A Multi-Agent Collaborative Business Requirements Generation Method Based on Task Acceptance Criteria

By constructing a weighted undirected graph and a dual-agent collaborative mechanism, the problem of automatically recovering the acceptance criteria from user stories in DDD was solved, achieving efficient and accurate semantic aggregation and generation, and reducing manual costs and error rates.

CN122816588BActive Publication Date: 2026-11-17ZHEJIANG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611331163.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17
Estimated Expiration
2046-08-31

AI Technical Summary

Technical Problem

In Domain-Driven Design (DDD), existing technologies struggle to automate and restore user stories with high quality from fragmented acceptance criteria. This is especially true in large-scale scenarios where computational complexity is high and accuracy is difficult to guarantee, making it impossible to effectively handle semantic associations and business intent recovery across modules.

Method used

A large language model is used to evaluate the semantic relevance between acceptance criteria, a weighted undirected graph is constructed, and graph partitioning is solved by integer linear programming. User stories are generated by combining a dual-agent collaborative mechanism to ensure the accuracy and scalability of semantic aggregation. The generation quality is optimized through human-computer interaction.

Benefits of technology

It achieves efficient and accurate recovery from acceptance criteria to user stories, ensuring that the generated user stories are highly semantically relevant and can maintain computational efficiency in large-scale scenarios, reducing manual costs and error rates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122816588B_ABST
    Figure CN122816588B_ABST
Patent Text Reader

Abstract

This invention discloses a multi-agent collaborative business requirement generation method based on task acceptance criterion aggregation, belonging to the interdisciplinary field of software engineering and artificial intelligence. The core features of this method are: (1) semantic perception association construction - using a large language model to evaluate the deep semantic correlation between acceptance criteria based on business semantic elements; (2) scale-adaptive optimal solution - for small-scale problems with a small number of acceptance criteria, the global optimal solution is obtained by solving; for large-scale problems with a large number of acceptance criteria, the approximate optimal solution is obtained by solving; (3) consensus-driven quality assurance - through a dual-agent collaborative mechanism, automatic output is only generated when consensus is reached and the quality meets the standard; otherwise, human-computer interaction is triggered to obtain user feedback, forming a quality closed loop. This invention achieves efficient and accurate grouping and quality assurance of acceptance criteria through semantic aggregation, adaptive solution and dual-agent consensus mechanism.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of software engineering and artificial intelligence, and in particular relates to a method for generating multi-agent collaborative business requirements based on task acceptance criteria. Background Technology

[0002] In agile software development practices based on Domain-Driven Design (DDD), user stories typically describe business requirements using the standard format of "As a <role>, I hope for <goals> so as to achieve <value>". Each user story is also associated with several acceptance criteria written in the Given / When / Then format. As a core bridge connecting business intent and implementation details, user stories play an irreplaceable role in long-term software iterative maintenance.

[0003] However, as projects evolve, original user stories are often broken down into numerous fine-grained acceptance criteria, scattered across test case libraries, defect tracking systems, and automated test scripts. Maintenance personnel are faced with a massive amount of isolated Given / When / Then descriptions, requiring them to reverse-engineer the original user stories to which these criteria belong in order to understand the overall business objectives of the functional modules. Currently, this restoration process relies entirely on manual intervention. Requirements analysts must read each acceptance criterion individually, relying on personal experience to determine which criteria belong to the same user story, and manually infer their roles, goals, and value. This approach is not only labor-intensive but also imposes a heavy mental burden—maintainers must reconstruct the business context from fragmented descriptions, identify state transitions, rule dependencies, and event causality, and different personnel have different judgment standards, making consistency difficult to guarantee. Some solutions use predefined templates to directly convert individual acceptance criteria into user story format or simply concatenate criteria from the same module. However, these methods only address format conversion and do not involve semantic aggregation, failing to restore the business intent of the original user stories or handle cross-module criterion relationships.

[0004] The common drawback of the aforementioned existing technologies lies in their failure to solve the core challenge of "recovering the whole from fragments." Specifically, this task faces three fundamental difficulties: First, the value dimension of user stories cannot be directly derived from acceptance criteria and must be inferred from the business context. However, while Large Language Models (LLMs) may provide reasonable results in general domain scenarios, they are prone to errors in specific domain scenarios such as insurance claims and medical compliance. Existing methods either rely entirely on the automatic output of LLMs, leading to high error rates, or require manual completion of templates, resulting in excessive costs. There is a lack of a minimal interaction mechanism for processing based on confidence levels. Second, the quality of aggregation directly determines the upper limit of subsequent inferences. If the semantic association criteria are not aggregated into the same group, LLM will generate roles based on incomplete context. If the target or value is misaligned, and semantically irrelevant criteria are incorrectly aggregated, LLM will inevitably produce chaotic inferences in contradictory contexts. However, existing clustering algorithms only group based on text similarity and cannot capture the deep business semantic relationships between acceptance criteria, such as state transitions, rule dependencies, and event causality. Third, aggregating N acceptance criteria into a user story can be mathematically modeled as a graph partitioning problem with capacity constraints. When N is small (N≤200), the global optimal solution can be obtained through integer linear programming. However, as N increases, the solution time increases exponentially. Existing heuristic algorithms do not utilize the specific structure of the graph in this scenario, such as sparsity and semantic locality, resulting in a significant decrease in recovery quality in large-scale scenarios. Therefore, how to achieve automated and high-quality recovery from acceptance criteria to user stories, while taking into account inference accuracy, aggregation rationality, and computational scalability, remains a pressing technical challenge in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation, which solves the technical problem of difficulty in recovering the original business intent from fragmented acceptance criteria after user stories are lost in software maintenance. This method is particularly suitable for the organization and management of large-scale acceptance criteria in Domain-Driven Design (DDD) projects.

[0006] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for generating multi-agent collaborative business requirements based on task acceptance criteria, comprising the following steps:

[0008] S1. The set of input acceptance criteria is transformed into a weighted undirected graph using a large language model-related discrimination method;

[0009] S2. Transform the weighted undirected graph into an integer linear programming problem. With the goal of maximizing the sum of the weights of the undirected edges in the weighted undirected graph, the optimal graph partitioning scheme is obtained by solving the integer linear programming problem. This graph partitioning scheme aggregates the acceptance criteria into several connected components, and each connected component corresponds to a subset of the acceptance criteria associated with a user story.

[0010] S3. A dual-agent collaborative mechanism is adopted to generate user stories for each connected component. The generating agent generates user stories in a standard format according to the acceptance criteria subset of each connected component, and the reviewing agent independently evaluates the generation quality of the user stories. If the generated user stories simultaneously meet the confidence difference condition and the confidence consensus condition, they are directly output; otherwise, human-computer interaction is triggered to obtain user feedback and use it as long-term memory context to guide the generating agent to output user stories that meet the conditions.

[0011] S4. The large language model extracts three key concepts from user stories and acceptance criteria: application service information, domain entity information, and business rule information, forming multiple candidate contexts. For each candidate context, the large language model evaluates its relevance to the user story from multiple dimensions, calculates a weighted sum of the evaluation scores of each dimension, and selects the candidate contexts that meet the requirements of the comprehensive score as domain contexts that have substantial supplementary value to the user story. After integration with the user story, a complete business requirement document is output.

[0012] Based on the above scheme, each step can be implemented in the following preferred manner.

[0013] As a preferred embodiment of the first aspect mentioned above, the specific process of generating a weighted undirected graph using a large language model-related discrimination method is as follows:

[0014] S11. Extract the business semantic elements from each acceptance criterion using the large language model, traverse all acceptance criteria, select a unique acceptance criterion as the target acceptance criterion during each traversal, and use all other acceptance criteria as candidate acceptance criteria. The large language model evaluates the semantic relevance between the target acceptance criterion and each candidate acceptance criterion based on the business semantic elements, and outputs the confidence that these two acceptance criteria belong to the same user story.

[0015] S12. Treat the acceptance criteria as nodes in a weighted undirected graph, and establish an undirected edge between two acceptance criteria with a bidirectional confidence score greater than zero. Use the bidirectional confidence score as the weight of the undirected edge to form a weighted undirected graph. The bidirectional confidence score is the sum of the two confidence scores obtained when the two acceptance criteria are used as the target acceptance criterion and the candidate acceptance criterion, respectively.

[0016] As a preferred option for the first aspect mentioned above, the business semantic elements extracted from the acceptance criteria include domain concept elements, scenario type elements, business rule elements, and domain event elements.

[0017] As a preferred option for the first aspect mentioned above, the large language model comprehensively considers four dimensions when evaluating semantic relevance: domain concept overlap, which indicates whether the domain entities and value objects involved in the acceptance criteria overlap; scenario type consistency, which indicates whether the acceptance criteria belong to the same scenario type; business rule dependency, which indicates whether the acceptance criteria have a dependency chain between business rule elements and the acceptance criteria; and event causal relationship, which indicates whether the acceptance criteria have a domain event and its trigger-response chain.

[0018] As a preferred embodiment of the first aspect, the integer linear programming problem includes three constraints: node exclusivity constraint, upper limit constraint on the size of connected components, and constraint on the retention of undirected edges.

[0019] As a preferred embodiment of the first aspect above, the node exclusivity constraint is that each acceptance criterion belongs to only one connected component; the upper limit constraint of the connected component size is that the number of acceptance criteria contained in each connected component cannot exceed the upper limit of the number of acceptance criteria contained in each user story; and the undirected edge retention condition constraint is that an undirected edge can only be retained in the connected component when the two nodes of an undirected edge belong to the same connected component.

[0020] As a preferred option of the first aspect mentioned above, the several connected components output by S2, the subset of acceptance criteria corresponding to each connected component, and the text content of each acceptance criterion are organized into prompt words, which are input into the generating agent. The generating agent generates user stories through multi-step reasoning and outputs self-rated confidence.

[0021] As a preferred option for the first aspect mentioned above, the user stories generated by the generating agent are input into the reviewing agent, which evaluates the quality of the generated user stories in multiple dimensions and outputs the review confidence level.

[0022] As a preferred embodiment of the first aspect mentioned above, the confidence level difference condition is that the difference between the self-assessed confidence level and the review confidence level is less than a preset difference threshold; the confidence level consensus condition is that the average confidence level is greater than or equal to a preset quality threshold, and the average confidence level is the average of the self-assessed confidence level and the review confidence level.

[0023] In a second aspect, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can realize the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation as described in any of the solutions in the first aspect above.

[0024] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation as described in any of the solutions of the first aspect above.

[0025] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0026] The memory is used to store computer programs;

[0027] The processor is configured to, when executing the computer program, implement the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation as described in any of the solutions of the first aspect above.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] First, precise semantic-aware aggregation. By using a large language model to evaluate the deep semantic relationships between acceptance criteria based on business semantic elements, and aiming to maximize the weight of undirected edges, the aggregation ensures that acceptance criteria within the same user story are highly related in terms of business semantics, avoiding the erroneous aggregation based on text similarity in traditional methods.

[0030] Second, efficient solution with adaptive scaling. Different solution strategies are adopted based on the number of acceptance criteria, maximizing aggregation quality while ensuring computational scalability, thus resolving the technical contradiction between the lack of scalability of exact methods and the lack of quality guarantee for heuristic methods.

[0031] Third, consensus-driven quality assurance. Through a dual-agent collaborative mechanism, output is automatically generated only when both parties reach a consensus and the quality meets the standards; otherwise, human-computer interaction is triggered to obtain verification information. This ensures the quality of the generated data while keeping human intervention to a minimum. Attached Figure Description

[0032] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0033] Figure 2 A flowchart illustrating the method for generating a weighted undirected graph according to the present invention;

[0034] Figure 3 This is a schematic diagram illustrating the three types of constraints involved in the integer linear programming problem in the method of this invention;

[0035] Figure 4 This is a flowchart illustrating the generation of user stories through a dual-agent collaborative mechanism in the method of this invention;

[0036] Figure 5This is a schematic diagram of a computer electronic device provided by the present invention. Detailed Implementation

[0037] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0038] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0039] The concept of this invention is: given a set of acceptance criteria (which can be generated manually or by automated methods), how to automatically generate user stories with complete structures—each user story contains a clear role, an understandable goal, quantifiable business value, and a semantically related set of acceptance criteria, with controllable granularity (not exceeding the upper limit) and guaranteed generation quality.

[0040] Therefore, this invention provides a method for generating multi-agent collaborative business requirements based on task acceptance criteria aggregation, the core idea of ​​which consists of two stages:

[0041] Phase 1: Semantic Aggregation. The acceptance criterion aggregation problem is modeled as a constrained integer linear programming problem. Using the acceptance criteria as graph nodes and the semantic relevance evaluated by the large language model as edge weights, the aggregation scheme that maximizes the intra-class edge weights is solved under constraints. This phase ensures both semantic relevance and controllable granularity.

[0042] Phase Two: User Story Generation and Quality Assurance. Based on the aggregated subset of acceptance criteria, a dual-agent collaborative mechanism is introduced—the generating agent self-evaluates and the reviewing agent independently evaluates, judging reliability by the difference in scores: automatic output when confidence is high, and user feedback is obtained through minimal necessary interaction when confidence is low, and is stored as long-term memory. This feedback is then used as contextual input when generating user stories in subsequent scenarios, making the generating agent's inferences more accurate in similar business scenarios.

[0043] like Figure 1As shown, in a preferred embodiment of the present invention, the multi-agent collaborative business requirement generation method based on task acceptance criteria includes the following steps S1 to S4. The specific implementation process of each step is described in detail below.

[0044] S1. The set of input acceptance criteria (AC) is transformed into a weighted undirected graph using a large language model-related discrimination method, providing a structured semantic relational basis for subsequent aggregation.

[0045] It should be noted that in S1 of this invention, as Figure 2 As shown, the specific process of generating a weighted undirected graph using the large language model correlation discrimination method is as follows:

[0046] S11. Extract the business semantic elements from each acceptance criterion using the large language model, traverse all acceptance criteria, select a unique acceptance criterion as the target acceptance criterion during each traversal, and use all other acceptance criteria as candidate acceptance criteria. The large language model evaluates the semantic relevance between the target acceptance criterion and each candidate acceptance criterion based on the business semantic elements, and outputs the confidence that these two acceptance criteria belong to the same user story.

[0047] In S11 of this invention, the acceptance criteria describe business rules in a Given / When / Then (GWT) three-part format, with each acceptance criterion corresponding to an atomic business behavior description. The business semantic elements extracted from the acceptance criteria include: (1) domain concept elements—domain entities (such as "order" and "customer") and value objects involved in the acceptance criterion; (2) scenario type elements—determining whether the acceptance criterion belongs to the main path scenario, abnormal path scenario, or edge scenario; (3) business rule elements—state transition conditions, data constraints, and business invariants; (4) domain event elements—if the acceptance criterion describes a state change, determining its corresponding domain event and its trigger-response chain. After extraction, the extraction results of the above business semantic elements are stored in key-value pairs, with each acceptance criterion corresponding to one business semantic element record, serving as the structured context input for subsequent semantic relevance evaluation, thereby improving the accuracy and interpretability of the large language model evaluation.

[0048] In S11 of this invention, when the large language model evaluates the semantic relevance between acceptance criteria, the input includes: (1) the original text of the Given / When / Then of the two acceptance criteria; (2) the extracted business semantic elements, including domain concept elements, scenario type elements, business rule elements, and domain event elements. When evaluating semantic relevance, the large language model comprehensively considers four dimensions, namely: domain concept overlap, which indicates whether the domain entities and value objects involved in the acceptance criteria overlap; scenario type consistency, which indicates whether the acceptance criteria belong to the same scenario type; business rule dependency, which indicates whether the acceptance criteria have a dependency chain between business rule elements and acceptance criteria; and event causal relationship, which indicates whether the acceptance criteria have a domain event and its trigger-response chain.

[0049] In this embodiment S11, each time the large language model is called, the large language model makes a judgment according to the following rules: (1) If the same role and operation are involved in the when condition of the two acceptance criteria, it should be selected; (2) If the same domain object is involved in the Then condition of the two acceptance criteria, it should be selected; (3) If the Then condition of the first acceptance criterion involves a domain object or state change, and if the domain object involved in the When condition of the second acceptance criterion or the state involved in the Given condition is consistent, it can be selected. Each call outputs a list of selected acceptance criteria and a confidence score that measures semantic relevance. For any two acceptance criteria and Acceptance criteria As the target acceptance criteria, acceptance criteria When used as a candidate acceptance criterion, the large language model outputs confidence. Regarding the target acceptance criteria After the traversal is complete, the final result will be a set consisting of one or more candidate acceptance criteria that have the highest semantic relevance. and satisfy the set Candidate acceptance criteria and target acceptance criteria The confidence level is greater than or equal to 0 and less than or equal to 1, and it is not in the set. Candidate acceptance criteria and target acceptance criteria The confidence level is 0, meaning that for any There is a corresponding confidence level. For any There is a corresponding confidence level. , This is the set of all acceptance criteria. It should be noted that in this invention, the large language model is ultimately a statistical model; therefore, even two related acceptance criteria may have a confidence level of 0, but this does not affect subsequent procedures.

[0050] Furthermore, to obtain sufficient bidirectional semantic information, this embodiment adopts a bidirectional independent evaluation mechanism: that is, each time the large language model is called, one acceptance criterion is selected as the target acceptance criterion, and all other acceptance criteria are selected as candidate acceptance criteria. For any two acceptance criteria... and Confidence level and Since the data originates from two independent traversals, bidirectional information is naturally present, eliminating the need for additional calls to a large language model. Therefore, the time complexity of the entire graph construction process is O(log n). Second largest language model call. This represents the total number of acceptance criteria.

[0051] S12. Treat the acceptance criteria as nodes in a weighted undirected graph, and establish an undirected edge between two acceptance criteria with a bidirectional confidence score greater than zero. Use the bidirectional confidence score as the weight of the undirected edge to form a weighted undirected graph. The bidirectional confidence score is the sum of the two confidence scores obtained when the two acceptance criteria are used as the target acceptance criterion and the candidate acceptance criterion, respectively.

[0052] In this embodiment S12, each node in the weighted undirected graph corresponds to an acceptance criterion, therefore the node set The size is equal to the total number of acceptance criteria. For any two nodes corresponding to any two different acceptance criteria... If the sum of the two-way confidence scores is greater than zero, that is... Then at node and Establish an undirected edge between them to obtain the edge set. For undirected edges Its weight That is, nodes The corresponding two-way confidence level, that is, the acceptance criteria As a target acceptance criterion, acceptance criteria The confidence score of the large language model output when used as a candidate acceptance criterion, and the acceptance criterion As a target acceptance criterion, acceptance criteria The sum of the confidence scores output by the large language model when used as candidate acceptance criteria is expressed as:

[0053]

[0054] This invention uses symmetric and formal undirected edge weights to simultaneously reflect the... arrive And from and The evaluation results in these two directions indicate that the range of values ​​for the undirected edge weights is... Through the above process, this invention achieves the transformation of the aforementioned set of acceptance criteria into a weighted undirected graph structure, resulting in a weighted undirected graph. Recorded as , It is a set of weights for undirected edges, and the weighted undirected graph structure is stored in the form of an adjacency list or adjacency matrix for use by the ILP solver interface.

[0055] S2. Transform the weighted undirected graph into an integer linear programming (ILP) problem. With the goal of maximizing the sum of weights of undirected edges in the retained weighted undirected graph, the optimal graph partitioning scheme is obtained by solving the integer linear programming problem. This graph partitioning scheme aggregates the acceptance criteria into several connected components, and each connected component corresponds to a subset of acceptance criteria associated with a user story.

[0056] It should be noted that in S2 of this invention, the integer linear programming problem includes three constraints: node exclusivity constraint, upper limit constraint on the size of connected components, and undirected edge retention constraint.

[0057] In this embodiment, as Figure 3 As shown, the specific forms of the above three constraints are further explained.

[0058] Constraint 1 – Node Exclusivity Constraint: Each acceptance criterion belongs to only one connected component, expressed as:

[0059]

[0060] in, It is a set of connected components; It is a connected component in the set of connected components; To indicate the acceptance criteria Does it belong to a connected component? Discrete decision variables. In this embodiment, the user story is modeled as a connected component in a weighted undirected graph, if the acceptance criteria... Belongs to connected components ,but =1, otherwise It is 0.

[0061] Constraint 2 – Maximum Size of Connected Components: The number of acceptance criteria in each connected component cannot exceed the maximum number of acceptance criteria that can be included in each user story. , is represented as:

[0062]

[0063] In this embodiment, This is a hyperparameter that can be set by those skilled in the art according to actual needs, and is not limited in this invention. It should be noted that this invention does not set a lower limit on the number of acceptance criteria, meaning that a user story is allowed to contain only one acceptance criterion.

[0064] Constraint 3 – Undirected Edge Retention Constraint: An undirected edge can only be retained in a connected component if both nodes of the undirected edge belong to the same connected component. This is expressed as:

[0065]

[0066]

[0067] in, To indicate the acceptance criteria Does it belong to a connected component? Discrete decision variables; To represent undirected edges Do the two nodes belong to a connected component? Discrete decision variables, if acceptance criteria and Both belong to connected components ,but =1, otherwise It is 0.

[0068] It should be noted that the conditions It is inherently nonlinear. Equivalent linearization is achieved by introducing the two linear inequalities mentioned above. When the objective function... When maximized, Will belong to both nodes It automatically takes 1, because when Setting the value to 0 will decrease the objective function value without affecting constraint satisfaction. Therefore, in the optimal solution, the discrete decision variables... It must be 1.

[0069] In the objective function described above, the optimization objective of this invention is to maximize the sum of the weights of the undirected edges retained after graph partitioning. Undirected edges with higher weights are retained because the corresponding two acceptance criteria are more likely to belong to the same user story. Undirected edges with lower weights may indicate that although the two acceptance criteria are relatively related, they may belong to different user stories (e.g., login and logout). If there is no upper limit on the maximum number of acceptance criteria, the optimization will be successful. This constraint causes the acceptance criteria for related user stories to be aggregated together. Therefore, the optimization conditions proposed in this invention aim to work with the constraint to eliminate correlations that do not belong to the same user story.

[0070] Based on the aforementioned objective function and three constraints, this invention constructs a solvable integer linear programming problem. Depending on the total number of acceptance criteria, this embodiment employs a hierarchical solution strategy to ensure that small-scale problems (…) To obtain the globally optimal solution, large-scale problems ( The approximate optimal solution is obtained within an acceptable timeframe, as detailed below:

[0071] (1) Small-scale problems

[0072] This embodiment uses a standard ILP solver (such as CBC, CPLEX, Gurobi) for direct solution. At this scale, these ILP solvers can obtain the global optimum or an approximate optimum with an optimality gap of less than 1% within a reasonable time. For example, it can be implemented using `pulp.LpProblem("Graph Partitioning", pulp.LpMaximize)` from Python's `pulp` library. The variables x, w, and z are defined using `pulp.LpVariable`, the constraints and objective function described above are defined using `pulp.lpSum`, and finally, `prob.solve` is used for the solution.

[0073] (2) Large-scale problems

[0074] For large-scale problems, this embodiment presents a specific algorithm:

[0075] Phase 1: Greedy Construction

[0076] The goal of this stage is to quickly generate an initial matrix that satisfies the constraints. .

[0077] definition: For connected components The current set of nodes during the iteration process; For connected components The current number of nodes during the iteration process; This is a mapping from the list of nodes to their respective connected components, that is, . Acceptance criteria Does it belong to a connected component? Discrete decision variables; Acceptance criteria Does it belong to its connected component? Discrete decision variables, according to the definition, It is always 1.

[0078] algorithm:

[0079] 1. State initialization, for ,initialization , .

[0080] 2. Set the edge set According to weight Sort in descending order.

[0081] 3. Traverse the edge set ,for ,if That is, the edges already fall within the same connected component, or If, then skip; if swap The node represented; ,make , ;make .in, Indicating acceptance criteria Does it belong to its connected component? The discrete decision variables are always 1 by definition; Indicates the acceptance criteria in the current iteration. The number of acceptance criteria contained in the connected component; Indicates the acceptance criteria in the current iteration. The number of acceptance criteria contained in the connected component; Indicates the acceptance criteria in the current iteration. The set of current nodes in the connected component during the iteration process; Indicating acceptance criteria Does it belong to its connected component? Discrete decision variables; Indicating acceptance criteria Does it belong to its connected component? Discrete decision variables.

[0082] Phase Two: FM Local Search

[0083] definition: For nodes The internal weights; For nodes Relative to a certain connected component External weights and; Define as a node Move to an external connected component The shift gain of the connected components. Wherein, Acceptance criteria Does it belong to The connected components Discrete decision variables.

[0084] algorithm:

[0085] 1. Traverse each node Iterate through nodes that are connected to the given node and satisfy the following conditions: external connected components Calculate the moving gain Select the one that maximizes the movement gain. .in, These represent the acceptance criteria for achieving maximum movement gain and the connected components, respectively, during the traversal process.

[0086] 2. Remember The currently connected component is For all neighboring nodes (in other words The following updates will be made:

[0087]

[0088]

[0089] if Belongs to connected components ,So ;

[0090] if Belongs to connected components ,So Otherwise, no updates will be made.

[0091] The maximum number of iterations for the above algorithm, or The iteration stops when the value is less than a certain set hyperparameter.

[0092] in, For nodes With nodes Edge weights between them; For nodes Does the corresponding acceptance criterion belong to a connected component? Decision variables; For nodes Does the corresponding acceptance criterion belong to a connected component? Decision variables; For nodes Does the corresponding acceptance criterion belong to a connected component? Decision variables; For neighboring nodes Relative to connected components External weights and; For neighboring nodes Relative to connected components External weights and; For neighboring nodes The internal weights; For nodes Move to connected component The shift gain of the connected components.

[0093] After the solution is completed, based on the decision variables The value of is used to assign acceptance criteria to the corresponding connected components. The actual number of connected components may be less than the preset total number of connected components. That is, some connected components may be empty, which is normal; the total number of connected components only serves as an upper limit for the potential number of connected components. For each connected component... It aggregates all the acceptance criteria it contains into a subset of acceptance criteria, which serves as input for subsequent steps.

[0094] S3. A dual-agent collaborative mechanism is adopted to generate user stories for each connected component. The generating agent generates user stories in a standard format according to a subset of acceptance criteria for each connected component, and the reviewing agent independently evaluates the generation quality of the user stories. If the generated user stories simultaneously meet the confidence difference condition and the confidence consensus condition, they are directly output; otherwise, human-computer interaction is triggered to obtain user feedback and use it as long-term memory context to guide the generating agent to output user stories that meet the conditions.

[0095] It should be noted that in S3 of this invention, the several connected components output by S2, the subset of acceptance criteria corresponding to each connected component, and the text content of each acceptance criterion are organized into prompt words, which are input into the generating agent. The generating agent generates user stories through multi-step reasoning and outputs self-rated confidence.

[0096] In this embodiment, application service information and business rule information can also be used as additional information to input the generated intelligent agent, which generates a three-part user story through five-step reasoning:

[0097] (1) Identify user goals (I want) – Extract behaviors from the When clause of the acceptance criteria and the application service method name, and transform technical behaviors into user intentions;

[0098] (2) Inferring role (As a) - First, infer from the user's historical feedback (long-term memory context), and second, from the When condition of the acceptance criteria, the source code snippet of the controller that calls the relevant application service, the interface path, and the permission check code of the relevant application service.

[0099] (3) Inferring business value (So that) - First, match from historical semantics, and second, infer the business purpose from the state change statements and acceptance criteria related to the Then clause;

[0100] (4) Consistency check - check whether it conflicts with the user's historical feedback (long-term memory context), whether all business concepts appearing in the user story appear in the common language vocabulary, and whether the business value in the "So that" statement is from the user's perspective rather than the system's perspective;

[0101] (5) Self-rated confidence level - The generating agent gives a self-rated confidence level based on factors such as whether it is consistent with user feedback, whether the context information on which the inference is based is sufficient, and whether the reason for the inference is singular and clear. .

[0102] It should be noted that in S3 of this invention, the user story generated by the generating agent is input into the reviewing agent, and the reviewing agent evaluates the quality of the generated user story in multiple dimensions and outputs the review confidence level.

[0103] In this embodiment, the review agent operates independently of the generator agent to avoid self-calibration bias in a single agent and improve the reliability of user story quality assessment. Specifically, the generator agent inputs the generated user stories into the review agent, which performs quality assessments across multiple review dimensions. As shown in Table 1, the review dimensions used in this embodiment include: role clarity, objective understandability, value quantifiability, testability, and consistency with AC. The review agent performs structured scoring on each review dimension, with the scoring criteria for each dimension clearly defined in the input prompts to ensure consistency and interpretability of the review. Taking role clarity as an example, 1 point represents that the generated user story role is clear and reasonable, 0.5 points represents that the generated user story role is correct but inaccurate (e.g., in a library borrowing management system, it should be "borrower" but is simply output as "user"), and 0 points represents that the generated user story role is missing or incorrect. After the review agent completes the scoring, the scores for each review dimension are weighted and summed to obtain the review confidence level. .

[0104] Table 1. Review Dimensions and Weights

[0105] It should be noted that in S3 of this invention, the confidence level difference condition is that the difference between the self-assessed confidence level and the review confidence level is less than a preset difference threshold; the confidence level consensus condition is that the average confidence level is greater than or equal to a preset quality threshold, and the average confidence level is the average of the self-assessed confidence level and the review confidence level.

[0106] In this embodiment, the core challenge of step S3 lies in the fact that the large language model needs to infer roles and business value from acceptance criteria and the behavior described in the source code, and this inference is uncertain. Therefore, this invention employs dual-agent cross-validation—the contexts of the generating agent and the reviewing agent are independent of each other, and the reliability is evaluated by the difference in confidence levels of the outputs of the two agents, reducing the illusion that the generating agent may have due to long contexts.

[0107] Therefore, such as Figure 4 As shown, this embodiment first calculates the difference between the self-assessed confidence level and the reviewed confidence level. With average confidence level This embodiment stipulates that: This indicates that the two agents have a high degree of consensus on the quality of the generated data. This indicates that the generated quality has reached an acceptable level. Therefore, this embodiment sets a difference threshold. Quality threshold However, both the difference threshold and the quality threshold can be set based on engineering experience; the above threshold values ​​are merely preferred options. Based on the above two thresholds, this invention designs two conditions for judging reliability: the confidence level difference condition and the confidence level consensus condition. The confidence level difference condition is... The confidence consensus condition is Both conditions must be met simultaneously for a user story to be directly output.

[0108] Specifically, when both conditions are met (i.e.) and When a certain condition is not met, or both conditions are not met (i.e., when the consensus between the two agents is high), no long-term memory is added; only the user story is output for use. or When the two agents are considered to have a significant divergence, human-computer interaction is triggered in this scenario. This embodiment combines the thought processes of the two agents at corresponding confidence levels, identifies the user story element with the highest uncertainty during the thought process, and generates multiple-choice questions based on the priority of business value, role, and goal. The large language model provides its own guesses for the options, requests confirmation from the user, and additionally provides an "other" option for the user to fill in. After the user answers or selects, this embodiment adds the user-confirmed role, goal, and business value to long-term memory as a subsequently generated long-term memory context. This long-term memory context only contains information confirmed by the user, and when there is a contradiction between past information and newly confirmed information, a new confirmation question will be asked.

[0109] Furthermore, when the generative agent is unable to generate a user story, this embodiment will also have the agent ask the user questions and receive user feedback until the agent is able to generate a user story. During this process, user feedback is also treated as long-term memory, used as a context for subsequent generation to ensure terminology consistency and reduce unnecessary user interactions.

[0110] S4. The large language model extracts three key concepts from user stories and acceptance criteria: application service information, domain entity information, and business rule information, forming multiple candidate contexts. For each candidate context, the large language model evaluates its relevance to the user story from multiple dimensions, calculates a weighted sum of the evaluation scores of each dimension, and selects the candidate contexts that meet the requirements of the comprehensive score as domain contexts that have substantial supplementary value to the user story. After integration with the user story, a complete business requirement document is output.

[0111] In this embodiment, step S4 aims to filter relevant domain context for each user story, enriching the descriptive background of the user story. After generating the user story in step S3, each user story only contains the role, goal, and business value inferred from the acceptance criteria, lacking the context of "why to do it" and "in what context to do it." Therefore, this step uses a multi-dimensional scoring method to filter the most relevant information from the candidate context, integrates it with the user story, and generates a complete requirements document containing the business background.

[0112] Specifically, this embodiment evaluates the relevance of candidate contexts to user stories in the following dimensions:

[0113] 1) Rule relevance: The degree of correlation between the technical actions constrained by the business rules in the candidate context and the behavior described in the "I want" statement of the user story;

[0114] 2) Object Relevance: The degree of overlap between the domain entities involved in the candidate context and the target object described in the "I want" statement of the user story;

[0115] 3) Event Relevance: The causal relationship between the domain events in the candidate context and the user story in the acceptance criteria, and the states described in the Given and Then conditions;

[0116] 4) Interface dependency: The dependency between interface or method calls in the candidate context and the behavior described in the "I want" statement of the user story;

[0117] 5) Concept coverage: The extent to which the domain concepts in the candidate context cover the business concepts involved in the user story.

[0118] When weighting and summing the evaluation scores of each dimension, the rule relevance has the highest weight. After calculating the comprehensive score, candidate contexts can be filtered by setting thresholds. For example, candidate contexts with a comprehensive score greater than or equal to a preset score threshold are considered as qualified candidate contexts, or the candidate context with the highest comprehensive score can be considered as a qualified candidate context.

[0119] To better demonstrate the specific implementation and technical effects of the present invention, the method for generating multi-agent collaborative business requirements based on task acceptance criteria aggregation shown in steps S1 to S4 of the above preferred implementation is applied to a specific example.

[0120] Example

[0121] The specific implementation process of the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation used in this embodiment is as described above and will not be repeated here.

[0122] This embodiment is validated based on the acceptance criteria and user stories of existing projects driven by five internal financial industry sectors. To demonstrate the differences in the effectiveness of this embodiment in different scenarios, it is divided into two categories according to functional type: general functions (such as login, registration, permission verification, etc.) and business functions (such as payment, credit approval, risk control, etc.), and statistics are presented separately as shown in Table 2. The following is a demonstration case of the user authentication scenario, showing the complete processing flow in detail. The statistical data of the experimental results are presented in the overall experimental results section. The basic statistical information of the experimental data is as follows:

[0123] Table 2. Experimental Data Information

[0124] This embodiment uses seven acceptance criteria in a user authentication scenario of a certain system as an example to demonstrate the complete aggregation and generation process of the present invention from acceptance criteria to user stories. The key parameter of the method is set as follows: the upper limit of the connected component size K=3. The input acceptance criteria are shown in Table 3.

[0125] From the perspective of manual annotation, AC1 / AC2 / AC3 should belong to the user stories corresponding to the core login authentication process, AC4 / AC5 should belong to the user stories corresponding to session management, and AC6 / AC7 should belong to the user stories corresponding to user account management. That is, the 7 ACs should be aggregated into 3 user stories.

[0126] Therefore, this embodiment first uses a large language model to extract the business semantic elements (domain concepts, scenario types, business rules, and domain events) of each of the seven acceptance criteria in Table 3, and then evaluates the semantic relevance between the acceptance criteria. For example, the confidence scores of AC1 and AC2 are conf = 0.85, and the confidence scores of AC2 and AC3 are conf = 0.75, both belonging to login-related categories. However, AC1 and AC4 belong to different scenarios of login and session management, with a confidence score of only 0.45. Based on the above confidence scores, a weighted undirected graph is constructed with AC as the node and the sum of the bidirectional confidence scores as the edge weight. The partitioning results are then obtained as follows: C1 = {AC1, AC2, AC3}, C2 = {AC4, AC5}, and C3 = {AC6, AC7}. C1 aggregates the main login path, abnormal path, and security restrictions; C2 aggregates session management; and C3 aggregates password management. Next, for each connected component, this embodiment generates a user story from each AC aggregation through a generating agent; the review agent evaluates the story and then reaches a consensus with the generating agent. Taking C1 as an example, the generating agent infers from AC1 / AC2 / AC3 that the role is "registered user", the goal is "login to the system with username and password", and the value is "secure access to personal account and prevention of unauthorized access". The review agent independently scores the story and determines whether the consensus conditions are met. If not, the process proceeds to manual intervention. The final generated user stories are shown in Table 4.

[0127] Table 3. Input Acceptance Criteria Information

[0128] Table 4. Generated User Stories

[0129] To evaluate the overall experimental results, this embodiment summarizes the experimental results data of the five projects according to two dimensions: general functions and business functions, and evaluates them from the aspects of aggregation quality and generation quality.

[0130] Regarding aggregation quality, the requirements design document contains the correspondence between user stories and acceptance criteria. Based on this correspondence, this embodiment groups acceptance criteria belonging to the same user story; this group is considered the ground truth. The consistency between the acceptance criterion divisions generated in this embodiment and the ground truth acceptance criterion groupings is measured using ARI (Adjusted RAND Index). ARI compares the consistency of each acceptance criterion pair in the "same group / different group" of the two divisions and adjusts for random expectation, with a value range of [-1, 1], where 1 represents perfect consistency and 0 represents a random level. The aggregation results for each dimension are shown in Table 5.

[0131] Table 5. Aggregation Results

[0132] Regarding the quality of generated content, each original user story is manually compared with all generated user stories in the same project, and the degree to which its core business intent is covered is determined according to the following three levels:

[0133] Full coverage: The core intent contained in the "So that" part of the original user story has a clear correspondence in the generated user story; and the role in the "As" condition is consistent with the action in the "I Wish" condition.

[0134] Incomplete coverage: The core intent contained in the "So that" part of the original user story has some corresponding elements in the generated user story, but there are omissions or deviations; and the role in the "As" condition is consistent with the action in the "I Wish" condition.

[0135] Complete mismatch: The core intent has no corresponding counterpart in any of the generated user stories.

[0136] After all judgments are completed, the coverage rate is calculated as "Coverage = (Number of stories fully covered + 0.5 × Number of stories partially covered) / Total number of original user stories". That is, for original user stories, a story fully covered scores 1 point, partially covered scores 0.5 points, and a complete mismatch scores 0 points. Accuracy is evaluated using a similar method: "Accuracy = (Number of stories fully covered + 0.5 × Number of stories partially covered) / Total number of generated user stories". That is, for generated user stories, a story fully covering a given original user story scores 1 point, partially covered scores 0.5 points, and a complete mismatch scores 0 points. Simultaneously, the number of times manual intervention is triggered is recorded, noting instances where the dual-agent consensus mechanism triggers manual intervention in low-confidence scenarios, and where the large language model generates invalid options. These situations incur high costs. In the general functionality dimension, the original user stories totaled 117, with 86 fully covered and 27 partially covered, resulting in a coverage rate of 85.04%. 125 user stories were generated, with 86 achieving full coverage and 29 achieving partial coverage, resulting in an accuracy rate of 80.40%. Manual input was triggered 12 times. In the business functionality dimension, the original user stories totaled 242, with 143 fully covered and 74 partially covered, resulting in a coverage rate of 74.38%. 284 user stories were generated, with 143 achieving full coverage and 87 achieving partial coverage, resulting in an accuracy rate of 65.67%. Manual input was triggered 33 times.

[0137] In summary, the method of this invention can effectively achieve semantic-level aggregation and user story generation in both general and business functional scenarios of projects driven by five internal domains of the financial industry. From an aggregation perspective, the generation segmentation for general scenarios is more consistent with the manual annotation scheme, with an ARI of 0.79. The segmentation effect for business scenario generation is reduced but still usable, with an ARI of 0.66. In terms of generation quality, the core intent coverage of user stories generated in general functional scenarios reaches 85.04%, with an accuracy of 80.40%; the core intent coverage of user stories generated in business functional scenarios reaches 74.38%, with an accuracy of 65.67%. Furthermore, with the guarantee of the dual-agent consensus mechanism, the number of manual input triggers is controlled to an average of 9 times per project.

[0138] It is understood that the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation described in S1-S4 above can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation provided in the above embodiments. This product includes a computer program / instruction, which, when executed by a processor, can implement the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation as described in the above embodiments.

[0139] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation provided in the above embodiments, such as... Figure 5 As shown, it includes a memory and a processor;

[0140] The memory is used to store computer programs;

[0141] The processor is configured to, when executing the computer program, implement the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation in the above embodiments.

[0142] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0143] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can realize the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation in the above embodiments.

[0144] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S4.

[0145] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0146] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0147] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for generating multi-agent collaborative business requirements based on task acceptance criteria, characterized in that, Includes the following steps: S1. The set of input acceptance criteria is transformed into a weighted undirected graph using a large language model-related discrimination method; S2. Transform the weighted undirected graph into an integer linear programming problem. With the goal of maximizing the sum of the weights of the undirected edges in the weighted undirected graph, the optimal graph partitioning scheme is obtained by solving the integer linear programming problem. This graph partitioning scheme aggregates the acceptance criteria into several connected components, and each connected component corresponds to a subset of the acceptance criteria associated with a user story. S3. A dual-agent collaborative mechanism is adopted to generate user stories for each connected component. The generating agent is used to generate user stories in a standard format according to a subset of acceptance criteria for each connected component, and the reviewing agent is used to independently evaluate the generation quality of user stories. If the generated user story satisfies both the confidence difference condition and the confidence consensus condition, it is output directly; otherwise, human-computer interaction is triggered to obtain user feedback and use it as long-term memory context to guide the generating agent to output user stories that meet the conditions. S4. The large language model extracts three key concepts from user stories and acceptance criteria: application service information, domain entity information, and business rule information, forming multiple candidate contexts. For each candidate context, the large language model evaluates its relevance to the user story from multiple dimensions, calculates a weighted sum of the evaluation scores of each dimension, and selects the candidate contexts that meet the requirements of the comprehensive score as domain contexts that have substantial supplementary value to the user story. After integration with the user story, a complete business requirement document is output. The specific process of generating a weighted undirected graph using the large language model correlation discrimination method is as follows: S11. Extract the business semantic elements from each acceptance criterion using the large language model, traverse all acceptance criteria, select a unique acceptance criterion as the target acceptance criterion during each traversal, and use all other acceptance criteria as candidate acceptance criteria. The large language model evaluates the semantic relevance between the target acceptance criterion and each candidate acceptance criterion based on the business semantic elements, and outputs the confidence that these two acceptance criteria belong to the same user story. S12. Treat the acceptance criteria as nodes in a weighted undirected graph, and establish an undirected edge between two acceptance criteria with a bidirectional confidence score greater than zero. Use the bidirectional confidence score as the weight of the undirected edge to form a weighted undirected graph. The bidirectional confidence score is the sum of the two confidence scores obtained when the two acceptance criteria are used as the target acceptance criterion and the candidate acceptance criterion, respectively.

2. The method for generating multi-agent collaborative business requirements based on task acceptance criteria as described in claim 1, characterized in that, The business semantic elements extracted from the acceptance criteria include domain concept elements, scenario type elements, business rule elements, and domain event elements.

3. The method for generating multi-agent collaborative business requirements based on task acceptance criteria as described in claim 2, characterized in that, When evaluating semantic relevance, the large language model comprehensively considers four dimensions: domain concept overlap, which indicates whether the domain entities and value objects involved in the acceptance criteria overlap; scenario type consistency, which indicates whether the acceptance criteria belong to the same scenario type; business rule dependency, which indicates whether the acceptance criteria have a dependency chain between business rule elements and the acceptance criteria; and event causal relationship, which indicates whether the acceptance criteria have a domain event and its trigger-response chain.

4. The method for generating multi-agent collaborative business requirements based on task acceptance criteria as described in claim 1, characterized in that, The integer linear programming problem includes three constraints: node exclusivity constraint, upper limit constraint on the size of connected components, and undirected edge retention constraint.

5. The method for generating multi-agent collaborative business requirements based on task acceptance criteria as described in claim 4, characterized in that, The node exclusivity constraint is that each acceptance criterion belongs to only one connected component; the upper limit constraint for the size of a connected component is that the number of acceptance criteria contained in each connected component cannot exceed the maximum number of acceptance criteria contained in each user story. The constraint for retaining undirected edges is that an undirected edge can only be retained in a connected component if both nodes of the undirected edge belong to the same connected component.

6. The method for generating multi-agent collaborative business requirements based on task acceptance criteria as described in claim 1, characterized in that, The connected components output by S2, the subset of acceptance criteria corresponding to each connected component, and the text content of each acceptance criterion are organized into prompt words, which are then input into the generating agent. The generating agent generates user stories through multi-step reasoning and outputs the self-rated confidence level.

7. The method for generating multi-agent collaborative business requirements based on task acceptance criteria as described in claim 6, characterized in that, The user stories generated by the generating agent are input into the reviewing agent. The reviewing agent evaluates the quality of the generated user stories across multiple dimensions and outputs the review confidence score.

8. The method for generating multi-agent collaborative business requirements based on task acceptance criteria as described in claim 7, characterized in that, The confidence level difference condition is that the difference between the self-assessed confidence level and the review confidence level is less than the preset difference threshold; the confidence level consensus condition is that the average confidence level is greater than or equal to the preset quality threshold, and the average confidence level is the average of the self-assessed confidence level and the review confidence level.

9. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can realize the multi-agent collaborative business requirement generation method based on task acceptance criteria aggregation as described in any one of claims 1 to 8.

Citation Information

Patent Citations

  • Automatic software development system based on multi-agent cooperation

    CN121349876A

  • Hierarchical function demand generation method and system based on UI layer code

    CN122470152A