Intelligent agent cooperative control method and device, equipment and medium

Abstract

The application provides an agent cooperation control method and device, equipment and a medium. The method comprises the following steps: in the process of multiple agents discussing a discussion problem for multiple rounds, after each agent speaks, convergence detection is performed based on a first historical message in the current round of discussion to obtain a convergence detection result; if the convergence detection result triggers convergence, the current round of discussion is ended; in response to detecting that the current round of discussion is ended, quality evaluation is performed based on a second historical message in the current round of discussion to obtain a quality evaluation result; based on a confidence level and a recommended action, an execution strategy after the current round of discussion is ended is determined; and the execution strategy is processed until the discussion is ended, and a discussion result for the discussion problem is obtained. In this way, convergence detection is performed after each agent speaks, the current round of discussion can be directly ended when convergence is triggered, and the execution strategy is determined after each round of discussion is ended, so that the accuracy of convergence judgment is improved while the control cost is reduced.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method, apparatus, device, and medium for intelligent agent collaborative control. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, Large Language Models (LLMs), with their powerful language understanding, knowledge learning, and complex reasoning capabilities, have become a crucial foundation for achieving Artificial Intelligence (AI) and are widely applied in various natural language processing scenarios. Agents built upon LLMs possess core capabilities for autonomously perceiving the environment, planning and making decisions, and executing interactions, enabling them to break free from the passive response mode of traditional AI and autonomously complete tasks. To further enhance the ability to solve complex tasks, agent collaboration models have emerged. Through interactive discussions among multiple agents, simulating human collaborative decision-making processes, this effectively compensates for the shortcomings of single agents in terms of knowledge coverage and reasoning depth, becoming one of the important development directions for current agent technology.

[0003] In multi-agent collaborative discussions, convergence judgment is often rather coarse, mostly only executing a fixed number of rounds, or simply using keyword matching to determine whether to end the discussion, which affects the quality of the discussion results. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, device and medium for intelligent agent collaborative control, which performs convergence detection after each intelligent agent speaks, and can directly end the current round of discussion when convergence is triggered. After each round of discussion, the execution strategy can be determined through quality evaluation, thereby improving the accuracy of convergence judgment while controlling costs.

[0005] Specifically, this application is implemented through the following technical solution: According to a first aspect of this application, a method for intelligent agent cooperative control is provided, the method comprising: During the process of multiple agents engaging in multiple rounds of discussion on a topic, after each agent speaks, a convergence detection is performed based on the first historical message in the current round of discussion to obtain the convergence detection result. If the convergence detection result is a triggered convergence, the current round of discussion ends. In response to the detection that the current round of discussion has ended, a quality assessment is performed based on the second historical message in the current round of discussion to obtain a quality assessment result; the quality assessment result includes confidence level and suggested action. Based on the confidence level and the suggested action, an execution strategy is determined after the current round of discussion ends; the execution strategy includes continuing to the next round of discussion, entering a closed round, or ending the discussion, wherein the closed round is used to indicate the last round of discussion before ending the discussion; The process continues according to the described execution strategy until the discussion concludes, yielding the results of the discussion on the aforementioned issues.

[0006] In one optional implementation, the method further includes: Once the first agent to speak in the first round of discussion is determined, a prompt word corresponding to the agent is generated based on the agent's role, the discussion question, the quality standard, and the discussion restrictions. The prompt word is then sent to the agent, who then speaks based on the prompt word. The quality standards include at least one of the following: consensus verification standard, quantitative anchoring standard, scheme specificity standard, dimension richness standard, and exit standard. The consensus verification standard instructs agents to raise objections when their conclusions align with those of other agents. The quantitative anchoring standard instructs agents to output conclusive messages including numerical indicators. The scheme specificity standard instructs agents to output scheme messages including names, parameter ranges, and resource estimates. The dimension richness standard instructs agents to cover at least two of several preset dimensions during multi-round discussions on the issue, including technical feasibility, business logic, regulatory pathways, alternative solutions, and exit criteria. The exit standard instructs agents to determine the failure conditions and loss-mitigation decisions corresponding to the scheme messages discussed. The discussion restrictions include at least one of the following: prohibiting the repetition of opinions output by other agents, prohibiting self-introduction, prohibiting the repetition of the discussion question, providing a message length less than a preset message length, and including a question marker in the message output when there is a question about the discussion question.

[0007] In one optional implementation, the convergence detection based on the first historical message in the current round of discussion to obtain the convergence detection result includes: Detect whether a first preset number of agents output historical messages during this round of discussion; If no first preset number of agents output historical messages during this round of discussion, the convergence detection result is determined to be that convergence has not been triggered. In this round of discussion, if there is a first preset number of agents outputting historical messages, the most recent first preset number of agents outputting historical messages is determined as the first historical message. Protocol keyword matching and content similarity detection are performed on the first historical message. The protocol keyword matching is used to detect the number of messages in the first historical message that contain protocol keywords, and the content similarity detection is used to detect the intersection ratio between any two messages in the first historical message. If the number of messages containing protocol keywords in the first historical messages is greater than the second preset number, and / or the number of message pairs in the first historical messages with an intersection ratio greater than a preset intersection ratio is greater than the third preset number, the convergence detection result is determined to be triggered convergence; otherwise, the convergence detection result is determined to be not triggered convergence.

[0008] In one alternative implementation, the suggested action includes continuing to the next round of discussion, entering a closed round, or ending the discussion; The process of determining the execution strategy after the current round of discussion, based on the confidence level and the suggested action, includes: Based on the confidence level and the suggested action, determine whether the first condition or the second condition is met; the first condition is that the suggested action is to end the discussion and the confidence level is greater than the first confidence level; the second condition is that the suggested action is to enter a closed loop and the confidence level is greater than the second confidence level, and the first confidence level is greater than the second confidence level. If the first condition is met, the execution strategy after the end of this round of discussion is to end the discussion; If the second condition is met, the execution strategy after the end of this round of discussion is to enter the closed loop. If neither the first nor the second condition is met, the execution strategy after the end of this round of discussion is to continue to the next round of discussion.

[0009] In one optional implementation, the step of responding to detecting the end of the current discussion round by performing a quality assessment based on second historical messages from the current discussion round to obtain a quality assessment result includes: In response to the detection that the current round of discussion has ended, determine the next round of discussion; If the current round of discussion is greater than or equal to the minimum number of rounds and less than or equal to the maximum number of rounds, a quality assessment is performed based on the second historical message in the current round of discussion to obtain the quality assessment result.

[0010] In one optional implementation, after each agent speaks, a convergence detection is performed based on the first historical message in the current round of discussion to obtain a convergence detection result, including: After each agent speaks, a divergence detection is performed based on the third historical message in the current round of discussion to obtain the divergence detection result. If the divergence detection result indicates that there is no divergence, a convergence detection is performed based on the first historical message in this round of discussion to obtain a convergence detection result.

[0011] In one optional implementation, the method further includes: If the disagreement detection result indicates that a disagreement exists, a debate mode is triggered to identify the disagreeing agents and the topics of disagreement. When it is determined that the divergent agent is about to speak, a prompt word corresponding to the divergent agent is generated based on the divergent topic, and the prompt word corresponding to the divergent agent is sent to the divergent agent; the prompt word corresponding to the divergent agent is used to instruct the divergent agent to engage in debate; The debate mode ends when each of the dissenting agents has finished speaking.

[0012] In one optional implementation, the method further includes: If any agent's output message includes a question marker, control the discussion mode to pause and display a question message for the user; In response to receiving a reply message from the user, the reply message is added to the history messages and the discussion mode is restored.

[0013] In one optional implementation, the method further includes: Upon receiving a user interruption message, pause the discussion mode and identify the agent whose speech was interrupted. The system controls the interrupted agent to output a reply to the user's interrupt message, restores the discussion mode, and controls the interrupted agent to speak again.

[0014] In one optional implementation, the method further includes: For each agent other than the one who has spoken in the previous round of discussion, determine the topic relevance score, speaking interval score, and debate context score for that agent; the topic relevance score is determined based on the matching degree between the most recent historical message and the agent; the speaking interval score is determined based on the time of the agent's last speech; the debate context score is determined based on whether the agent is a divergent agent and the agent's role style; The target score for the agent is determined based on the topic relevance score, speaking interval score, and debate context score corresponding to the agent. The agent with the highest target score is selected as the agent to speak.

[0015] According to a second aspect of this application, an intelligent agent cooperative control device is provided, the device comprising: The convergence detection module is used to perform convergence detection based on the first historical message in the current round of discussion after each agent speaks during multiple rounds of discussion on a topic, and to obtain the convergence detection result. The discussion end module is used to control the end of this round of discussion if the convergence detection result is a trigger convergence. The quality assessment module is used to perform a quality assessment based on the second historical messages in the current round of discussion in response to the detection that the current round of discussion has ended, and to obtain a quality assessment result; the quality assessment result includes confidence level and suggested action; The strategy determination module is used to determine the execution strategy after the current round of discussion based on the confidence level and the suggested action; the execution strategy includes continuing to the next round of discussion, entering a closed round, or ending the discussion, wherein the closed round is used to indicate the last round of discussion before ending the discussion; The strategy execution module is used to process according to the execution strategy until the discussion ends and obtain the discussion results for the discussion issue.

[0016] According to a third aspect of this application, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the intelligent agent cooperative control method described in the first aspect above.

[0017] According to a fourth aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent agent cooperative control method described in the first aspect above.

[0018] The agent collaboration control method, apparatus, device, and medium provided in this application embodiment, in the process of multiple agents conducting multi-round discussions on a discussion issue, performs convergence detection after each agent speaks, and directly ends the current round of discussion upon triggering convergence. After each round of discussion, an execution strategy can be determined through quality assessment, and the process is carried out according to the execution strategy until the discussion result is obtained. Compared with the coarse convergence judgment method that relies solely on a fixed number of rounds or simple keyword matching, it can achieve refined convergence judgment and comprehensive evaluation of content quality in the discussion process, effectively improving the accuracy and rationality of convergence judgment, reducing invalid and redundant discussions and the consumption of computing power and time resources, while taking into account the cost and efficiency of agent collaboration control.

[0019] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this disclosure.

[0020] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present application of a method for intelligent agent cooperative control; Figure 2 This is a schematic diagram illustrating a disagreement detection and debate triggering process according to an exemplary embodiment of this application; Figure 3 This is one of the schematic diagrams of an intelligent agent cooperative control device shown in an exemplary embodiment of this application; Figure 4 This is a second schematic diagram of an intelligent agent cooperative control device shown in an exemplary embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer device shown in an exemplary embodiment of this application. Detailed Implementation

[0022] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0023] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0024] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0025] Research has revealed that agent collaboration patterns mainly fall into the following categories: (1) Fixed-stage (Waterfall) approach: Multi-agent collaboration is divided into a fixed sequence of stages (e.g., requirements analysis → design → coding → testing), with each stage performed by an agent with a specific role. The limitation of this approach is that the stage division is rigid and cannot adapt to new dimensions or unexpected disagreements that emerge during the discussion; when the discussion has already produced high-quality conclusions in the early stages, the system is still forced to execute subsequent stages, resulting in a waste of computing resources and time. (2) Hierarchical management approach: An LLM with a manager role is introduced to dynamically allocate tasks and determine the direction of the dialogue. The limitation of this approach is that the manager LLM itself requires a lot of reasoning resources, and its decision quality is limited by the judgment ability of a single model; the manager may become a single point of failure and a performance bottleneck. (3) Free dialogue approach: Allows agents to have relatively free and organic dialogue. Although this approach avoids rigid stage division, it lacks structured convergence detection and it is difficult to determine when the discussion has reached a "sufficiently sufficient" level. (4) Academic research style, the Adaptive Stability Detection method proposes a convergence detection based on mathematical indicators, but its design goal is to evaluate the correctness of positions in the debate, rather than to judge the sufficiency and quality of multi-dimensional discussions.

[0026] In multi-agent collaborative discussions, significant shortcomings remain in the technology, with the issue of coarse convergence judgment being particularly prominent, affecting the quality of the discussion results. The methods used by multi-agent agents to determine whether a discussion has converged, whether the task objective has been achieved, and to terminate the discussion are generally quite simplistic. Most methods simply execute a fixed number of rounds, forcibly terminating the discussion after the set number of rounds, regardless of whether the discussion content achieves the expected results or whether there is redundancy or repetition. This easily leads to insufficient discussion or wasted resources. Alternatively, they rely solely on simple keywords to determine whether to end the discussion, such as only checking whether the discussion content contains keywords like "completed" or "end," resulting in low accuracy.

[0027] Based on the above research, this application provides an agent collaborative control method. In the process of multiple agents conducting multiple rounds of discussion on a problem, convergence detection is performed after each agent speaks. When convergence is triggered, the current round of discussion can be ended directly. After each round of discussion, the execution strategy can be determined through quality evaluation, thereby improving the accuracy of convergence judgment while controlling costs.

[0028] To facilitate understanding of this embodiment, a detailed description of the intelligent agent cooperative control method disclosed in this application embodiment is provided first. The executing entity of the intelligent agent cooperative control method provided in this application embodiment is generally a computer device with a certain computing power. The computer device can be a server, which can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data, and artificial intelligence platforms. In some possible implementations, the computer device can also be a terminal device, which can be a mobile device, terminal, handheld device, computing device, vehicle-mounted device, etc. In other implementations, the intelligent agent cooperative control method can be applied to an implementation environment composed of a terminal device and a server. Furthermore, the intelligent agent cooperative control method can also be implemented by a processor calling computer-readable instructions stored in memory.

[0029] The following description, in conjunction with the accompanying drawings, illustrates an agent cooperative control method provided in an embodiment of this application.

[0030] See Figure 1 The diagram shown is a flowchart illustrating an exemplary embodiment of this application of a method for intelligent agent cooperative control. Figure 1 As shown in the figure, the intelligent agent cooperative control method provided in this embodiment includes steps S101 to S105, wherein: S101: During the process of multiple agents conducting multiple rounds of discussion on the issue, after each agent speaks, convergence detection is performed based on the first historical message in this round of discussion to obtain the convergence detection result.

[0031] Here, before initiating a discussion, the user-inputted discussion question can be obtained, and multiple agents can be identified to participate in the discussion. These agents can be selected from a pool of pre-defined agents in response to the user's selection, or automatically selected based on the discussion question if not specified by the user. Each pre-defined agent has its own role and personality description. The agents can conduct a structured discussion around the discussion question, which may be an issue requiring in-depth exploration from multiple perspectives, such as research project review, technical solution demonstration, or investment decision analysis.

[0032] Before the first round of discussion begins, the first speaker can be determined from among the multiple agents. Similarly, when it is determined that an agent needs to speak, the agent to speak can also be determined from among the multiple agents.

[0033] In some possible implementations, the method further includes: For each agent other than the one who has spoken in the previous round of discussion, determine the topic relevance score, speaking interval score, and debate context score for that agent; the topic relevance score is determined based on the matching degree between the most recent historical message and the agent; the speaking interval score is determined based on the time of the agent's last speech; the debate context score is determined based on whether the agent is a divergent agent and the agent's role style; The target score for the agent is determined based on the topic relevance score, speaking interval score, and debate context score corresponding to the agent. The agent with the highest target score is selected as the agent to speak.

[0034] In the above steps, when determining the agent to speak on behalf of another, the agent who just spoke in the previous round of discussion is excluded, thereby avoiding continuous speaking by the same agent and improving the comprehensiveness and richness of multi-agent discussions. For each agent other than the one who spoke in the previous round of discussion, the most recent historical message is determined. Here, if the first speaker in the first round of discussion is identified, the most recent historical message includes the discussion question; otherwise, the most recent historical message includes the previous non-system message. The previous non-system message can be from the current round of discussion or from the previous round. The non-system message refers to a message not output by the discussion engine master editor, that is, the non-system message includes messages input by the user and messages output by the agent. The discussion engine master editor is used for lifecycle management, round control, convergence decision-making, and discussion result generation in multi-agent collaborative discussions.

[0035] The most recent historical message is matched with the agent's domain description and personality description for keywords. A topic relevance score is determined based on whether the keywords from the most recent historical message appear in the domain description and personality description. For example, three points are added for each keyword match with the domain description, and one point is added for each keyword match with the personality description. These scores are accumulated to obtain the topic relevance score.

[0036] The time of the agent's last speech is determined. Based on whether the agent's last speech is included in the most recent fourth preset number of messages, a speech interval score is determined for the agent. The specific value of the fourth preset number depends on the actual selection of speaking agents and is not limited here. For example, the fourth preset number is 4. If the agent's speech is not present in the most recent 4 messages, two points are added to the agent's score to obtain the speech interval score.

[0037] The system determines whether the agent is a divergent agent and detects whether the agent's role style includes a critical style. Based on the determination and detection results, a debate context score is determined for the agent. For example, if the agent is a divergent agent, three points are added; if the agent's role style includes a critical style, three points are added. The scores can be accumulated to obtain the debate context score.

[0038] The topic relevance score, speaking interval score, and debate context score corresponding to the agent are summed to obtain the target score for the agent. The agent with the highest target score is selected as the agent to speak. If there are multiple agents with the highest target scores, the agent ranked first among them in a preset speaking order is selected as the agent to speak. The preset speaking order can be user-specified or determined based on discussion needs.

[0039] In this way, by comprehensively considering multiple dimensions such as topic relevance, speaking intervals, and debate context, the agent to speak is determined from other agents besides the one that has already spoken in the previous round. Compared with the traditional random polling or fixed-order agent speaking method, this method can achieve refined selection of multiple agent speakers. This results in a higher degree of matching between the determined agents to speak and the current discussion topic, a more reasonable arrangement of speaking time, and a better fit with the roles and disagreements in the debate scenario. This effectively improves the coherence of the agent collaboration process, the focus of the topic, and the logic of interaction, reduces invalid speeches and redundant expressions that deviate from the topic, and improves the overall discussion progress efficiency and content quality.

[0040] Here, after determining the agent to speak, a prompt word corresponding to the agent is generated before the agent speaks, that is, the prompt word is dynamically constructed before each speech.

[0041] In traditional agent collaboration, for fixed-stage processes, transitions between stages are triggered by preset rules rather than based on the actual quality of the discussion content; for hierarchical management, the manager's control logic remains based on the process's "what to do next," rather than on quality's "is the discussion good enough"; for free-flowing dialogue, there is a lack of clear quality standard injection mechanisms, and discussions may remain superficial for a long time without in-depth exploration. It is evident that traditional process-driven architectures have low flexibility. Using process-driven rather than quality-driven approaches only tells the agent "which stage we are in now and what to do," rather than "what constitutes a good discussion." This violates the characteristic of large language models, which excel at autonomous decision-making under open constraints. In this embodiment, descriptive quality standards replace imperative process control. Instead of issuing instructions like "now execute the exploration phase" or "now begin convergence," a discussion orchestration paradigm is proposed, replacing state machines with prompts based on quality standards and discussion constraints. This allows the agent to autonomously decide its speaking content and manner based on "what a high-quality discussion should be like," rather than passively executing preset stage instructions.

[0042] Specifically, the method further includes: Once the first agent to speak in the first round of discussion is determined, a prompt word corresponding to the agent is generated based on the agent's role, the discussion question, the quality standard, and the discussion restrictions. The prompt word is then sent to the agent, who then speaks based on the prompt word. The quality standards include at least one of the following: consensus verification standard, quantitative anchoring standard, scheme specificity standard, dimension richness standard, and exit standard. The consensus verification standard instructs agents to raise objections when their conclusions align with those of other agents. The quantitative anchoring standard instructs agents to output conclusive messages including numerical indicators. The scheme specificity standard instructs agents to output scheme messages including names, parameter ranges, and resource estimates. The dimension richness standard instructs agents to cover at least two of several preset dimensions during multi-round discussions on the issue, including technical feasibility, business logic, regulatory pathways, alternative solutions, and exit criteria. The exit standard instructs agents to determine the failure conditions and loss-mitigation decisions corresponding to the scheme messages discussed. The discussion restrictions include at least one of the following: prohibiting the repetition of opinions output by other agents, prohibiting self-introduction, prohibiting the repetition of the discussion question, providing a message length less than a preset message length, and including a question marker in the message output when there is a question about the discussion question.

[0043] The agent's role tells it who it is, for example, "You are Einstein, an expert in physics." The discussion question tells the agent the core issue to be discussed. The quality standards and discussion restrictions are called injection standards, adhering to the Standards Inject, Standards Uphold (SiSu) principle. Optionally, the prompt corresponding to the consensus testing standard is specifically "Consensus will be tested, not accepted; if multiple people reach similar conclusions, it indicates a possible collective blind spot," thus guiding the agent to actively question when it finds itself agreeing with others' views, rather than simply conforming.

[0044] The specific prompt for the quantitative anchoring standard is: "Every statement (i.e., conclusive message) must have a quantitative anchor, requiring outputs to include specific quantitative indicators, rather than vague qualitative descriptions." For example, "good results" is not a conclusion; "Kd < 5nM is." This requires agents to provide specific numerical indicators, preventing discussions from remaining at a vague qualitative level.

[0045] The specific prompt for the Solution Specificity standard is that "Solution-level recommendations (i.e., solution messages) include specific names, parameter ranges, and resource estimates; otherwise, they are just empty talk," thus ensuring that the output of the discussion is actionable rather than vague.

[0046] The specific prompt for the Dimensional Richness criterion is: "The value of a discussion comes from the richness of its dimensions, including technical feasibility, business logic, regulatory pathways, alternatives, and exit criteria. A discussion is incomplete when there are fewer than two dimensions." This guides the agent to cover multiple evaluation dimensions in its discussions. Here, the agent can determine the dimensions to be discussed in each round of discussion independently, rather than pre-setting fixed dimensions.

[0047] The specific prompt for the exit criteria (Kill Criteria) is "Exit criteria are as important as the solution itself; projects without pre-set stop-loss decisions are not worth starting," thus requiring the agent to clearly discuss the failure conditions and stop-loss decisions for the solution. The failure conditions are predefined, quantifiable indicators; when these indicators are met, the solution is considered to have failed. The stop-loss decisions are adjustment decisions triggered by the failure conditions. For example, for a drug project, the failure condition is that the affinity does not reach Kd < 10nM within 6 months; the stop-loss decision is to terminate the candidate molecule and switch to an alternative solution. For an AI chip project, the failure condition is that the inference latency is still > 50ms after 3 rounds of selection; the stop-loss decision is to stop self-development and purchase a third-party solution. For an investment project, the failure condition is that the cumulative loss reaches 38% of the invested capital; the stop-loss decision is to force liquidation to stop the loss. For clinical trials, the failure condition is that the interim analysis efficacy rate is less than 20%; the stop-loss decision is to terminate the trial and not proceed to the next phase.

[0048] The specific value of the preset message length depends on the actual needs of intelligent agent collaboration and is not limited here. For example, the preset message length is 500 characters. The question marker is, for example, "[Questioning User]" or "[ASK_USER]".

[0049] It is understood that the role of the agent, the discussion issue, the quality standards, and the discussion restrictions are injected only once in the prompts of the first agent to speak in the first round of discussion. This content is remembered by each agent through the context mechanism of LLM, so there is no need to inject it repeatedly in the prompts of subsequent agents.

[0050] In this way, when determining the prompt word for the first agent to speak in the first round of discussion, quality standards and discussion constraints are introduced, thereby transforming the traditional engine-instruction-based control into a quality standard-guided approach. This fully leverages the autonomous reasoning capabilities of the large language model under open constraints, allowing the agent to autonomously decide its speaking strategy. Compared to the traditional fixed-stage orchestration control method, this reduces rigid constraints and significantly improves architectural flexibility. At the same time, this approach enables discussions to autonomously adapt their focus based on the type of discussion issue. For example, technical discussions can focus on quantitative anchors and the specificity of solutions, while strategic discussions can focus on dimensional richness and exit criteria, covering a wide range of discussion issues without modifying the engine orchestration logic.

[0051] In some possible implementations, for each agent other than the first agent to speak in the initial round of discussion, a prompt word corresponding to that agent is generated based on the discussion context and response method. The response method is a pre-defined format. This prompt word guides the agent to continue the discussion.

[0052] Multiple agents are controlled to perform multi-round adaptive discussions, in which each agent speaks in turn during each round.

[0053] In some possible implementations, the convergence detection based on the first historical message in this round of discussion to obtain the convergence detection result includes: Detect whether a first preset number of agents output historical messages during this round of discussion; If no first preset number of agents output historical messages during this round of discussion, the convergence detection result is determined to be that convergence has not been triggered. In this round of discussion, if there is a first preset number of agents outputting historical messages, the most recent first preset number of agents outputting historical messages is determined as the first historical message. Protocol keyword matching and content similarity detection are performed on the first historical message. The protocol keyword matching is used to detect the number of messages in the first historical message that contain protocol keywords, and the content similarity detection is used to detect the intersection ratio between any two messages in the first historical message. If the number of messages containing protocol keywords in the first historical messages is greater than the second preset number, and / or the number of message pairs in the first historical messages with an intersection ratio greater than a preset intersection ratio is greater than the third preset number, the convergence detection result is determined to be triggered convergence; otherwise, the convergence detection result is determined to be not triggered convergence.

[0054] In the above steps, it is detected whether a first preset number of agent output history messages exist in the current round of discussion. The specific value of the first preset number depends on the actual convergence detection needs and is not limited here. For example, the first preset number is 4, and it is detected whether there are 4 agent output history messages in the current round of discussion. If there are no first preset number of agent output history messages in the current round of discussion, the convergence detection result is determined to be that convergence has not been triggered. If there are first preset number of agent output history messages in the current round of discussion, the most recent first preset number of agent output history messages are determined as the first history message, and two checks are performed on the first history message: protocol keyword matching and content similarity detection.

[0055] For protocol keyword matching, the first historical messages are matched for protocol keywords to determine whether each message in the first historical messages contains a protocol keyword. Here, the protocol keyword is a word indicating agreement. For example, the protocol keyword includes words such as agree, agree, completely agree, approve, are right, I am convinced, unanimously believe, consensus, agree, convinced, exactly, well said, etc. If the number of messages containing the protocol keyword in the first historical messages is greater than a second preset number, it is determined that there is a duplicate consensus state. The specific value of the second preset number depends on the actual convergence detection needs and the value of the first preset number, and is not limited here. For example, the second preset number is 2, that is, at least 3 out of the most recent 4 messages contain the protocol keyword, thus determining that there is a duplicate consensus.

[0056] For content similarity detection, a fifth preset number of characters are extracted from each message in the first historical messages to construct a word set. Here, for each message in the first historical messages, the extracted fifth preset number of characters can be located at any position such as the beginning, middle, or end of the message. They can be extracted continuously or obtained by splicing them together. The specific value of the fifth preset number is determined according to the actual convergence detection needs and the preset message length, and is not limited here. For example, the fifth preset number is 50. Based on the word set, the intersection ratio between any two messages is determined by methods such as regular expression matching and cosine similarity. The intersection ratio can be determined by the following formula (1): (1) in, Indicates the intersection ratio. This represents the number of words included in the intersection of the word sets corresponding to two messages. and These represent the number of words in the word sets corresponding to the two messages. This indicates taking the minimum value.

[0057] The specific value of the preset intersection ratio depends on the actual convergence detection needs and is not limited here. For example, the preset intersection ratio is 0.5.

[0058] If at least one of the following conditions is met: the number of messages containing protocol keywords in the first historical messages is greater than a second preset number, and the number of message pairs in the first historical messages with an intersection ratio greater than a preset intersection ratio is greater than a third preset number, then the convergence detection result can be determined as triggered convergence; otherwise, the convergence detection result is determined as not triggered convergence. The specific value of the third preset number depends on the actual convergence detection needs and the value of the first preset number, and is not limited here. For example, the third preset number is 1, that is, at least 2 pairs of messages in the most recent 4 messages have an intersection ratio greater than 0.5, thus determining that duplicate consensus exists.

[0059] In this way, when performing convergence detection, protocol keyword matching and content similarity detection are used, which does not require calling a large language model, resulting in extremely low cost and fast response speed. Detection can be completed in milliseconds, effectively reducing computing power consumption. At the same time, compared with the traditional coarse judgment method of fixed number of rounds and simple keyword matching, the embodiments of this disclosure can accurately capture the consensus content and content relevance in multi-agent discussions, adapting to the core requirement of "convergence when the discussion quality is sufficient" in multi-dimensional open discussions. This effectively improves the accuracy and rationality of convergence judgment, avoiding insufficient content due to premature termination of discussion and preventing resource waste caused by redundant discussions, thus balancing discussion efficiency and content quality.

[0060] In traditional agent collaboration, the lack of disagreement detection and debate triggering mechanisms for free-flowing dialogue makes it difficult to effectively break groupthink when agents prematurely reach a superficial consensus. It is evident that due to the lack of proactive disagreement mechanisms, when multiple agents tend to reach agreement quickly due to sharing similar training data, it is difficult to detect and break such false consensus. However, this embodiment implements an automated disagreement detection and debate triggering mechanism. When a disagreement is detected between agents, a debate mode is automatically triggered, ensuring that different perspectives are fully explored.

[0061] In some possible implementations, after each agent speaks, a convergence detection is performed based on the first historical messages in the current round of discussion to obtain a convergence detection result, including: After each agent speaks, a divergence detection is performed based on the third historical message in the current round of discussion to obtain the divergence detection result. If the divergence detection result indicates that there is no divergence, a convergence detection is performed based on the first historical message in this round of discussion to obtain a convergence detection result.

[0062] In the above steps, after each agent speaks, a divergence detection is performed based on the third historical message in the current round of discussion to obtain the divergence detection result.

[0063] When performing divergence detection based on the third historical message in the current round of discussion, specifically, it checks whether there are a sixth preset number of agent output historical messages in the current round of discussion; if there are no sixth preset number of agent output historical messages in the current round of discussion, the divergence detection result is determined to be no divergence; if there are a sixth preset number of agent output historical messages in the current round of discussion, the most recent sixth preset number of agent output historical messages is determined as the third historical message, and divergence keyword matching is performed on the third historical message; if there are divergence keywords in the third historical message, the divergence detection result is determined to be divergence; if there are no divergence keywords in the third historical message, the divergence detection result is determined to be no divergence.

[0064] The specific value of the sixth preset quantity depends on the actual needs of divergence detection and is not limited here. For example, the sixth preset quantity is 3, which detects whether there are 3 agent output historical messages in the current round of discussion.

[0065] Here, the "disagreement keywords" are words that express disagreement. For example, these disagreement keywords include: disagree, I think not, but I feel, I'm afraid not, I have a different opinion, this might be problematic, needs discussion, not quite agree, needs to be discussed, I disagree, I don't think, however, but I believe, that's not quite right, I have concerns, problemmatic, etc.

[0066] Only if the divergence detection result is that there is no divergence will a convergence detection be performed based on the first historical message in this round of discussion to obtain a convergence detection result.

[0067] In this way, after each agent speaks, a divergence detection is performed first. Only if there is no divergence is found is a convergence detection performed. This avoids blindly performing convergence detection when agents still have opposing views or cognitive differences, reducing the possibility of prematurely determining that the discussion has converged and forcibly terminating the discussion. It ensures that the convergence judgment process only begins when the views of all agents tend to be consistent and there are no disagreements. This makes the timing of convergence detection more reasonable and rigorous, effectively improving the accuracy and comprehensiveness of agent collaborative control, and enhancing the integrity and reliability of the discussion results.

[0068] In some possible implementations, the method further includes: If the disagreement detection result indicates that a disagreement exists, a debate mode is triggered to identify the disagreeing agents and the topics of disagreement. When it is determined that the divergent agent is about to speak, a prompt word corresponding to the divergent agent is generated based on the divergent topic, and the prompt word corresponding to the divergent agent is sent to the divergent agent; the prompt word corresponding to the divergent agent is used to instruct the divergent agent to engage in debate; The debate mode ends when each of the dissenting agents has finished speaking.

[0069] In the above steps, if the disagreement detection result indicates a disagreement exists, a debate mode is triggered. The disagreeing agents and the disagreement topics are identified, and the disagreeing agents and topics are written to the pending Disagreement state. Here, the number of disagreeing agents is at least two. If the agent to speak is determined to be one of the disagreeing agents, the pending Disagreement state can be detected. Based on the disagreement topic, a prompt word corresponding to the disagreeing agent is generated. This injects a debate instruction when generating the prompt word corresponding to the disagreeing agent, guiding the agent to articulate its position using specific reasoning, evidence, or examples. For example, the prompt word corresponding to the disagreeing agent might be: "A disagreement has arisen in the discussion… Please elaborate on your position using specific reasoning, evidence, or examples. Respond to the other party's viewpoint directly but respectfully."

[0070] Here, the debate will not interrupt the current discussion round. That is, the disagreement detection occurs within the current discussion round, and the debate is also carried out within the current discussion round. There will be no restart of the discussion round.

[0071] The debate topic (debateAddressedBy) array tracks the speaking agents in each opposing viewpoint. Once each agent has finished speaking, the pendingDisagreement state is cleared, the debate mode is exited, and the discussion mode is restored. Subsequent speaking agents resume using prompts generated from standard speaking commands.

[0072] It's understandable that exiting debate mode and resuming discussion mode allows the agent to still see previous debate content in context during subsequent speeches and choose to delve deeper. Debate commands are one-time guidance, not continuous control.

[0073] In this way, when the disagreement detection result indicates the existence of a disagreement, the debate mode is automatically triggered, the disagreeing agents and the disagreement topics are identified, and exclusive debate prompts are generated for the disagreeing agents waiting to speak based on the disagreement topics. This guides the disagreeing agents to conduct targeted debates around the points of contention, ensuring that different viewpoints are fully developed rather than being covered up by superficial polite agreement. The debate mode is exited only after all the disagreeing agents have finished speaking. This effectively reduces the problem that large language models tend to reach a consensus quickly due to similar training data, and improves the depth and diversity of multi-agent discussions.

[0074] To more clearly illustrate the process of disagreement detection and debate triggering, see [link to relevant documentation]. Figure 2 This is a schematic diagram illustrating a disagreement detection and debate triggering process, as shown in an exemplary embodiment of this application. Figure 2 As shown, if the disagreement detection result indicates a disagreement, a debate mode is triggered. The disagreeing agents and their topics are identified. While the disagreeing agents are waiting to speak, a prompt word corresponding to them is generated. The mode tracks the disagreeing agents who have already spoken. The debate mode exits after all disagreeing agents have finished speaking. Specific steps are described in the aforementioned embodiment and will not be repeated here.

[0075] S102: If the convergence detection result is a triggered convergence, control the end of this round of discussion.

[0076] In this step, if the convergence detection result is determined to be a convergence trigger, the discussion engine's main editor can be controlled to issue a moderator intervention event, skipping the remaining speakers in the current round and directly controlling the end of the current round of discussion.

[0077] Optionally, if the convergence detection result is determined to be that convergence has not been triggered, the current round of discussion can continue until the current round of discussion ends.

[0078] S103: In response to detecting the end of the current discussion, a quality assessment is performed based on the second historical messages in the current discussion to obtain a quality assessment result; the quality assessment result includes confidence level and suggested action.

[0079] In this step, in response to the detection that the current round of discussion has ended, the second historical message from the current round of discussion is collected, and a quality assessment is performed based on the second historical message to obtain the quality assessment result.

[0080] Optionally, the end of this round of discussion can be determined by the following steps: If the convergence detection result indicates that convergence has been triggered, then this round of discussion is considered complete. Alternatively, the number of non-system historical messages in the current discussion can be detected, and if the number of non-system historical messages in the current discussion reaches the first preset number of messages, the current discussion can be determined to end. Alternatively, detect the number of non-system historical messages in the current discussion. If the number of non-system historical messages in the current discussion reaches the second preset number of messages, determine whether the current discussion has ended based on the message content of the non-system historical messages in the current discussion.

[0081] The specific values ​​of the first preset message quantity and the second preset message quantity are determined based on the number of collaborating agents and the actual needs of agent collaboration control, and are not limited here.

[0082] Here, the second historical message includes a greater number of messages than the first historical message. The specific number of messages included in the second historical message depends on the actual needs of intelligent agent collaborative control and is not limited here. For example, the second historical message includes the 8 most recent non-system messages.

[0083] When conducting a quality assessment based on the second historical messages discussed in this round, and obtaining the quality assessment results, specifically, a seventh preset number of characters are extracted from each message in the second historical messages to form transcribed text; judgment prompts are generated, and the transcribed text and the judgment prompts are input into the quality assessment model to obtain the quality assessment results output by the quality assessment model.

[0084] The specific value of the seventh preset quantity is determined based on the actual quality assessment needs and the preset message length, and is not limited here. For example, the seventh preset quantity is 200.

[0085] The judgment prompts include role settings, quality assessment rules, and output format. The role settings indicate that the quality assessment model is a discussion quality assessor. Optionally, the quality assessment rules include: if the current discussion introduces a new dimension compared to previous discussions, the suggested action is to continue to the next round of discussion; if the current discussion repeats existing arguments and the discussion issue has been covered, the suggested action is to enter a closed round; if the current discussion has produced an actionable conclusion, exit criteria, or next step plan, the suggested action is to end the discussion. The output format, for example, requires the quality assessment results to be output in a preset format, such as JSON format, and the output results include two fields: confidence level and suggested action.

[0086] Here, the inference cost of the quality assessment model is lower than that of the agent being discussed. For example, the agent could be the Claude Opus series, the GPT-4 series, etc., and the quality assessment model could be DeepSeek Chat, Claude Haiku, GPT-4o Mini, Gemini Flash, Qwen Plus, etc.

[0087] For example, the quality assessment model uses transcribed text truncated to 200 characters and an output of up to 200 tokens, with each execution costing approximately 1 / 20th of an agent's speech. Compared to using an agent for quality assessment, the overall LLM call cost is reduced by approximately 95%. Thus, relatively simple but frequent quality assessments are handled through quality assessment lines, while collaborative discussions requiring deep reasoning are handled through agents, thereby controlling costs while ensuring discussion quality.

[0088] The quality assessment results include confidence level and recommended actions. The confidence level, ranging from 0 to 100, indicates a degree of certainty that "the discussion has been sufficient." Recommended actions include continuing to the next round of discussion, entering a closed round, or ending the discussion. Here, the closed round refers to the final round of discussion, i.e., the transition from normal discussion to the end of the discussion.

[0089] In some possible implementations, if the current round of discussion ends due to convergence, the number of non-system historical messages in the current round of discussion may be less than the number of messages included in the second historical message. In this case, all non-system historical messages in the current round of discussion can be identified as the second historical message.

[0090] In some possible implementations, the response to detecting the end of the current round of discussion, performing a quality assessment based on the second historical messages in the current round of discussion, and obtaining a quality assessment result, includes: In response to the detection that the current round of discussion has ended, determine the next round of discussion; If the current round of discussion is greater than or equal to the minimum number of rounds and less than or equal to the maximum number of rounds, a quality assessment is performed based on the second historical message in the current round of discussion to obtain the quality assessment result.

[0091] In the above steps, in response to detecting the end of the current discussion round, the number of discussion rounds is determined. This number of discussion rounds is compared with both the minimum and maximum number of rounds. If the number of discussion rounds is less than the minimum number of rounds, the execution strategy after the end of this round is determined to be to continue to the next round, thus ensuring that each round of discussion has at least enough rounds for each agent to express their views. If the number of discussion rounds is greater than the maximum number of rounds, the execution strategy after the end of this round is determined to be to end the discussion; only when the number of discussion rounds is greater than or equal to the minimum number of rounds and less than or equal to the maximum number of rounds is a quality assessment performed.

[0092] The specific values ​​of the minimum and maximum number of rounds are determined according to the actual needs of intelligent agent collaborative control, and are not limited here. For example, the minimum number of rounds is 2, and the maximum number of rounds is 8.

[0093] In this way, after the current round of discussion ends, the current discussion round is first determined. Quality assessment is only conducted if the current discussion round falls within the range of the minimum and maximum number of rounds. By setting round range admission conditions to pre-constrain the quality assessment, we can reduce the situation where the discussion rounds are too few, resulting in insufficient development of viewpoints and inadequate assessment basis, leading to distorted quality assessments. At the same time, we can also reduce the situation where too many rounds result in excessive redundant content, leading to ineffective assessments and wasted computing power. This ensures that quality assessment is only conducted when the discussion process has reached a reasonable level, guaranteeing that the quality assessment has sufficient information and a valid assessment basis, and significantly improving the accuracy and rationality of the quality assessment results.

[0094] S104: Based on the confidence level and the suggested action, determine the execution strategy after the end of this round of discussion; the execution strategy includes continuing to the next round of discussion, entering a closed round, or ending the discussion, wherein the closed round is used to indicate the last round of discussion before ending the discussion.

[0095] In this step, the execution strategy after the end of this round of discussion can be determined comprehensively based on the confidence level and the suggested actions.

[0096] In some possible implementations, the suggested actions include continuing to the next round of discussion, entering a closed round, or ending the discussion; The process of determining the execution strategy after the current round of discussion, based on the confidence level and the suggested action, includes: Based on the confidence level and the suggested action, determine whether the first condition or the second condition is met; the first condition is that the suggested action is to end the discussion and the confidence level is greater than the first confidence level; the second condition is that the suggested action is to enter a closed loop and the confidence level is greater than the second confidence level, and the first confidence level is greater than the second confidence level. If the first condition is met, the execution strategy after the end of this round of discussion is to end the discussion; If the second condition is met, the execution strategy after the end of this round of discussion is to enter the closed loop. If neither the first nor the second condition is met, the execution strategy after the end of this round of discussion is to continue to the next round of discussion.

[0097] In the above steps, if the suggested action is to end the discussion and the confidence level is greater than the first confidence level, the execution strategy after the end of this round of discussion is determined to be to end the discussion. If the suggested action is to enter a closed loop and the confidence level is greater than the second confidence level, the execution strategy after the end of this round of discussion is determined to be to enter a closed loop. In cases other than the suggested action being to end the discussion and the confidence level being greater than the first confidence level, or the suggested action being to enter a closed loop and the confidence level being greater than the second confidence level, the execution strategy after the end of this round of discussion is determined to be to continue to the next round of discussion.

[0098] The specific values ​​of the first confidence level and the second confidence level are determined according to the actual needs of intelligent agent collaborative control, and are not limited here. For example, the first confidence level is 85 and the second confidence level is 75.

[0099] In this way, the suggested actions are divided into three categories: continuing to the next round of discussion, entering a closed round, and ending the discussion. By combining the confidence level and judging whether the first or second condition is met, the execution strategy after the end of this round of discussion is determined. The joint judgment of the confidence level threshold grading and the suggested actions helps to improve the accuracy and flexibility of the subsequent discussion.

[0100] S105: Process according to the execution strategy until the discussion ends and the discussion results for the discussed issues are obtained.

[0101] In this step, after obtaining the execution strategy, the execution strategy is executed until the discussion ends, resulting in a discussion result for the discussed issue. The discussion result includes a structured summary.

[0102] Optionally, if the execution strategy is to enter a closed round, after entering the closed round, a closing prompt is added to the prompts before all agents speak in the closed round, thereby guiding agents to summarize core viewpoints, point out unresolved disagreements, and recommend next steps. After the closed round discussion ends, the discussion results for the discussed issues are obtained.

[0103] When the execution strategy is to end the discussion, it means that the discussion has produced actionable conclusions, exit criteria, and next steps, so the discussion can be ended directly, and the discussion results for the discussed issues are obtained.

[0104] In traditional agent collaboration, user intervention during discussions typically requires interrupting the entire process rather than seamlessly integrating into the discussion context, resulting in fragmented user involvement. However, this embodiment provides a non-disruptive user intervention interface, supporting agents to proactively ask questions and allowing users to interject at any time.

[0105] In some possible implementations, the method further includes: If any agent's output message includes a question marker, control the discussion mode to pause and display a question message for the user; In response to receiving a reply message from the user, the reply message is added to the history messages and the discussion mode is restored.

[0106] Here, during multiple rounds of discussion among multiple agents on a topic, when an agent has questions about the topic, the output message includes a question marker to ask the user a question.

[0107] Here, a question timeout can be set. After displaying a question to the user, the timeout period begins. If a reply is received from the user within the timeout period, the reply is added to the history messages and the discussion mode is restored. If no reply is received from the user within the timeout period, the discussion mode is automatically restored. If a question marker is included in subsequent messages output by the agent, the agent can proactively ask the user a question again. The specific value of the question timeout can be determined based on the actual agent collaboration control needs and is not limited here. For example, the question timeout is 180 seconds.

[0108] In some possible implementations, the method further includes: Upon receiving a user interruption message, pause the discussion mode and identify the agent whose speech was interrupted. The system controls the interrupted agent to output a reply to the user's interrupt message, restores the discussion mode, and controls the interrupted agent to speak again.

[0109] Here, during multiple rounds of discussion among various agents on a particular issue, all messages can be displayed to the user, who can interrupt at any time. The agent that interrupted the current speaker will respond to the user first. Optionally, if the user's interrupt message specifies a responding agent, the user-specified agent can be controlled to output a reply message in response to the interrupt message.

[0110] Here, an interruption timeout can be set. After the interrupted agent outputs a reply to the user's interruption message, the interruption timeout is started. If no further user interruption message is received within the timeout period, the discussion mode is restored and the interrupted agent is allowed to speak again. If a further user interruption message is received within the timeout period, the interrupted agent outputs a reply to the new message and the interruption timeout is restarted until no further message is received within the timeout period. Then, the discussion mode is restored and the interrupted agent is allowed to speak again. The specific value of the interruption timeout can be determined according to the actual needs of agent collaboration control and is not limited here. For example, the interruption timeout is 30 seconds.

[0111] Thus, through the embodiments of this disclosure, it is possible for intelligent agents to proactively ask questions of users and for users to interrupt at any time, without disrupting the contextual coherence of the discussion. User-inputted messages are naturally added to the message history, and all subsequent intelligent agents can see the user-inputted messages when constructing the speaking context. Compared to the fragmented interaction of traditional methods that require pausing, switching modes, and restarting, this non-disruptive human-computer interaction mode effectively improves the user experience and enhances the continuity of intelligent agent collaborative control.

[0112] To better understand the agent collaborative control process, an example is provided below. In this example, three agents (Agent-A: a protein engineering expert, Agent-B: an AI / machine learning expert, and Agent-C: an investment decision-making expert) discuss the topic "Feasibility of an AI-based protein drug design platform." In the first round of discussion, prompts are generated for the first agent to speak. Taking Agent-A as the first agent to speak in the first round as an example, its prompts are: "You are Agent-A, an expert in the field of protein engineering. The discussion topic is..., the quality standard is..., and the discussion restrictions are...". After Agent-A speaks, a disagreement detection is performed, and the result is no disagreement. Then Agent-B speaks, and another disagreement detection is performed. Agent-B outputs "The current prediction accuracy of AlphaFold is insufficient for drug design." After Agent-B speaks, a disagreement detection is performed, and the result is no disagreement. Agent-C speaks last, analyzing from the perspective of return on investment. After the first round of discussion, since the number of rounds in this round is less than the minimum number of rounds, the execution strategy after the end of this round of discussion is to continue to the next round of discussion.

[0113] In the second round of discussion, Agent-B's statement contained the dissenting keyword "I disagree," contradicting Agent-A's view that "protein design is mature." This triggered a debate mode, identifying Agent-A and Agent-B as the dissenting agents, with the topic being "protein design maturity assessment." When Agent-A was scheduled to speak, a debate instruction was injected into the generated prompt for Agent-A. Agent-A responded to Agent-B's challenges with specific data (Kd value, success rate statistics), meeting the "quantitative anchor" requirement in the quality standard. After Agent-B also provided a debate response, it exited the debate mode. This round of discussion did not trigger convergence. After the second round of discussion concluded, because the number of rounds exceeded the minimum but was less than the maximum, a quality assessment was conducted, determining that the execution strategy after this round was to continue to the next round of discussion.

[0114] In the third round of discussions, Agent-C proposed an exit criterion: "If the affinity of the preclinical candidate does not reach Kd < 10 nM within 6 months, it is recommended to terminate the project," echoing the "exit criterion requirement" in the quality standards. Agent-A and Agent-B expressed their agreement and supplemented the exit indicators with technical and commercial dimensions. No convergence was triggered during this round of discussions. After the third round of discussions, because the number of rounds was greater than the minimum number of rounds but less than the maximum number of rounds, a quality assessment was conducted, and it was determined that the execution strategy after this round of discussions would be to enter the closed round.

[0115] In the fourth round of discussion, all agents received a closing prompt: "This is the final round of this discussion. Please summarize your core points, identify unresolved disagreements, and provide your most important next steps." Agent-A summarized the technical roadmap and proposed a three-month milestone. Agent-B pointed out that data quality remains an unresolved disagreement. Agent-C provided an investment decision-making framework and a phased funding plan. After the closing round, the entire discussion concluded, yielding results on the discussed issues, including a discussion summary, core points, consensus reached, existing disagreements, proposed solutions, and issues requiring clarification.

[0116] The agent collaboration control method provided in this application performs convergence detection after each agent speaks during multiple rounds of discussion on a topic. Upon triggering convergence, the current round of discussion ends directly. After each round, an execution strategy is determined through quality assessment, and processing continues according to the strategy until the discussion result is obtained. Compared to coarse convergence judgment methods that rely solely on a fixed number of rounds or simple keyword matching, this method achieves refined convergence judgment and comprehensive evaluation of content quality during the discussion process. This effectively improves the accuracy and rationality of convergence judgment, reduces invalid and redundant discussions, and minimizes the consumption of computing power and time resources, while also balancing the cost and efficiency of agent collaboration control.

[0117] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0118] Corresponding to the aforementioned embodiments of the intelligent agent cooperative control method, this application also provides embodiments of the intelligent agent cooperative control device.

[0119] Please see Figure 3 and Figure 4 , Figure 3 This is one of the schematic diagrams of an intelligent agent cooperative control device provided in an embodiment of this disclosure. Figure 4 This is a second schematic diagram of an intelligent agent cooperative control device provided in an embodiment of this disclosure. Figure 3 As shown in the figure, the intelligent agent cooperative control device 300 provided in this embodiment includes: The convergence detection module 301 is used to perform convergence detection based on the first historical message in the current round of discussion after each agent speaks during multiple rounds of discussion on a discussion issue, and to obtain the convergence detection result. The discussion termination module 302 is used to control the end of this round of discussion if the convergence detection result is a trigger convergence. The quality assessment module 303 is used to perform a quality assessment based on the second historical messages in the current round of discussion in response to detecting the end of the current round of discussion, and to obtain a quality assessment result; the quality assessment result includes confidence level and suggested action; The strategy determination module 304 is used to determine the execution strategy after the current round of discussion based on the confidence level and the suggested action; the execution strategy includes continuing to the next round of discussion, entering a closed round, or ending the discussion, wherein the closed round is used to indicate the last round of discussion before ending the discussion; The strategy execution module 305 is used to process according to the execution strategy until the discussion ends and obtain the discussion results for the discussion issues.

[0120] In one alternative implementation, such as Figure 4 As shown, the intelligent agent collaborative control device 300 further includes a prompt word generation module 306, which is used for: Once the first agent to speak in the first round of discussion is determined, a prompt word corresponding to the agent is generated based on the agent's role, the discussion question, the quality standard, and the discussion restrictions. The prompt word is then sent to the agent, who then speaks based on the prompt word. The quality standards include at least one of the following: consensus verification standard, quantitative anchoring standard, scheme specificity standard, dimension richness standard, and exit standard. The consensus verification standard instructs agents to raise objections when their conclusions align with those of other agents. The quantitative anchoring standard instructs agents to output conclusive messages including numerical indicators. The scheme specificity standard instructs agents to output scheme messages including names, parameter ranges, and resource estimates. The dimension richness standard instructs agents to cover at least two of several preset dimensions during multi-round discussions on the issue, including technical feasibility, business logic, regulatory pathways, alternative solutions, and exit criteria. The exit standard instructs agents to determine the failure conditions and loss-mitigation decisions corresponding to the scheme messages discussed. The discussion restrictions include at least one of the following: prohibiting the repetition of opinions output by other agents, prohibiting self-introduction, prohibiting the repetition of the discussion question, providing a message length less than a preset message length, and including a question marker in the message output when there is a question about the discussion question.

[0121] In one optional implementation, when the convergence detection module 301 performs convergence detection based on the first historical message in the current round of discussion to obtain the convergence detection result, it is specifically used for: Detect whether a first preset number of agents output historical messages during this round of discussion; If no first preset number of agents output historical messages during this round of discussion, the convergence detection result is determined to be that convergence has not been triggered. In this round of discussion, if there is a first preset number of agents outputting historical messages, the most recent first preset number of agents outputting historical messages is determined as the first historical message. Protocol keyword matching and content similarity detection are performed on the first historical message. The protocol keyword matching is used to detect the number of messages in the first historical message that contain protocol keywords, and the content similarity detection is used to detect the intersection ratio between any two messages in the first historical message. If the number of messages containing protocol keywords in the first historical messages is greater than the second preset number, and / or the number of message pairs in the first historical messages with an intersection ratio greater than a preset intersection ratio is greater than the third preset number, the convergence detection result is determined to be triggered convergence; otherwise, the convergence detection result is determined to be not triggered convergence.

[0122] In one alternative implementation, the suggested action includes continuing to the next round of discussion, entering a closed round, or ending the discussion; The strategy determination module 304 is specifically used for: Based on the confidence level and the suggested action, determine whether the first condition or the second condition is met; the first condition is that the suggested action is to end the discussion and the confidence level is greater than the first confidence level; the second condition is that the suggested action is to enter a closed loop and the confidence level is greater than the second confidence level, and the first confidence level is greater than the second confidence level. If the first condition is met, the execution strategy after the end of this round of discussion is to end the discussion; If the second condition is met, the execution strategy after the end of this round of discussion is to enter the closed loop. If neither the first nor the second condition is met, the execution strategy after the end of this round of discussion is to continue to the next round of discussion.

[0123] In one optional implementation, the quality assessment module 303 is specifically used for: In response to the detection that the current round of discussion has ended, determine the next round of discussion; If the current round of discussion is greater than or equal to the minimum number of rounds and less than or equal to the maximum number of rounds, a quality assessment is performed based on the second historical message in the current round of discussion to obtain the quality assessment result.

[0124] In one optional implementation, when the convergence detection module 301 performs convergence detection based on the first historical message in the current round of discussion after each agent speaks, and obtains the convergence detection result, it is specifically used for: After each agent speaks, a divergence detection is performed based on the third historical message in the current round of discussion to obtain the divergence detection result. If the divergence detection result indicates that there is no divergence, a convergence detection is performed based on the first historical message in this round of discussion to obtain a convergence detection result.

[0125] In one alternative implementation, such as Figure 4 As shown, the intelligent agent collaborative control device 300 further includes a disagreement debate module 307, which is used for: If the disagreement detection result indicates that a disagreement exists, a debate mode is triggered to identify the disagreeing agents and the topics of disagreement. When it is determined that the divergent agent is about to speak, a prompt word corresponding to the divergent agent is generated based on the divergent topic, and the prompt word corresponding to the divergent agent is sent to the divergent agent; the prompt word corresponding to the divergent agent is used to instruct the divergent agent to engage in debate; The debate mode ends when each of the dissenting agents has finished speaking.

[0126] In one alternative implementation, such as Figure 4 As shown, the intelligent agent collaborative control device 300 further includes a user intervention module 308, which is used for: If any agent's output message includes a question marker, control the discussion mode to pause and display a question message for the user; In response to receiving a reply message from the user, the reply message is added to the history messages and the discussion mode is restored.

[0127] In an optional implementation, the user intervention module 308 is further configured to: Upon receiving a user interruption message, pause the discussion mode and identify the agent whose speech was interrupted. The system controls the interrupted agent to output a reply to the user's interrupt message, restores the discussion mode, and controls the interrupted agent to speak again.

[0128] In one alternative implementation, such as Figure 4 As shown, the intelligent agent collaborative control device 300 further includes a speech selection module 309, which is used for: For each agent other than the one who has spoken in the previous round of discussion, determine the topic relevance score, speaking interval score, and debate context score for that agent; the topic relevance score is determined based on the matching degree between the most recent historical message and the agent; the speaking interval score is determined based on the time of the agent's last speech; the debate context score is determined based on whether the agent is a divergent agent and the agent's role style; The target score for the agent is determined based on the topic relevance score, speaking interval score, and debate context score corresponding to the agent. The agent with the highest target score is selected as the agent to speak.

[0129] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0130] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0131] Based on the same technical concept, this application also provides a computer device 500, referring to... Figure 5 The diagram shown is a schematic representation of the structure of a computer device according to an exemplary embodiment of this application, comprising: The processor 510, memory 520, and bus 530 are included. The memory 520 is used to store execution instructions and includes main memory 521 and external memory 522. The main memory 521, also known as internal memory, is used to temporarily store the operation data in the processor 510 and the data exchanged with external memory 522 such as hard disk. The processor 510 exchanges data with external memory 522 through main memory 521.

[0132] In this embodiment, the memory 520 is specifically used to store application code that executes the solution of this application, and its execution is controlled by the processor 510. That is, when the computer device 500 is running, the processor 510 communicates with the memory 520 through the bus 530, or the processor 510 communicates with the memory 520 through other means, so that the processor 510 executes the application code stored in the memory 520, and then executes the steps of the intelligent agent cooperative control method described in any of the foregoing embodiments.

[0133] The memory 520 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0134] Processor 510 may be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0135] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the computer device 500. In other embodiments of this application, the computer device 500 may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0136] This disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the intelligent agent cooperative control method described in the above-described method embodiments. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0137] This disclosure also provides a computer program product, which stores a computer program. When the computer program is run by a processor, it executes the steps of the intelligent agent cooperative control method provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0138] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0139] Furthermore, embodiments of the subject matter and functional operation described in this specification can be implemented in the following ways: digital electronic circuits, tangibly embodied computer software or firmware, computer hardware including the structures disclosed in this specification and their structural equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by a data processing apparatus or for controlling the operation of a data processing apparatus. Alternatively or additionally, program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are generated to encode information and transmit it to a suitable receiving device for execution by the data processing apparatus. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or combinations thereof.

[0140] The processing and logic flow described in this specification can be executed by one or more programmable computers that execute one or more computer programs to perform corresponding functions by operating on input data and generating output. The processing and logic flow can also be executed by dedicated logic circuitry—such as FPGAs (Field-Programmable Gate Arrays) or ASICs (Application-Specific Integrated Circuits), and the device can also be implemented as dedicated logic circuitry.

[0141] Suitable computers for executing computer programs include, for example, general-purpose and / or special-purpose microprocessors, or any other type of central processing unit. Typically, the central processing unit receives instructions and data from read-only memory and / or random access memory. The basic components of a computer include a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices for storing data, such as disks, magneto-optical disks, or optical disks, or the computer will be operatively coupled to such mass storage devices to receive data from or transfer data to them, or both. However, a computer is not required to have such devices. Furthermore, a computer can be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device such as a universal serial bus (USB) flash drive, to name a few.

[0142] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, such as semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks. Processors and memory may be supplemented by or incorporated into dedicated logic circuitry.

[0143] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily intended to describe features of specific embodiments of a particular invention. Certain features described in the various embodiments herein may also be implemented in combination in a single embodiment. Conversely, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation thereof.

[0144] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0145] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0146] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent agent cooperative control, characterized in that, The method includes: During the process of multiple agents engaging in multiple rounds of discussion on a topic, after each agent speaks, a convergence detection is performed based on the first historical message in the current round of discussion to obtain the convergence detection result. If the convergence detection result is a triggered convergence, the current round of discussion ends. In response to the detection that the current round of discussion has ended, a quality assessment is performed based on the second historical message in the current round of discussion to obtain a quality assessment result; the quality assessment result includes confidence level and suggested action. Based on the confidence level and the suggested action, an execution strategy is determined after the current round of discussion ends; the execution strategy includes continuing to the next round of discussion, entering a closed round, or ending the discussion, where the closed round is used to indicate the last round of discussion before ending the discussion; The process continues according to the described execution strategy until the discussion concludes, yielding the results of the discussion on the aforementioned issues.

2. The method according to claim 1, characterized in that, The method further includes: Once the first agent to speak in the first round of discussion is determined, a prompt word corresponding to the agent is generated based on the agent's role, the discussion question, the quality standard, and the discussion restrictions. The prompt word is then sent to the agent, who then speaks based on the prompt word. The quality standards include at least one of the following: consensus verification standard, quantitative anchoring standard, scheme specificity standard, dimension richness standard, and exit standard. The consensus verification standard instructs agents to raise objections when their conclusions align with those of other agents. The quantitative anchoring standard instructs agents to output conclusive messages including numerical indicators. The scheme specificity standard instructs agents to output scheme messages including names, parameter ranges, and resource estimates. The dimension richness standard instructs agents to cover at least two of several preset dimensions during multi-round discussions on the issue, including technical feasibility, business logic, regulatory pathways, alternative solutions, and exit criteria. The exit criteria instruct agents to determine the failure conditions and loss-mitigation decisions corresponding to the scheme messages discussed. The discussion restrictions include at least one of the following: prohibiting the repetition of opinions output by other agents, prohibiting self-introduction, prohibiting the repetition of the discussion question, providing a message length less than a preset message length, and including a question marker in the message output when there is a question about the discussion question.

3. The method according to claim 1, characterized in that, The convergence detection based on the first historical message in this round of discussion, and the resulting convergence detection results, include: Detect whether a first preset number of agents output historical messages during this round of discussion; If no first preset number of agents output historical messages during this round of discussion, the convergence detection result is determined to be that convergence has not been triggered. In this round of discussion, if there is a first preset number of agents outputting historical messages, the most recent first preset number of agents outputting historical messages is determined as the first historical message. Protocol keyword matching and content similarity detection are performed on the first historical message. The protocol keyword matching is used to detect the number of messages in the first historical message that contain protocol keywords, and the content similarity detection is used to detect the intersection ratio between any two messages in the first historical message. If the number of messages containing protocol keywords in the first historical messages is greater than the second preset number, and / or the number of message pairs in the first historical messages with an intersection ratio greater than a preset intersection ratio is greater than the third preset number, the convergence detection result is determined to be triggered convergence; otherwise, the convergence detection result is determined to be not triggered convergence.

4. The method according to claim 1, characterized in that, The suggested actions include continuing to the next round of discussion, entering a closed round, or ending the discussion; The process of determining the execution strategy after the current round of discussion, based on the confidence level and the suggested action, includes: Based on the confidence level and the suggested action, determine whether the first condition or the second condition is met; the first condition is that the suggested action is to end the discussion and the confidence level is greater than the first confidence level; the second condition is that the suggested action is to enter a closed loop and the confidence level is greater than the second confidence level, and the first confidence level is greater than the second confidence level. If the first condition is met, the execution strategy after the end of this round of discussion is to end the discussion; If the second condition is met, the execution strategy after the end of this round of discussion is to enter the closed loop. If neither the first nor the second condition is met, the execution strategy after the end of this round of discussion is to continue to the next round of discussion.

5. The method according to claim 1, characterized in that, The response, upon detecting the end of the current discussion round, performs a quality assessment based on the second historical messages from the current discussion round, obtaining a quality assessment result, including: In response to the detection that the current round of discussion has ended, determine the next round of discussion; If the current round of discussion is greater than or equal to the minimum number of rounds and less than or equal to the maximum number of rounds, a quality assessment is performed based on the second historical message in the current round of discussion to obtain the quality assessment result.

6. The method according to claim 1, characterized in that, After each agent speaks, a convergence detection is performed based on the first historical message in the current round of discussion to obtain the convergence detection result, including: After each agent speaks, a divergence detection is performed based on the third historical message in the current round of discussion to obtain the divergence detection result. If the divergence detection result indicates that there is no divergence, a convergence detection is performed based on the first historical message in this round of discussion to obtain a convergence detection result.

7. The method according to claim 6, characterized in that, The method further includes: If the disagreement detection result indicates that a disagreement exists, a debate mode is triggered to identify the disagreeing agents and the topics of disagreement. When it is determined that the divergent agent is about to speak, a prompt word corresponding to the divergent agent is generated based on the divergent topic, and the prompt word corresponding to the divergent agent is sent to the divergent agent; the prompt word corresponding to the divergent agent is used to instruct the divergent agent to engage in debate; The debate mode ends when each of the dissenting agents has finished speaking.

8. The method according to claim 1, characterized in that, The method further includes: If any agent's output message includes a question marker, control the discussion mode to pause and display a question message for the user; In response to receiving a reply message from the user, the reply message is added to the history messages and the discussion mode is restored.

9. The method according to claim 1, characterized in that, The method further includes: Upon receiving a user interruption message, pause the discussion mode and identify the agent whose speech was interrupted. The system controls the interrupted agent to output a reply to the user's interrupt message, restores the discussion mode, and controls the interrupted agent to speak again.

10. The method according to claim 1, characterized in that, The method further includes: For each agent other than the one who has spoken in the previous round of discussion, determine the topic relevance score, speaking interval score, and debate context score for that agent; the topic relevance score is determined based on the matching degree between the most recent historical message and the agent; the speaking interval score is determined based on the time of the agent's last speech; the debate context score is determined based on whether the agent is a divergent agent and the agent's role style; The target score for the agent is determined based on the topic relevance score, speaking interval score, and debate context score corresponding to the agent. The agent with the highest target score is selected as the agent to speak.

11. A collaborative control device for intelligent agents, characterized in that, The device includes: The convergence detection module is used to perform convergence detection based on the first historical message in the current round of discussion after each agent speaks during multiple rounds of discussion on a topic, and to obtain the convergence detection result. The discussion end module is used to control the end of this round of discussion if the convergence detection result is a trigger convergence. The quality assessment module is used to perform a quality assessment based on the second historical messages in the current round of discussion in response to the detection that the current round of discussion has ended, and to obtain a quality assessment result; the quality assessment result includes confidence level and suggested action; The strategy determination module is used to determine the execution strategy after the current round of discussion based on the confidence level and the suggested action; the execution strategy includes continuing to the next round of discussion, entering a closed round, or ending the discussion, wherein the closed round is used to indicate the last round of discussion before ending the discussion; The strategy execution module is used to process according to the execution strategy until the discussion ends and obtain the discussion results for the discussion issue.

12. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the intelligent agent cooperative control method according to any one of claims 1 to 10.

13. 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 steps of the agent cooperative control method according to any one of claims 1 to 10.

Images