A software development process automatic execution method based on multi-agent cooperation

Abstract

The application discloses a kind of based on multi-agent cooperation's software research and development process automatic execution method, belong to software development technical field.The application includes: generating task negotiation specification according to target task in software research and development;Each candidate agent generates initial bid package, and aggregation generates first round negotiation view containing bid distribution density, carries out multiple rounds of negotiation and bid update;Determine main force agent and standby candidate sequence, carry out preplan credible discrimination verification to candidate agent in standby candidate sequence, and update to obtain final standby agent sequence;Main force agent executes target task, and real-time acquisition task execution state data, if there is breach of contract risk, then carry out hot switching, from standby agent sequence in order to select suitable standby agent to replace execution target task;After task is completed, settlement and reputation update are carried out.The application can realize agent prior credible discrimination, in-process performance monitoring and post-smooth replacement.

Description

Technical Field

[0001] This invention belongs to the field of software development technology, specifically relating to an automatic execution method for software development processes based on multi-agent collaboration. Background Technology

[0002] As the complexity of software development continues to increase, multi-agent systems, with their autonomous decision-making and collaborative capabilities, are being increasingly applied to the automation of development processes. Existing solutions mainly achieve task allocation and execution through sequential collaboration or static capability matching between agents, but they still have significant shortcomings in dynamic negotiation, trust identification, and anomaly succession.

[0003] A search revealed that patent CN121349876A discloses an automated software development system based on multi-agent collaboration. This system achieves code output through a pipeline-style collaboration of agents in requirements analysis, detailed design, and project development. However, its task allocation is fixed, lacking consideration for the real-time load and historical performance of agents, and it lacks a backup mechanism. If an agent deviates from its assigned task, the entire process is forced to stop. Another search revealed that patent CN121326306A discloses a low-code development system and method for multi-agent collaborative applications. This system assigns tasks based on capability matching through a management and scheduling engine. However, its matching relies on static capability tags, lacking multi-round negotiation and pre-emptive verification of contingency plans. This makes it impossible to identify exaggerated claims by agents in their capability declarations, and it is also difficult to perceive the accumulation of performance risks during task execution.

[0004] The aforementioned existing technologies generally lack dynamic negotiation mechanisms between agents and adversary behavior modeling mechanisms, making it impossible to achieve optimal allocation under information asymmetry. Furthermore, they lack pre-execution credible verification of agents' claimed capabilities, making them prone to adverse selection. During task execution, they lack real-time performance monitoring and default warning capabilities, resulting in only passive responses after failures occur, and the handover process lacks pre-planned readiness verification, leading to progress rollbacks and resource waste. Therefore, there is an urgent need for a fully automated execution method that enables pre-event credible agent identification, in-event performance monitoring, and smooth post-event handover. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an automatic execution method for software development process based on multi-agent collaboration.

[0006] The objective of this invention is achieved through the following technical solution: an automated execution method for software development processes based on multi-agent collaboration, comprising:

[0007] Extract the task execution context of the target task to be assigned in software development and encapsulate it into a task negotiation tender document that carries a capability requirement vector, relative deadline, and dependency constraints.

[0008] The task negotiation proposal is broadcast to the set of candidate agents that meet the capability requirements in the multi-agent registry. Each candidate agent generates an initial bid package, and the packages are aggregated to generate a first-round negotiation view containing the bid distribution density.

[0009] Each candidate agent obtains the bidding strategies of other candidate agents based on the first round of negotiation view, and at the same time evaluates the marginal gain after withdrawing its own bid based on counterfactual reasoning, updates its own initial bid package, and injects it into the first round of negotiation view; multiple rounds of negotiation iteration are performed to output a stable set of bid packages;

[0010] Based on the set of stable bid packages, the candidate agents are sorted to obtain the main agent and a standby candidate sequence containing a preset number of candidate agents.

[0011] Perform contingency plan credibility verification on the candidate agents in the standby candidate sequence and generate a verification report; the candidate agents in the standby candidate sequence pledge a preset amount of credit assets to the credit asset custody contract and calculate the contingency plan fulfillment status after the candidate agent is activated; update the standby candidate sequence according to the verification report and the contingency plan fulfillment status to obtain a standby sequence containing standby agents.

[0012] The main agent executes the target task and collects task execution status data. If there is a commitment offset, a standby agent is selected from the standby sequence to execute the target task.

[0013] Furthermore, the task execution context of the target task to be assigned in software development is extracted and encapsulated into a task negotiation tender carrying a capability requirement vector, relative deadline, and dependency constraints, including: extracting the pre-dependent outputs, associated requirement slices, and current pipeline status of the target task to be assigned from the shared context pool.

[0014] Furthermore, each candidate agent generates an initial bid package, and these packages are aggregated to generate a first-round negotiation view containing the bid distribution density, including:

[0015] The initial bid package includes the candidate agent's current operating load, historical distribution of completion times for similar tasks, committed completion time, and commitment confidence level. The committed completion time and commitment confidence level are calculated by understanding the task requirements.

[0016] Each candidate agent submits its initial bid package and its digitally signed capability self-certification digest to the temporary sandbox of the negotiation node. The negotiation node aggregates the initial bid packages and generates a first-round negotiation view containing the bid distribution density, and broadcasts it to all candidate agents.

[0017] Furthermore, each candidate agent obtains the bidding strategies of other candidate agents based on the first-round negotiation view, and simultaneously evaluates the marginal gain after withdrawing its own bid based on counterfactual reasoning, updates its initial bid package, and injects it into the first-round negotiation view; multiple rounds of negotiation iterations are performed to output a stable bid package set, including:

[0018] After receiving the first round of negotiation view, each candidate agent enters a multi-round negotiation phase. In each round of negotiation, each candidate agent dynamically infers the private constraints and concession bottom lines of other candidate agents based on the previous round of negotiation view and the bidding strategies of other candidate agents in the previous round, and updates its own bidding strategy.

[0019] Based on counterfactual reasoning, the marginal impact of withdrawing one's own bid on the ranking of the negotiation results is estimated, and a counterfactual gain value is generated. Each candidate agent combines the counterfactual gain value with its own preference weight for the target task to determine the bidding strategy for this round. The bidding strategy includes sticking to the original bid, making concessions, or giving up the bid.

[0020] Each candidate agent injects its updated bidding strategy, along with a causal explanation record containing the bidding strategy adjustment logic, into the negotiation view until bidding convergence or the upper limit of negotiation rounds is reached. Then, it outputs a set of stable bidding packages and the corresponding causal chain log of the negotiation process.

[0021] Furthermore, based on the stable bid package set, the candidate agents are sorted to obtain the main agent and a standby candidate sequence containing a preset number of candidate agents. This includes: sorting all candidate agents based on a multi-factor decision function that includes the commitment completion time, commitment confidence, and the reasonableness of explanation in the causal chain log; taking the candidate agent at the top of the sort as the main agent; and selecting a preset number of candidate agents immediately following the main agent in the sort to form the standby candidate sequence.

[0022] Further, the candidate agents in the standby candidate sequence are subjected to contingency plan credibility verification, and a verification report is generated; the candidate agents in the standby candidate sequence pledge a preset amount of credit assets to the credit asset custody contract, and the contingency plan fulfillment status after the candidate agent is activated is calculated; the standby candidate sequence is updated according to the verification report and the contingency plan fulfillment status to obtain a standby sequence containing standby agents, including:

[0023] For each candidate agent in the standby candidate sequence, it is required to submit an emergency response plan and corresponding emergency response executable script for the declared abnormal scenario type within a specified time limit;

[0024] The negotiation node retrieves historical fault instances and extracts parameter boundaries using the abnormal scenario types declared in the emergency response plan as seeds, generates a set of verification test cases covering the parameter boundaries, executes the emergency response executable script in the isolation sandbox, generates a plan sandbox verification report, and extracts the mapping rules from defect types to runtime monitoring indicators as risk labeling tags during the verification process.

[0025] Each candidate agent in the standby candidate sequence pledges a preset amount of credit assets to the credit asset custody contract. If the agent is not activated for replacement, the pledge will be automatically released and returned after the target task is completed. If the agent is activated for replacement, the agent will be penalized or returned based on the fulfillment of the plan.

[0026] The negotiation node uses the consistency score of the verification report, the historical emergency integrity profile, and the sufficiency of the pledge as correction terms to correct the multi-factor decision function and reorder the standby candidate sequence to obtain a standby sequence containing the standby agent.

[0027] The consistency score of the verification report is calculated from the performance indicators of the emergency response executable script under several verification test cases. The historical emergency integrity profile is obtained from the exponential decay weighted average of the fulfillment of each contingency plan. The sufficiency of the pledge is the upper limit cutoff value of the ratio of the actual pledged credit assets to the minimum pledge amount corresponding to the task.

[0028] The main intelligent agent, standby sequence and its emergency response plan are bound into a task execution contract, and the task execution contract, plan sandbox verification report, risk labeling and negotiation process causal chain log are solidified into the task negotiation evidence storage.

[0029] Furthermore, the main agent executes the target task and collects task execution status data. If a commitment offset exists, a standby agent is selected from the standby sequence to execute the target task, including:

[0030] The main intelligent agent executes the target task according to the task execution contract;

[0031] By setting state listening probes in the shared context pool, the phased output rate, quality metadata and resource consumption fluctuations of the main intelligent agent are collected in real time and input into a pre-trained commitment drift detector.

[0032] When the probability of default output by the commitment drift detector exceeds a preset threshold, a succession request is sent sequentially to the standby sequence in the task negotiation and storage.

[0033] After receiving the replacement request, each standby agent reports its current availability and selects the first standby agent that reports that it is available to take over as the replacement agent.

[0034] If the risk label of the replacement agent contains a mapping rule that matches the current abnormal symptoms, then a lightweight contingency plan readiness verification is performed on the emergency response executable script for the replacement agent. After the verification is passed, the replacement is completed. After the replacement is completed, the actual execution trajectory of the replacement agent and the abnormal parameters that triggered the replacement are recorded in the task negotiation and storage and the task execution contract is updated.

[0035] Furthermore, the method also includes: after the target task is completed, calculating the commitment fulfillment deviation based on the task negotiation and storage, the actual completion time and the quality assessment results, and performing time-series fusion of the commitment fulfillment deviation with the historical reputation record of the corresponding intelligent agent, and updating the reliability dimension and bid deviation prediction parameter in the capability vector of each intelligent agent in the registry center;

[0036] For standby agents that have pledged a preset amount of credit assets to the credit asset custody contract, if they are not activated and replaced, their credit assets will be released and returned. If they are activated and replaced, the decision to confiscate or return the assets will be made based on the fulfillment of the contingency plan, and the historical emergency integrity profile of the standby agent will be updated.

[0037] The present invention provides an electronic device, including a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the method.

[0038] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0039] The beneficial effects of this invention are as follows: 1. This invention proposes a reliable screening mechanism for emergency response plans for selecting standby agents. By requiring agents entering the standby candidate sequence to submit emergency response plans in the form of executable scripts, and by having an anomaly trigger sampler automatically generate a set of verification test cases based on real fault instances in the historical fault mode library that match the scenario type declared in the plan, the plan is offline simulated and verified in an isolated sandbox. This transforms the traditional ex-post reputation penalty into an ex-ante verifiable proof of performance capability, eliminating the adverse selection space of candidate agents in emergency capability declarations, and ensuring that only agents with genuine anomaly handling capabilities can pass the verification and enter the standby sequence.

[0040] 2. This invention proposes a method for ranking and incentivizing standby agents by integrating sandbox simulation results with asset pledging game theory. After determining the standby candidate sequence, the consistency score of the contingency plan sandbox verification report, the historical emergency integrity profile, and the sufficiency of assets pledged by the candidate agents to the credit asset custody contract are used as additional correction items to reorder the original task bidding rankings to form the final standby sequence. At the same time, a conditional pledging closed loop is established throughout the task lifecycle, which automatically unpleases and returns the pledged assets if the agent is not activated, and settles the accounts based on the contingency plan fulfillment rate after activation and replacement. This ensures that when submitting contingency plans, candidate agents have no incentive to exaggerate their emergency response capabilities, and have sufficient incentives to ensure the actual executability of the contingency plans.

[0041] 3. This invention proposes a dual hot-swap triggering and execution method based on commitment drift detection and contingency plan readiness verification. During the execution of tasks by the main agent, the method continuously monitors the performance status of dimensions such as output rate, defect density, and resource consumption through a time-series encoder and calculates the probability of default. When the probability of default exceeds a threshold, a replacement inquiry is initiated sequentially according to the standby sequence. At the same time, before the formal replacement is executed, a lightweight readiness review is performed on the emergency response executable script of the replacement agent based on the abnormal precursor mapping rules contained in the risk label extracted during the contingency plan verification stage. This upgrades the simple passive switching to a hierarchical switching process of "default prediction-driven active triggering and contingency plan matching degree secondary verification". This ensures that the emergency plan of the replacement agent is still effective in the current abnormal state while minimizing progress rollback, and feeds back the actual execution trajectory after the replacement to the evolutionary closed loop of the emergency integrity profile and the abnormal trigger sampler. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a main flowchart of the overall method provided in the embodiments of the present invention;

[0044] Figure 2 A flowchart for task tender generation and initial negotiation provided in this embodiment of the invention;

[0045] Figure 3 This is a flowchart of the restricted multi-round negotiation and bidding strategy update provided in an embodiment of the present invention;

[0046] Figure 4 A flowchart for contingency plan reliability verification and standby sequence rearrangement provided in this embodiment of the invention;

[0047] Figure 5 The flowchart for commitment drift detection and emergency hot-switching provided in this embodiment of the invention is shown. Detailed Implementation

[0048] To enable those skilled in the art to better understand the technical solutions in the embodiments of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art should fall within the protection scope of the embodiments of this application.

[0049] like Figure 1 As shown, this embodiment of the invention provides a method for automatically executing a software development process based on multi-agent collaboration, including:

[0050] S1: For the target task to be assigned in the software development process, extract the pre-dependent deliverables, related requirement slices and current pipeline status of the target task from the shared context pool, encapsulate them into a task negotiation proposal carrying a capability requirement vector, relative deadline and dependency constraints, and broadcast the task negotiation proposal to the set of candidate agents in the multi-agent registry that meet the basic capability screening conditions to start a task negotiation session.

[0051] In this embodiment, as Figure 2 As shown, the specific process for generating the task negotiation proposal is as follows:

[0052] Extract the dependency subgraph associated with the target task from the shared context pool. The dependency subgraph is represented as follows: , among which, nodes Directed edges corresponding to the finished or ongoing upstream task products From node Pointing to node ,express The execution of the corresponding task requires The output is used as the input condition.

[0053] Traverse the dependency subgraph, extract the version snapshot, quality check status, and output storage path of each upstream task's output, and construct a dependency constraint vector. , among which, components The value is taken as the ready flag bit of the upstream product. This indicates that the upstream product has passed quality verification and can be read. This indicates that it is not yet ready.

[0054] Furthermore, a set of functional semantic labels is extracted from the associated demand slices, and this set of functional semantic labels is input into a pre-trained word embedding matrix. Encode, where, For embedded dimensions, To obtain the required semantic vector, the embedding vectors of each tag are averaged bitwise to determine the size of the semantic tag vocabulary. .

[0055] Demand semantic vector Through a parameter matrix The defined linear mapping layer yields the capability requirement vector. ,in, The number of capability dimensions defined in the multi-agent registry. The Each component This indicates that completing this objective task is crucial for the first... The intensity of demand for this skill.

[0056] Among them, relative deadline Subtract the dependency constraint vector from the total time budget of the current pipeline stage. The expected readiness time for upstream products that are not yet ready has been obtained.

[0057] Finally, the capability requirement vector Relative deadline and dependency constraint vector The target task identifier is encapsulated together as a task negotiation tender.

[0058] S2: After receiving the task negotiation proposal, each candidate agent parses the task context and generates its own initial bid package, such as... Figure 2 As shown.

[0059] Specifically, the initial bid package should include at least the promised completion time and confidence level calculated based on its current operating load, the distribution of completion times for similar historical tasks, and its understanding of the task requirements.

[0060] The initial bid package and its digitally signed capability self-certification digest are submitted to the temporary sandbox of the negotiation node. The negotiation node aggregates the initial bid package and generates a first-round negotiation view containing the bid distribution density, which is then broadcast to all candidate agents.

[0061] In this embodiment, the negotiation node is the coordinating and executing entity responsible for matching bids from multiple agents. It does not participate in the bidding itself, but is only responsible for collecting the initial bid packages from each candidate agent, calculating the bid distribution density, and then generating the first round of negotiation view and broadcasting it to all agents.

[0062] S3: Entering a restricted multi-round negotiation phase, such as Figure 3As shown, in each round of negotiation, any candidate agent dynamically infers the private constraints and bottom lines of other agents based on the negotiation view and the bid adjustment magnitude and strategy tendency of other candidate agents in the previous round, and updates its own bid package accordingly.

[0063] Simultaneously, a counterfactual reasoning module is introduced. This module generates a counterfactual gain value by simulating the marginal impact on the ranking of the negotiation results after withdrawing its own bid in the current bidding sequence. Candidate agents combine the counterfactual gain value with their own preference weight for the task to decide whether to stick to the original bid, make concessions, or give up bidding in this round. The updated bid, along with a causal explanation record containing the bidding strategy adjustment logic, is injected into the negotiation view until the bidding convergence or the upper limit of the negotiation round is reached. The output is a set of stable bid packages and the corresponding causal chain log of the negotiation process.

[0064] For example, if the change in the commitment completion time of two adjacent rounds of bidding packages is less than 1% to 5% of the relative deadline, and the change in commitment confidence is less than 0.01 to 0.05, then the bidding is considered to have converged; the upper limit of the negotiation rounds can be 10 rounds.

[0065] The counterfactual reasoning module is a ranking prediction model that calculates counterfactual gain values ​​by removing / perturbing its own bids and observing changes in ranking results.

[0066] Specifically, the implementation of the adversary modeling network based on the attention mechanism is as follows: the negotiation node is at the... Maintain a bid history matrix during rounds of negotiation. ,in, The number of candidate agents participating in the negotiation. The historical bidding feature dimension retained for each candidate agent, the first... OK Storage agent in the past The hidden state of the bid sequence in the round after being encoded by the gated cyclic unit.

[0067] In this embodiment, the adversary modeling network is applicable to any intelligent agent. Regarding the adversary intelligent agent The process of inferring private constraints is as follows:

[0068] First, computational agents For intelligent agents query vector and intelligent agents key vector Sum value vector ,in, For a trainable parameter matrix, , which represents the dimension of the attention space.

[0069] Then, calculate the target agent. For all other intelligent agents Normalized attention weights .

[0070] Finally, the attention weights and value vectors are weighted and summed to obtain the adversary context representation. , After feedforward network Mapped to adversary intelligent agents Private constraint inferred value , Including the inferred adversary intelligent agents Acceptable latest completion time and lowest resource offer, intelligent agent Adjust your concessions accordingly.

[0071] S4: Based on a stable set of bid packages, the negotiation node uses a multi-factor decision function that takes into account the commitment completion time, commitment confidence, and the reasonableness of the explanation in the causal chain log to rank all candidate agents. The candidate agent ranked first is determined as the main agent, and the preset number of candidate agents immediately following it in the ranking form a standby candidate sequence.

[0072] In this embodiment, as Figure 4 As shown, for each candidate agent in the standby candidate sequence, the contingency plan credibility verification sub-process is triggered: requiring it to submit an emergency response plan and the corresponding emergency response executable script within a specified time limit.

[0073] Among them, the emergency response plan declaration targets the abnormal scenario type, the emergency response executable script contains the processing logic for dealing with the abnormal scenario type, and declares with a digital signature that it can complete the predefined compensation action within the promised time window under the abnormal scenario type and the resource consumption does not exceed the declared limit.

[0074] The negotiation node invokes the anomaly trigger sampler, using the anomaly scenario type declared in the emergency response plan as a seed, to retrieve real fault instances belonging to that type from the historical fault mode library and extract parameter boundaries. Based on this, a set of verification test cases matching the anomaly scenario type and covering its parameter boundaries is generated. The set of verification test cases is injected into the isolation sandbox to execute the emergency response executable script, collect the execution trajectory, and generate a tamper-proof plan sandbox verification report. At the same time, during the verification process, the mapping rules from defect types to runtime monitoring indicator sets are automatically extracted, and a risk label carrying the mapping rules is attached to each plan.

[0075] In a further implementation, the candidate intelligent agent pledges a preset amount of credit assets to the credit asset custody contract. If the credit assets are not activated and replaced, they are automatically unpledged and returned after the task is completed. If the credit assets are activated and replaced, the penalty or return is determined according to the fulfillment of the plan.

[0076] The negotiation node uses the consistency score of the contingency plan sandbox verification report, the historical emergency integrity profile of the candidate agents, and the sufficiency of the pledge as additional correction items to correct and reorder the original decision function values ​​of the standby candidate sequences, forming the final standby sequence that includes each standby agent.

[0077] The main intelligent agents, standby sequences and their emergency response plans are bound into task execution contracts, and the task execution contracts, plan sandbox verification reports, risk labeling tags and negotiation process causal chain logs are solidified into the task negotiation evidence storage.

[0078] The task negotiation evidence storage is a record of key process data such as task execution contracts, contingency plan sandbox verification reports, risk labeling tags, and negotiation process causal chain logs formed during the negotiation and performance of collaborative tasks of multiple intelligent agents. It is a structured encapsulation and tamper-proof solidified storage of evidence records, which is used to provide credible data basis for in-process hot switching decisions, post-event responsibility determination, credit settlement and audit traceability.

[0079] In this embodiment, the specific implementation method of the exception trigger sampler generating the verification test case set is as follows: the exception trigger sampler receives the exception scenario type identifier declared in the emergency response plan. As a seed, access the historical fault mode library, where each fault record contains a fault type label and a snapshot of the system state parameters at the time of triggering. And the fault evolution trajectory sequence.

[0080] Filter fault type tags from the historical fault mode library and All fault records that are precisely matched constitute a set of fault records of the same type. .

[0081] right Parameter snapshots for each fault record Each dimension Extract the minimum value of this dimension across all records. and maximum value This abnormal scenario type is constituted in the first The value boundary range of each parameter dimension .

[0082] Latin hypercube sampling is used to generate values ​​within the boundary interval. Verification parameter vector , ,in, It is determined by preset verification adequacy parameters.

[0083] For each validation parameter vector ,from The evolution trajectory sequence of the real fault record that has the closest Euclidean distance to the parameter vector is selected as the reference trajectory, and the verification parameter vector and the reference trajectory are concatenated to form a complete verification test case. .

[0084] All verification cases As the output of the verification test case set; each verification test case is sequentially placed in the isolation sandbox. The verification parameter vector is injected into the execution entry point of the emergency response executable script. The actual execution path length, action coverage, and peak resource consumption of the emergency response executable script after injection are collected and compared with the declared values ​​of the script item by item. The comparison results are written into the contingency plan sandbox verification report.

[0085] In this embodiment, the specific form of the multi-factor decision function and its extension method in the contingency plan credibility verification sub-process are as follows:

[0086] In the phase of determining the main agent, a multi-factor decision function is used. For candidate intelligent agents The original decision function value is .

[0087] in, Candidate intelligent agents The moment of fulfillment of the promise Candidate intelligent agents The confidence level of the commitment Candidate intelligent agents The causal explanation rationality score , , To meet The weighting coefficients.

[0088] In the contingency plan reliability screening sub-process, candidate agents in the standby candidate sequence are... Its score, after verification and correction according to the contingency plan, is calculated as follows: .

[0089] in, The consistency score of the contingency plan sandbox validation report is determined by the emergency response executable script in... The weighted harmonic average of the action coverage rate, completion time satisfaction rate, and resource consumption compliance rate under each verification test case is calculated. Candidate intelligent agents The historical emergency integrity profile value of the candidate intelligent agent The exponential decay weighted average of the fulfillment rates of each contingency plan is given below: ,in, Candidate intelligent agents The total number of times in history that the contingency plan has been verified For the first The timing of the second contingency plan verification This represents the degree of fulfillment of the contingency plan confirmed by actual succession after this verification. The attenuation coefficient; The sufficiency of the pledge is determined by the value of the candidate agent. The upper limit cutoff value for the ratio of the actual amount of pledged credit assets to the preset minimum pledge amount required for this task level. , , To meet Non-negative correction coefficients; classify each candidate agent in the standby candidate sequence according to... Arrange them in descending order to form the final standby sequence.

[0090] S5: The main agent executes the target task and collects task execution status data. If there is a commitment offset, a standby agent is selected from the standby sequence to take over the execution of the target task.

[0091] Specifically, such as Figure 5 As shown, during the process of the main agent executing tasks according to the task execution contract, the phased output rate, quality metadata and resource consumption fluctuations of the main agent are collected in real time by the state listening probe set in the shared context pool, and input into a pre-trained commitment drift detector.

[0092] When the probability of default output by the promise drift detector exceeds the preset threshold, the hot switch preparation process is triggered directly without waiting for the task to fail. The replacement query is sent to the standby sequence in the task negotiation and storage in sequence. After receiving the replacement query, the standby agent reports its current availability status. The negotiation node selects the first standby agent that reports that it can be replaced as the replacement agent according to the order of the standby sequence.

[0093] For example, the preset threshold is between 0.7 and 0.85.

[0094] Before the formal takeover is executed, if the risk label associated with the takeover agent in the task negotiation and storage contains a mapping rule that matches the currently detected abnormal signs, a lightweight contingency plan readiness verification will be performed on the emergency response executable script for the takeover agent. After the verification is passed, the takeover will be completed. After the takeover is completed, the actual execution trajectory of the takeover agent and the abnormal parameters that triggered the takeover will be recorded in the task negotiation and storage and the task execution contract will be updated.

[0095] In the above embodiments, the specific construction and operation method of the commitment drift detector are as follows: the state monitoring probe in each acquisition cycle Write a monitoring snapshot record to the inward shared context pool.

[0096] For example, the monitoring snapshot record contains the cumulative output completion rate since the task began. Cumulative code defect density and the ratio of cumulative resource consumption to budget ,in, For the first Each data collection moment.

[0097] The commitment drift detector maintains a temporal feature buffer window, which covers the most recent features. The monitoring data from each acquisition cycle constitutes the local observation matrix. The observation vector at each time step is .

[0098] Local observation matrix The input is fed into a temporal encoder composed of a Long Short-Term Memory network, and the temporal encoder outputs the hidden state vector at the current time step. ; the hidden state vector Phased target vectors agreed upon in the task execution contract Concatenate the vectors to form a phased target vector. The promise made by the main intelligent agent at the time of S2 bidding It consists of the required output completion rate, the upper limit of defect density, and the upper limit of resource consumption.

[0099] The concatenated phased target vector The input is fed into a binary classifier consisting of a fully connected layer and a sigmoid activation function, and the output is the default probability. ,in, For the Sigmoid function, This represents a vector concatenation operation. and represents the weight vector and bias term of the binary classifier.

[0100] when Exceeding the preset threshold At that time, the drift detector is promised to output a default alarm signal, driving the start of the hot-switching preparation process.

[0101] As a preferred embodiment, the method further includes S6: after the target task is completed, the commitment fulfillment deviation is calculated based on the task negotiation and storage, the actual completion time and the quality assessment results, and the commitment fulfillment deviation is fused with the historical reputation record of the corresponding intelligent agent in a time series, and the reliability dimension and bid deviation prediction parameter in the capability vector of each intelligent agent in the registry center are updated.

[0102] For standby agents who have submitted credit asset pledges, if they are not activated for replacement, their pledged assets will be automatically released and returned. If a real replacement is triggered during the execution of the task, the preset root cause analysis interface will be called to determine the fulfillment rate of the plan based on the actual execution trajectory of the replacement agent, the declaration in the emergency response plan, and the abnormal triggering parameters. Based on the fulfillment rate, the credit asset custody contract will be triggered to settle the pledged assets of the replacement agent, and the fulfillment rate will be written into the emergency integrity profile of the corresponding agent.

[0103] Meanwhile, the causal chain logs generated from this negotiation process will be used to optimize the adversary modeling network and counterfactual reasoning module. The comparison records between the contingency plan sandbox verification report and the contingency plan fulfillment rate in actual succession will be used to optimize the verification test case generation strategy of the abnormal trigger sampler and the evaluation weight of the emergency integrity profile.

[0104] In this embodiment, the specific implementation method for determining the fulfillment rate of the contingency plan and settling credit assets is as follows:

[0105] The root cause analysis interface receives four types of input data:

[0106] a. The set of predefined compensation actions declared in the emergency response plan for taking over the intelligent agent. Declaration time window and resource usage limit ;

[0107] b. The actual execution trajectory of the succeeding agent during the actual succession process, the actual execution trajectory is recorded by S5 up to the timestamp action sequence of task negotiation and notification. and the corresponding resource consumption sequence composition;

[0108] c. The exception parameter vector that triggers the succession ;

[0109] d. The criteria for determining successful compensation under this abnormal scenario as specified in the task execution contract.

[0110] The root cause analysis interface first calculates the action fulfillment rate. ,in, The cardinality of the set is used to determine the intersection of the sets, and the semantic equivalence of the actions is used as the criterion.

[0111] Then, calculate the time compliance rate. ,in, For indicator functions, take the condition if it is true. Otherwise take .

[0112] Finally, calculate the resource compliance rate. .

[0113] Contingency plan fulfillment rate It is given by the weighted product of the three. ,in, and To meet The weight.

[0114] like If the contingency plan is deemed successfully implemented, the credit asset custody contract will return the pledged assets in full to replace the intelligent agent; if If the plan is deemed unsuccessful, the credit asset custody contract will hold the pledged assets in accordance with... The proportion of the fines will be confiscated and paid to the account of the party whose task was damaged.

[0115] Will The value is written into the emergency integrity profile of the successor agent and compared with the historical fulfillment sequence. After merging, update according to the exponential decay weighted formula in S4. .

[0116] Among them, after the historical emergency integrity profile value is updated, subsequent target tasks can more accurately identify intelligent agents with high historical fulfillment rates when screening and sorting standby intelligent agents, reducing information asymmetry and adverse selection risks, while making it easier for high-quality intelligent agents to be selected or even reducing staking pressure, thus forming a positive incentive.

[0117] This invention also provides an electronic device, including a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the method described in any embodiment.

[0118] This invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the embodiments.

[0119] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0120] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for automatically executing a software development process based on multi-agent collaboration, characterized in that, include: Extract the task execution context of the target task to be assigned in software development and encapsulate it into a task negotiation tender document that carries a capability requirement vector, relative deadline, and dependency constraints. The task negotiation proposal is broadcast to the set of candidate agents that meet the capability requirements in the multi-agent registry. Each candidate agent generates an initial bid package, and the packages are aggregated to generate a first-round negotiation view containing the bid distribution density. Each candidate agent obtains the bidding strategies of other candidate agents based on the first round of negotiation view, and at the same time evaluates the marginal gain after withdrawing its own bid based on counterfactual reasoning, updates its own initial bid package, and injects it into the first round of negotiation view. Perform multiple rounds of negotiation iterations to output a stable set of bid packages; Based on the set of stable bid packages, the candidate agents are sorted to obtain the main agent and a standby candidate sequence containing a preset number of candidate agents. Perform contingency plan credibility verification on the candidate agents in the standby candidate sequence and generate a verification report; the candidate agents in the standby candidate sequence pledge a preset amount of credit assets to the credit asset custody contract and calculate the contingency plan fulfillment status after the candidate agent is activated; update the standby candidate sequence according to the verification report and the contingency plan fulfillment status to obtain a standby sequence containing standby agents. The main agent executes the target task and collects task execution status data. If there is a commitment offset, a standby agent is selected from the standby sequence to take over the execution of the target task.

2. The method according to claim 1, characterized in that, Extract the task execution context of the target task to be assigned in software development and encapsulate it into a task negotiation tender document carrying a capability requirement vector, relative deadline and dependency constraints. This includes: extracting the pre-dependent outputs, associated requirement slices and current pipeline status of the target task to be assigned from the shared context pool.

3. The method according to claim 1, characterized in that, Each candidate agent generates an initial bid package, which is then aggregated to generate a first-round negotiation view containing the bid distribution density, including: The initial bid package includes the candidate agent's current operating load, historical distribution of completion times for similar tasks, committed completion time, and commitment confidence level. The committed completion time and commitment confidence level are calculated by understanding the task requirements. Each candidate agent submits its initial bid package and its digitally signed capability self-certification digest to the temporary sandbox of the negotiation node. The negotiation node aggregates the initial bid packages and generates a first-round negotiation view containing the bid distribution density, and broadcasts it to all candidate agents.

4. The method according to claim 1, characterized in that, Each candidate agent obtains the bidding strategies of other candidate agents based on the first round of negotiation view, and at the same time evaluates the marginal gain after withdrawing its own bid based on counterfactual reasoning, updates its own initial bid package, and injects it into the first round of negotiation view. Perform multiple rounds of negotiation iterations to output a stable set of bid packages, including: After receiving the first round of negotiation view, each candidate agent enters a multi-round negotiation phase. In each round of negotiation, each candidate agent dynamically infers the private constraints and concession bottom lines of other candidate agents based on the previous round of negotiation view and the bidding strategies of other candidate agents in the previous round, and updates its own bidding strategy. Based on counterfactual reasoning, the marginal impact of withdrawing one's own bid on the ranking of the negotiation results is estimated, and a counterfactual gain value is generated. Each candidate agent combines the counterfactual gain value with its own preference weight for the target task to determine the bidding strategy for this round. The bidding strategy includes sticking to the original bid, making concessions, or giving up the bid. Each candidate agent injects its updated bidding strategy, along with a causal explanation record containing the bidding strategy adjustment logic, into the negotiation view until bidding convergence or the upper limit of negotiation rounds is reached. Then, it outputs a set of stable bidding packages and the corresponding causal chain log of the negotiation process.

5. The method according to claim 4, characterized in that, Based on the set of stable bid packages, the candidate agents are sorted to obtain the main agent and a standby candidate sequence containing a preset number of candidate agents. This includes: sorting all candidate agents based on a multi-factor decision function that includes commitment completion time, commitment confidence and the reasonableness of explanation in the causal chain log; taking the first-ranked candidate agent as the main agent; and selecting a preset number of candidate agents that are immediately after the main agent in the sorting to form the standby candidate sequence.

6. The method according to claim 5, characterized in that, The candidate agents in the standby candidate sequence are subjected to contingency plan credibility verification, and a verification report is generated; the candidate agents in the standby candidate sequence pledge a preset amount of credit assets to the credit asset custody contract, and the contingency plan fulfillment status after the candidate agent is activated is calculated; the standby candidate sequence is updated according to the verification report and the contingency plan fulfillment status to obtain a standby sequence containing standby agents, including: For each candidate agent in the standby candidate sequence, it is required to submit an emergency response plan and corresponding emergency response executable script for the declared abnormal scenario type within a specified time limit; The negotiation node retrieves historical fault instances and extracts parameter boundaries using the abnormal scenario types declared in the emergency response plan as seeds, generates a set of verification test cases covering the parameter boundaries, executes the emergency response executable script in the isolation sandbox, generates a plan sandbox verification report, and extracts the mapping rules from defect types to runtime monitoring indicators as risk labeling tags during the verification process. Each candidate agent in the standby candidate sequence pledges a preset amount of credit assets to the credit asset custody contract. If the agent is not activated for replacement, the pledge will be automatically released and returned after the target task is completed. If the agent is activated for replacement, the agent will be penalized or returned based on the fulfillment of the plan. The negotiation node uses the consistency score of the verification report, the historical emergency integrity profile, and the sufficiency of the pledge as correction terms to correct the multi-factor decision function and reorder the standby candidate sequence to obtain a standby sequence containing the standby agent. The consistency score of the verification report is calculated from the performance indicators of the emergency response executable script under several verification test cases. The historical emergency integrity profile is obtained from the exponential decay weighted average of the fulfillment of each contingency plan. The sufficiency of the pledge is the upper limit cutoff value of the ratio of the actual pledged credit assets to the minimum pledge amount corresponding to the target task. The main intelligent agent, standby sequence and its emergency response plan are bound into a task execution contract, and the task execution contract, plan sandbox verification report, risk labeling and negotiation process causal chain log are solidified into the task negotiation evidence storage.

7. The method according to claim 6, characterized in that, The primary agent executes the target task and collects task execution status data. If a commitment offset exists, a backup agent is selected from the standby sequence to take over the execution of the target task, including: The main intelligent agent executes the target task according to the task execution contract; By setting state listening probes in the shared context pool, the phased output rate, quality metadata and resource consumption fluctuations of the main intelligent agent are collected in real time and input into a pre-trained commitment drift detector. When the probability of default output by the commitment drift detector exceeds a preset threshold, a succession request is sent sequentially to the standby sequence in the task negotiation and storage. After receiving the replacement request, each standby agent reports its current availability and selects the first standby agent that reports that it is available to take over as the replacement agent. If the risk label of the replacement agent contains a mapping rule that matches the current abnormal symptoms, then a lightweight contingency plan readiness verification is performed on the emergency response executable script for the replacement agent. After the verification is passed, the replacement is completed. After the replacement is completed, the actual execution trajectory of the replacement agent and the abnormal parameters that triggered the replacement are recorded in the task negotiation and storage and the task execution contract is updated.

8. The method according to claim 7, characterized in that, The method further includes: after the target task is completed, calculating the commitment fulfillment deviation based on the task negotiation and storage, the actual completion time and the quality assessment results, and performing time-series fusion of the commitment fulfillment deviation with the historical reputation record of the corresponding intelligent agent to update the reliability dimension and bid deviation prediction parameter in the capability vector of each intelligent agent in the registry center; For standby agents that have pledged a preset amount of credit assets to the credit asset custody contract, if they are not activated and replaced, their credit assets will be released and returned. If they are activated and replaced, the decision to confiscate or return the assets will be made based on the fulfillment of the contingency plan, and the historical emergency integrity profile of the standby agent will be updated.

9. An electronic device comprising a memory and a processor, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.

Images