load in an isolated environment to test the system limit, dynamically replace module components to verify output consistency, check whether all-link metrics meet the service level agreement, and generate an encrypted audit report with an attached integrity check hash value;

[0193] In one embodiment of the present invention, it specifically includes the following steps:

[0194] S610, Extreme pressure test: Inject chaos engineering events to simulate extreme loads and record system degradation and self-healing performance;

[0195] Its calculation formula is as follows:

[0196] ;

[0197] where represents the set of pressure test events, represents the class request intensity, represents the parameter Pareto distribution of, represents the test duration, represents the maximum delay of the service level agreement, represents the second dimension (recovery efficiency) of the three-dimensional fault tolerance index in S510;

[0198] S620, Hot replacement consistency verification: Dynamically replace the agent module and verify output consistency;

[0199] Its calculation formula is as follows:

[0200] ;

[0201] where represents the hot replacement consistency verification result (1 for consistent, 0 for inconsistent), represents the new output after hot replacement, represents the old output before hot replacement, represents the generation model function, represents the new model parameters after evolution, represents the governance feature set, Indicates data distribution offset noise ( ), Indicates the feature concatenation operator, Indicates the maximum allowable deviation threshold, Indicates the norm of the vector / parameter, Indicates the indicator function;

[0202] S630, SLA compliance verification: Verify that all-link metrics meet the enterprise service level agreement;

[0203] Its calculation formula is as follows:

[0204] ;

[0205] Where Indicates the SLA compliance verification result (1 for compliance, 0 for non-compliance), Indicates the total number of performance metrics, Indicates the decision weight matrix slice of S330, Indicates the th performance metric includes of S440, of S510, Indicates the th performance target value, Indicates the number of performance metrics, Indicates the indicator function;

[0206] S640, Verify audit report generation: Generate a verifiable encryption efficiency certificate;

[0207] Its calculation formula is as follows:

[0208] ;

[0209] Where Indicates the encryption efficiency verification certificate, Indicates the encryption operation using the report key, Indicates the report key derived based on S430, Indicates the SLA compliance verification result, Indicates the hot swap consistency verification result, Indicates 's SHA3 hash value (integrity check), input of S340 is used to calculate the stress test downgrade policy, inheriting the constraint test boundary conditions;

[0210] Output verification:

[0211] Hot swap success rate:

[0212] ;

[0213] wherein represents probability;

[0214] Full SLA compliance:

[0215] ;

[0216] Audit non-tamperability:

[0217] ;

[0218] wherein represents the Hamming distance to verify the non-tamperability of the audit report.

[0219] The mapping relationship of the agentic workflow is shown in Table 1 as follows:

[0220] Table 1: Mapping Relationship of Agentic Workflow

[0221]

[0222] As Figure 2 shown, based on the above-mentioned method for constructing an agent based on the agentic workflow, an agent construction system based on the agentic workflow is given, including the following modules:

[0223] Architecture Management Module 101: Define the functional boundaries and interaction protocols of the core components of the agent, assign unique identifiers and version numbers to modules such as perception, alignment, decision-making, and execution, formulate message formats, priority rules, and lifecycle policies for inter-module communication, establish a resource quota allocation mechanism and an elastic scaling policy, and generate a module dependency relationship graph and an architecture metadata description file;

[0224] Feature Governance Module 102: Real-time monitor the integrity and timeliness of multi-modal input data, perform context-aware compensation for missing or delayed data, calculate the quality confidence weights of each modal feature, perform dynamic weighted fusion and filter low-trust data, record the latency and quality metrics of the feature governance process, and provide a standardized feature set for downstream tasks;

[0225] Decision Coordination Module 103: Evaluate the historical task contribution and real-time load status of each functional module, verify the credibility of nodes through a distributed consensus protocol, construct a dynamic decision weight topology network between modules, adjust the resource quota allocation policy according to environmental changes, and maintain the load balance and fault tolerance of the decision network;

[0226] Federated Generation Module 104: Inject privacy protection noise into the basic model parameters at the central node, send lightweight parameters to the edge nodes and integrate domain knowledge, generate multi-modal content through encryption algorithms and embed digital watermarks, verify the compliance of the generation results with privacy compliance requirements, and achieve a balance between data security and generation efficiency;

[0227] Fault Tolerant Evolution Module 105: Construct a multi-dimensional evaluation model to quantify the quality degradation, recovery time and evolution cost of the system, dynamically optimize the decision weights and resource allocation of the module based on the evaluation results, directionally search the parameter space and execute incremental hot updates of the model, and continuously verify the compliance and fairness of the evolution process;

[0228] Performance Verification Module 106: Inject high-concurrency loads in an isolated sandbox environment to test the extreme performance of the system, dynamically replace functional modules to verify service continuity, check whether the full-link metrics meet the predefined service level agreement, generate encrypted audit reports and attach integrity verification information to ensure that the system meets the production environment deployment standards.

[0229] Based on the above construction method and system, the following process example is proposed: Enterprises need to deploy AG1 application scenarios such as intelligent customer service and digital human live broadcast;

[0230] Input Information:

[0231] Multi-modal Data: User consultation text, product images, voice instructions;

[0232] Business Goals: Response speed < 2 seconds, content accuracy > 95%;

[0233] Network Conditions: Bandwidth fluctuation of edge nodes (packet loss rate ≤ 15%);

[0234] The execution unit includes: CarrotAl industry large model, Agentic dynamic scheduling engine, federated security generation node;

[0235] Execution steps are as follows:

[0236] Semantic Alignment and Feature Fusion:

[0237] Extract text keywords and image region associations, and construct a cross-modal mapping dictionary;

[0238] Dynamically analyze the multi-modal data dependency relationship and adjust the feature fusion weight;

[0239] Correct the generated semantics based on the enterprise knowledge base to ensure compliance with industry terminology specifications;

[0240] Dynamic Task Collaboration:

[0241] Decompose the user request into atomic tasks that can be executed in parallel (such as intent recognition, image parsing);

[0242] Monitor the loads of each module in real time and dynamically allocate the priorities of computing resources;

[0243] Select the optimal execution path through an autonomous negotiation mechanism and support mid-task policy adjustment;

[0244] Privacy protection generation:

[0245] Add noise protection to the model at the central node to generate encryption parameters resistant to reverse inference;

[0246] Edge nodes fine-tune the model based on local data to generate customized content with invisible watermarks;

[0247] Encrypt and transmit the generation results across the entire link and detect potential privacy leakage risks in real time;

[0248] Execution results:

[0249] Aligned content: A customer service response solution with synchronized graphics / voice and pictures / video;

[0250] Optimization strategy: Dynamic resource allocation topology diagram (including fault rollback path);

[0251] Secure deliverable: A multi-modal content package with enterprise digital watermarks;

[0252] Verification effect:

[0253] Accuracy rate of advertisement graphic matching: 89% → 94%;

[0254] Response speed for emergency tasks: 4.1 seconds → 1.7 seconds;

[0255] Pass rate of privacy compliance audit: 100%;

[0256] This process has been applied to an e-commerce customer service system, supporting 7×24-hour multilingual digital human live broadcasts, with a 67% reduction in labor costs.

[0257] The above describes the embodiments of the present invention. However, the present invention is not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms, all of which fall within the protection scope of the present invention.

Claims

1. An agent construction method based on an agentic workflow, characterized in that, The following steps are involved: S100, define the core functional modules of the intelligent agent and assign unique identifiers, establish inter-module communication protocol standards, clarify message formats and priority rules, allocate computing resources to each module and set elastic expansion and contraction strategies, and generate module dependency graphs and architecture metadata files; S200, detects loss and delay of multimodal input data, generates compensation features based on historical context, calculates quality confidence of each modality and performs dynamic weighted fusion, filters low-confidence features according to preset quality thresholds, and records time consumption and quality indicators; S300, quantifies the historical task contribution of each module, verifies the node credibility through the consensus protocol, builds a decision weight topology diagram between modules, and dynamically adjusts resource quotas according to the weights to balance the load; S400 injects differential noise into the model parameters at the central node, sends lightweight parameters to the edge nodes and integrates local domain knowledge, uses encryption algorithms to generate multimodal content and binds digital watermarks; S500 builds a three-dimensional evaluation model of quality degradation, recovery efficiency and evolution cost, dynamically adjusts module decision weights according to the evaluation results, searches the model parameter space in a targeted manner and performs incremental hot updates; S600 injects high-intensity loads in an isolated environment to test system limits, dynamically replaces module components to verify output consistency, verifies whether full-link indicators comply with service level agreements, and generates encrypted audit reports with attached integrity verification hash values.

2. The method for constructing an agent based on an agentic workflow according to claim 1, characterized in that, In S100, the specific steps are as follows: S110, component instantiation registration: define a set of intelligent agent function modules and assign a unique identifier and version number to each module; S120, communication protocol modeling: define the communication message format between modules and establish priority and lifetime rules; S130, resource quota allocation: allocate computing resources to each module and set elastic expansion and contraction strategies; S140, architecture metadata generation: output standardized architecture description files and generate runtime dependency graphs; The calculation formula for schema metadata generation is as follows: ； ； Among them represents a set of architecture metadata, including modules, message types, and dependencies, represents a set of agent function modules, represents a message type definition, represents a directed graph of module dependencies, represents a set of nodes, represents a set of edges.

3. The method for constructing an agent based on an agentic workflow according to claim 2, wherein, In S200, the specific steps are as follows: S210, spatiotemporal feature compensation: detect sensor data loss and delay, and generate compensation features based on historical context; S220, credibility weighted fusion: calculating the confidence weight of each modality feature and performing dynamic weighted feature fusion; S230, quality threshold filtering: evaluate the credibility of fusion features and trigger feature resampling mechanism; S240, Governance feature output: Generate standardized feature descriptions and record governance process metadata; The calculation formula for the governance feature output is as follows: ； ； Among them represents the governance feature set, including features, weights, quality scores, and elapsed time represents the feature output after governance represents the set of confidence weights for each modality represents the set of quality scores for each modality represents the total governance delay, and the constraint of the delay is one-third of the total SLA delay represents the maximum delay specified by the service level agreement represents the elapsed time for modality processing represents the total number of modalities 4. The method for constructing an agent based on an agentic workflow according to claim 3, wherein, In S300, the following steps are specifically included: S310, Agent Capability Evaluation: Calculate the Shapley contribution value of each agent module and quantify the effectiveness of the module in historical tasks; S320, BFT consensus protocol: implements Byzantine fault-tolerant consensus at the decision-making level to verify node credibility and weight; S330, dynamic decision topology generation: construct a weighted decision relationship graph and assign decision weights between modules; The calculation formula for dynamic decision topology generation is as follows: ； Among them represents a decision weight directed graph, represents a set of modules, represents a decision dependence edge, represents a weight matrix, and the elements are , represents the decision weight from to , represents the ability values of modules 、 , represents an indicator function, which is 1 when the condition is satisfied and 0 otherwise, represents a consensus reached flag; S340, load balancing optimization: adjust resource quotas based on decision weights to prevent single point overload.

5. The agent construction method based on an agentic workflow according to claim 4, wherein In S400, the following steps are specifically included: S410, Central Privacy Parameter Processing: Inject differential noise into the Carrot AI base model to generate privacy-preserving parameters resistant to reverse attacks; ； Among them represents the central model parameters after injecting differential privacy noise, represents the original parameters of the central model, represents Gaussian noise with a mean of 0 and a variance of ; represents the standard deviation of differential privacy noise, represents the differential privacy budget parameter, represents the differential privacy failure probability, represents the parameter cascade noise operator; S420, Federated Edge Adaptation Training: Inject domain knowledge at edge nodes and perform privacy-preserving transfer learning; S430, Secure Content Generation: Perform multimodal generation using encrypted parameters, binding digital watermarks and access control; S440, Cross-Domain Content Verification: Verify the compliance and privacy of the generated content and generate an auditable evidence chain.

6. The method for constructing an agent based on an agentic workflow according to claim 5, wherein In S430, the calculation formula for secure content generation is as follows: ； Among them represents the encrypted secure multi-modal content output represents the encryption operation using the secret key represents the generation model function represents the central parameter after differential privacy processing represents the local adapter parameter increment represents the processed feature output represents the encryption key generated based on zero-knowledge proof represents the zero-knowledge proof - public key generation algorithm represents the actuator drive module instance represents the parameter fusion / adaptation operator​ 7. An agent construction method based on an agentic workflow according to claim 6, wherein, In S500, it specifically includes the following steps: S510, Three-Dimensional Fault Tolerance Index Analysis: Quantify quality degradation, recovery time, and evolution cost; ； wherein represents a three-dimensional fault tolerance index vector represents the decision-making weight of S330 aggregation value represents the modal quality score represents the maximum value of the quality score represents the fault recovery time represents the maximum recovery delay specified by the service level indicator represents the evolution cost represents the baseline evolution cost represents the total number of modes S520, Adaptive Adjustment Strategy: Dynamically update module weights according to fault tolerance indicators; S530, Incremental Model Evolution: Perform directional search in the parameter space and hot update; S540, Compliance Closed-Loop Verification: Check the legal and ethical compliance of the evolution process.

8. The method for constructing an agent based on an agentic workflow according to claim 7, wherein, The calculation formula for the adaptive adjustment strategy is as follows: ； wherein represents the decision-making weight, represents the learning rate, represents the current three-dimensional fault tolerance index vector, represents the enterprise preset target vector, represents the square of the two-norm.

9. A method for constructing an agent based on an agentic workflow according to claim 8, characterized in that, In S600, it specifically includes the following steps: S610, Extreme Stress Testing: Inject chaos engineering events to simulate extreme loads and record system degradation and self-healing performance; S620, Hot Replacement Consistency Verification: Dynamically replace agent modules and verify output consistency; S630, SLA Compliance Verification: Verify that all-link metrics meet the enterprise service level agreement; S640, Verification Audit Report Generation: Generate a verifiable encrypted performance certificate; The calculation formula for the verification audit report is as follows: ； Among them represents the encryption performance verification certificate represents the encryption operation using the reporting key represents the reporting key derived based on the steps in S430 represents the SLA compliance verification result represents the hot replacement consistency verification result represents the SHA3 hash value of represents the encrypted secure multi-modal content output 10. An agent construction system based on an agentic workflow, characterized in that, Execute the steps in an agent construction method based on an agentic workflow as described in any one of claims 1-9, including: Architecture Management Module: Define the core functional modules of the agent, assign unique identifiers, establish communication protocol standards between modules, clarify message formats and priority rules, allocate computing resources to each module and set elastic scaling policies, generate a module dependency relationship graph and an architecture metadata file; Feature Governance Module: Detect the loss and delay of multimodal input data, generate compensation features based on historical context, calculate the quality confidence of each modality and perform dynamic weighted fusion, filter low-trust features according to a preset quality threshold, and record time consumption and quality metrics; Decision Coordination Module: Quantify the historical task contribution of each module, verify the credibility of nodes through a consensus protocol, construct a decision weight topology graph between modules, and dynamically adjust resource quotas according to the weights to balance the load; Federated Generation Module: Inject differential noise into model parameters at the central node, send lightweight parameters to edge nodes and fuse local domain knowledge, and use encryption algorithms to generate multimodal content and bind digital watermarks; Fault Tolerance Evolution Module: Construct a three-dimensional evaluation model of quality degradation, recovery efficiency, and evolution cost, dynamically adjust module decision weights according to the evaluation results, perform directional search in the model parameter space and execute incremental hot updates; Performance Verification Module: Inject high-intensity loads in an isolated environment to test the system limit, dynamically replace module components to verify output consistency, verify whether all-link metrics comply with the service level agreement, and generate an encrypted audit report with an attached integrity check hash value.

Images