Cooperative method and system for enhancing interoperability of AI agents

Abstract

The invention provides a collaboration method and system for enhancing interoperability of AI agents. The collaboration method comprises the steps that S1, the agents are described through structured metadata; s2, decomposing the target task; s3, extracting capability types required by all the sub-tasks according to all the sub-tasks, broadcasting the capability types required by all the sub-tasks, and screening to obtain intelligent agents meeting the capability types required by the sub-tasks; s4, comprehensively scoring the multiple agents corresponding to each sub-task obtained through screening, and screening the agent with the highest score as an executor of the current sub-task; s5, screening a minimum candidate agent subset according to all agents corresponding to all screened subtasks, so that a union set of capabilities of the minimum candidate agent subset can completely cover the total demand of the target task; and S6, based on the current minimum candidate agent subset, executing and finishing the target task according to the dependency relationship among all the sub-tasks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a collaborative method and system for enhancing the interoperability of AI agents. BACKGROUND

[0002] With the continuous development of artificial intelligence (AI) technology, agents are increasingly widely used in various fields. Breakthroughs in AI in speech recognition, image processing, data retrieval, natural language generation, etc. provide strong technical support for the construction of intelligent systems. Multi-agent collaborative working mode is gradually becoming an important technical direction for promoting smart industry, intelligent manufacturing and automated management.

[0003] Currently, multi-agent systems (MAS) still face many challenges in practical applications. Traditional agent architecture often lacks standardized self-description capability, making it difficult for different agents to effectively collaborate. In addition, task allocation and information exchange mostly rely on preset rules, and the flexibility and adaptability of the system are low. When tasks involve multiple agent collaboration, how to efficiently find suitable agents and complete task allocation is still a problem to be solved.

[0004] Patent document CN116233174A (application number: 202211696096.0) discloses an artificial intelligence collaboration, state monitoring, information collection and interaction, data, and capacity sharing system between multiple agents. The technical solution is as follows: it includes a GPRS data transmitter installed on multiple agents, a cloud network server, and a mobile terminal. The multiple agents are wirelessly connected to the cloud network server through the installed GPRS data transmitter. The mobile terminal is wirelessly connected to the cloud network server. The cloud network server is provided with a collaboration control module, an agent state monitoring module, an information collection and interaction module, a data transmission module, and a capacity sharing module. This ensures the stability of multiple agent collaboration, improves the efficiency of multiple agent collaboration, and enhances the transmission effect of data and the connection between multiple agents. This patent relies on a centralized cloud server and a hardware transmission module (GPRS) and uses preset rules for task allocation. The present application realizes autonomous collaboration of agents through a decentralized self-description structure and introduces a dynamic capability combination mechanism, significantly improving the flexibility and adaptability of the system. SUMMARY

[0005] In view of the defects in the prior art, the purpose of the present application is to provide a collaborative method and system for enhancing the interoperability of AI agents.

[0006] According to the collaborative method for enhancing the interoperability of AI agents provided by the present application, the method comprises the following steps: Step S1: Describe the agent through structured metadata, including the agent's own capabilities, task processing scope, and resource conditions; Step S2: Decompose the target task to generate a task dependency graph G = (V, E); where V represents the set of all subtasks; E represents the dependency relationship between all subtasks; Step S3: Extract the required capability types of all subtasks based on all subtasks, and broadcast the required capability types of all subtasks. The agents are matched with the required capability types of all subtasks based on the structured metadata, and the agents that meet the required capability types of the subtasks are screened. Step S4: Comprehensively score the multiple agents corresponding to each subtask obtained by screening, and select the agent with the highest score as the executor of the current subtask; Step S5: Based on all the agents corresponding to all the screened subtasks, further screen the minimum candidate agent subset so that the combined capabilities of the minimum candidate agent subset can fully cover the total requirements of the target task; Step S6: Based on the current minimum subset of candidate agents, execute and complete the target task according to the dependencies between all subtasks.

[0007] Preferably, the structured metadata includes: static metadata and dynamic state information; The static metadata is the functional attributes and interface specifications of the agent; The dynamic status information is the real-time updated running status and availability of the agent.

[0008] Preferably, the step S4 includes:

[0009] in, represents the comprehensive score; 、 is the weight coefficient and satisfies ; Indicates semantic matching degree; Indicates subtask capability requirements; Represents structured metadata of an agent; Representing an agent The real-time load rate; ϵ represents the smoothing constant;

[0010] in, represents the agent with the highest score; Assign subtasks to the corresponding highest-scoring agent in a unicast manner .

[0011] Preferably, the method further comprises: performing confidence evaluation on the execution results of the agents in the minimum candidate agent subset;

[0012] in, Indicates the confidence level of the final calculated result; , is the weight coefficient; is the data support; The total amount of data used to process the task; A data subset that supports the current conclusion R; is the inherent accuracy of the model; The performance indicators of the algorithm model that the agent relies on to perform the current task are evaluated on the offline test set; When the confidence assessment result is less than the preset value, the current agent is automatically blocked and steps S4 to S6 are repeatedly triggered; when the confidence assessment of the execution results of all agents corresponding to the current subtask is less than the preset value, the operation is terminated and a notification of termination is pushed to the client.

[0013] Preferably, step S6 includes: allocating subtasks through autonomous negotiation through a consensus protocol according to the dependencies between all subtasks and the current minimum subset of candidate agents, and agreeing on communication rules.

[0014] A collaborative system for enhancing interoperability of AI agents provided by the present invention includes: Module 1: Describe the agent through structured metadata, including the agent's own capabilities, task processing scope, and resource status; The second module: Decompose the target task to generate a task dependency graph G = (V, E); where V represents the set of all subtasks; E represents the dependency relationship between all subtasks; The third module extracts the required capability types of all subtasks based on all subtasks and broadcasts them. The agents match the required capability types of all subtasks based on structured metadata and select agents that meet the required capability types of the subtasks. The fourth module: comprehensively scores the multiple agents corresponding to each subtask obtained by screening, and selects the agent with the highest score as the executor of the current subtask; The fifth module: Based on all the agents corresponding to all the screened subtasks, further screen the minimum candidate agent subset so that the combined capabilities of the minimum candidate agent subset can fully cover the total requirements of the target task; Module 6: Based on the current minimum subset of candidate agents, complete the target task according to the dependencies between all subtasks.

[0015] Preferably, the structured metadata includes: static metadata and dynamic state information; The static metadata is the functional attributes and interface specifications of the agent; The dynamic status information is the real-time updated running status and availability of the agent.

[0016] Preferably, the fourth module includes:

[0017] in, represents the comprehensive score; 、 is the weight coefficient and satisfies ; Indicates semantic matching degree; Indicates subtask capability requirements; Represents structured metadata of an agent; Representing an agent The real-time load rate; ϵ represents the smoothing constant;

[0018] in, represents the agent with the highest score; Assign subtasks to the corresponding highest-scoring agent in a unicast manner .

[0019] Preferably, the system further comprises: performing confidence evaluation on the execution results of the agents in the minimum candidate agent subset;

[0020] in, Indicates the confidence level of the final calculated result; , is the weight coefficient; is the data support; The total amount of data used to process the task; A data subset that supports the current conclusion R; is the inherent accuracy of the model; The performance indicators of the algorithm model that the agent relies on to perform the current task are evaluated on the offline test set; When the confidence assessment result is less than the preset value, the current agent is automatically blocked and the fourth to sixth modules are repeatedly triggered; when the confidence assessment of the execution results of all agents corresponding to the current subtask is less than the preset value, the operation is terminated and a notification of termination is pushed to the client.

[0021] Preferably, the sixth module includes: autonomously negotiating and allocating subtasks through a consensus protocol based on the dependencies between all subtasks and the current minimum subset of candidate agents, and agreeing on communication rules.

[0022] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention effectively addresses issues such as insufficient interoperability, inefficient task scheduling, and limited data sharing in AI agent systems by employing a self-description structure and an agent collaboration mechanism. The self-description structure provides each agent with a readable capability description and status information, ensuring precise matching of tasks between agents and improving the rationality and efficiency of task allocation. Furthermore, the agent collaboration mechanism includes both unicast and broadcast task distribution methods, enabling on-demand task assignment and enabling autonomous responses from qualified agents, thus achieving more flexible resource scheduling. 2. This invention solves the problems of limited single-agent capabilities, insufficient task coverage, and single data processing methods in traditional AI task execution through agent collaboration. Each agent has an independent self-description structure and professional capabilities in a specific field, thereby achieving efficient task execution and precise processing. At the same time, the agents can autonomously collaborate based on their self-description information to form an efficient task collaboration network, further enhancing the system's intelligent scheduling capabilities. 3. By enhancing the task matching, information sharing, and collaborative execution capabilities of intelligent agents, this invention improves the response speed, data processing efficiency, and task execution accuracy of AI systems in complex task environments, providing an innovative solution for the efficient operation of multi-agent systems. 4. After decomposing complex tasks, the present invention no longer assigns an independent intelligent agent to each subtask. Instead, it regards the entire task as a whole that requires multiple capabilities to complete. From a global optimization perspective, it forms a team of intelligent agents with the least settings, the most comprehensive capabilities, and the highest efficiency for the entire task. It solves the problems of inefficient task matching in multi-agent systems, difficulty in cross-domain collaboration, and insufficient credibility of results, and significantly improves the efficiency of intelligent agent collaboration in industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings: Figure 1 Flowchart of a collaborative approach to enhancing interoperability among AI agents. DETAILED DESCRIPTION

[0024] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0025] Example 1 According to a collaborative method for enhancing interoperability of AI agents provided by the present invention, Figure 1 Shown, including: Step S1: Describe the agent through structured metadata, including the agent's own capabilities, task processing scope, and resource conditions; Step S2: Decompose the target task to generate a task dependency graph G = (V, E); where V represents the set of all subtasks; E represents the dependency relationship between all subtasks; Step S3: Extract the required capability types of all subtasks based on all subtasks, and broadcast the required capability types of all subtasks. The agents are matched with the required capability types of all subtasks based on the structured metadata, and the agents that meet the required capability types of the subtasks are screened. Step S4: Comprehensively score the multiple agents corresponding to each subtask obtained by screening, and select the agent with the highest score as the executor of the current subtask; Step S5: Based on all the agents corresponding to all the screened subtasks, further screen the minimum candidate agent subset so that the combined capabilities of the minimum candidate agent subset can fully cover the total requirements of the target task; Step S6: Based on the current minimum subset of candidate agents, execute and complete the target task according to the dependencies between all subtasks.

[0026] Specifically, the structured metadata includes: static metadata and dynamic state information; The static metadata is the functional attributes and interface specifications of the agent; The dynamic status information is the real-time updated running status and availability of the agent.

[0027] Specifically, step S4 includes:

[0028] in, represents the comprehensive score; 、 is the weight coefficient and satisfies ; Indicates the semantic matching degree; Indicates subtask capability requirements; Represents structured metadata of an agent; Representing an agent The real-time load rate; ϵ represents the smoothing constant;

[0029] in, represents the agent with the highest score; Assign subtasks to the corresponding highest-scoring agent in a unicast manner .

[0030] Specifically, the method further comprises: performing confidence evaluation on the execution results of the agents in the minimum candidate agent subset;

[0031] in, Indicates the confidence level of the final calculated result; , is the weight coefficient; is the data support; The total amount of data used to process the task; A data subset that supports the current conclusion R; is the inherent accuracy of the model; The performance indicators of the algorithm model that the agent relies on to perform the current task are evaluated on the offline test set; When the confidence assessment result is less than the preset value, the current agent is automatically blocked and steps S4 to S6 are repeatedly triggered; when the confidence assessment of the execution results of all agents corresponding to the current subtask is less than the preset value, the operation is terminated and a notification of termination is pushed to the client.

[0032] Specifically, step S6 includes: allocating subtasks through autonomous negotiation through a consensus protocol based on the dependencies between all subtasks and the current minimum subset of candidate agents, and agreeing on communication rules.

[0033] Example 2 A collaborative system for enhancing interoperability of AI agents provided by the present invention includes: Module 1: Describe the agent through structured metadata, including the agent's own capabilities, task processing scope, and resource status; The second module: Decompose the target task to generate a task dependency graph G = (V, E); where V represents the set of all subtasks; E represents the dependency relationship between all subtasks; The third module: extract all the ability types required by all sub-tasks, and broadcast all the ability types required by all sub-tasks, and the agent matches the structured metadata with the ability types required by all sub-tasks, and filters to obtain the agent satisfying the ability types required by the sub-tasks; The fourth module: comprehensively score each agent corresponding to each sub-task, and selects the agent with the highest score as the execution body of the current sub-task; The fifth module: according to all the agents corresponding to all the sub-tasks, filter the minimum candidate agent subset, so that the union of the abilities possessed by the minimum candidate agent subset can completely cover the total demand of the target task; The sixth module: based on the current minimum candidate agent subset, according to the dependency relationship between all sub-tasks, execute the target task.

[0034] Specifically, the structured metadata includes static metadata and dynamic state information. The static metadata is the functional attribute and interface specification of the agent. The dynamic state information is the real-time updated running state and availability of the agent.

[0035] Specifically, the fourth module includes:

[0036] Among them, represents the comprehensive score; , is a weight coefficient, and satisfies ; represents the semantic matching degree; represents the sub-task ability demand; represents the agent structured metadata; represents the real-time load rate of the agent ; and ε represents a smoothing constant.

[0037] Among them, represents the agent with the highest score; assign the sub-task to the corresponding agent with the highest score in the form of unicast .

[0038] Specifically, the system further includes: performing confidence evaluation on the execution results of the agents in the minimum candidate agent subset;

[0039] Among them, represents the final calculated result confidence; , is the weight coefficient; is the data support; The total amount of data used to process the task; A data subset that supports the current conclusion R; is the inherent accuracy of the model; The performance indicators of the algorithm model that the agent relies on to perform the current task are evaluated on the offline test set; When the confidence assessment result is less than the preset value, the current agent is automatically blocked and the first to sixth modules are repeatedly triggered; when the confidence assessment of the execution results of all agents corresponding to the current subtask is less than the preset value, the operation is terminated and a notification of termination is pushed to the client.

[0040] Specifically, the sixth module includes: autonomously negotiating and allocating subtasks through a consensus protocol based on the dependencies between all subtasks and the current minimum subset of candidate agents, and agreeing on communication rules.

[0041] Example 3 Example 3 is a preferred example of Example 1 A collaborative method for enhancing interoperability of AI agents provided by the present invention includes: Step S1: Describe the agent through structured metadata, including the agent's own capabilities, task processing scope, and resource conditions; Specifically, the structured metadata includes: static metadata and dynamic state information; The static metadata defines the basic functional attributes and interface specifications of the agent; Among them, basic functional attributes: use structured formats such as JSON Schema to clearly declare the functional domain, executable operations and required parameters of the intelligent agent; for example: {"domain": "Equipment Maintenance","tools":["Vibration Analysis","Fault Diagnosis"],"input_params":["Equipment ID","Time Range"]}).

[0042] The interface specification: declares communication protocols, such as gRPC; data serialization formats, such as Protobuf / JSONSchema, and API endpoints to ensure technical interoperability.

[0043] The dynamic status information includes real-time updates of the operating status and availability of the agent; including: real-time load rate; Periodically broadcast status information packets, including network location, available computing resources, current task queue length, etc., to ensure information freshness.

[0044] This embodiment also includes a semantic description layer, which can give static capabilities machine-understandable semantics to achieve accurate understanding and reasoning between intelligent agents. Specifically, based on standards such as OWL (Web Ontology Language), the capability terms in the static metadata are mapped to concepts in a unified domain ontology (such as the Industrial Manufacturing Ontology IMOS).

[0045] The semantic description layer can eliminate ambiguity, for example, clarifying that "temperature control" and "thermal management" refer to the same concept; it supports reasoning and can infer that a task requesting "cooling system optimization" requires matching an agent with the ex:ThermalManagement capability, even if the agent is only labeled "temperature control" at the static layer; cross-domain association reveals the relationship between capabilities. For example, it is discovered that "high temperature anomaly detection" may have a collaborative relationship with the "cooling power adjustment" agent, thus providing a basis for forming collaborative groups for complex tasks.

[0046] Step S2: Decompose the target task to generate a task dependency graph G = (V, E); where V represents the set of all subtasks; E represents the dependency relationship between all subtasks; Step S3: Extract the required capability types of all subtasks based on all subtasks, and broadcast the required capability types of all subtasks. The agents are matched with the required capability types of all subtasks based on the structured metadata, and the agents that meet the required capability types of the subtasks are screened. Specifically, step S3 includes: Analyze all subtasks, extract the capability types required for each subtask, and obtain the set of all capability types required for the target task; for example, a "predictive maintenance" task may require C{required}={vibration signal analysis, temperature time series prediction, fault knowledge retrieval}.

[0047] Broadcast the set of all capability types, or query the global self-description structure library maintained by it. All agents that have any one or more capabilities in C{required} will respond, forming the candidate agent set {A1, A2, ..., An}.

[0048] Step S4: Comprehensively score the multiple agents corresponding to each subtask obtained by screening, and select the agent with the highest score as the executor of the current subtask; Specifically, step S4 includes: The filtered agent set { , ,..., } is activated and in a very short decision time window ,For example =Calculate its own comprehensive score within 500ms .

[0049] Comprehensive score Score( ) is calculated by weighting two key factors, and the calculation formula is as follows:

[0050] in, is the semantic matching degree; is the load health; For intelligent agents For tasks The overall score or bid price of the bid; the higher the score, the more suitable it is for performing the task; α , β is Weight coefficients, satisfying α + β = 1; used to adjust the relative importance of "capacity matching" and "load status" in the scoring. For example, in a scenario focusing on quality, α = 0.7 and β = 0.3; in a scenario focusing on throughput, α = 0.4 and β = 0.6. is the semantic matching degree. Computational task requirements and agent capabilities The similarity between them is usually in the range of [0,1]. The higher the value, the more consistent the agent's professional capabilities are with the task requirements. For the task A description of the capability requirements. This is usually keywords or embedding vectors extracted from the task description; For intelligent agents A description of a capability, derived from its structured metadata; For intelligent agents The real-time load rate is calculated as , the value range is [0,1], the higher the value, the busier the agent; ϵ is a very small smoothing constant, such as ϵ=0.001, to prevent the denominator from being zero and ensure that the formula is =0 is still meaningful. It is the inverse of load health; load The lower it is, the larger the value of this item is, indicating that the agent is "more free" to handle new tasks, thus obtaining a higher score.

[0051] Waiting time window After the end, collect all the returned scores .

[0052] Select the agent with the highest overall score as the winner:

[0053] Then, the task T is formally assigned to .

[0054] Step S5: Based on all the agents corresponding to all the screened subtasks, further screen the minimum candidate agent subset so that the combined capabilities of the minimum candidate agent subset can fully cover the total requirements of the target task; Specifically, step S5 includes: finding a minimum candidate agent subset Team so that the union of the capabilities possessed by Team can completely cover the total demand C{required}.

[0055] In practice, approximate algorithms such as greedy algorithms are usually used to find the optimal solution: Prioritize agents that can cover the most uncovered capabilities to join the team. Update the remaining uncovered capabilities; repeat the process until all capabilities are covered.

[0056] The final Team = {Ax, Ay, ...} is the required minimum (or approximately minimum) combination of agents with the required coverage capability set.

[0057] Step S6: Based on the current minimum candidate agent subset, execute and complete the target task according to the dependencies between all subtasks; Specifically, step S6 includes: notifying all agents of the formed Team; allocating subtasks through autonomous negotiation based on the generated task dependency graph G and the specific capabilities of each agent, and agreeing on communication rules through a consensus protocol such as PBFT.

[0058] In this embodiment, it also includes: After completing a task, each agent doesn't directly return a result. Instead, it initiates an internal self-evaluation process. This process comprehensively analyzes two dimensions: data support and the credibility of the model itself. Using a quantifiable formula, it ultimately generates a confidence score (R) between [0 and 1].

[0059]

[0060] Among them, Confidence (R) is the confidence of the final calculated result, which is a value between [0,1]. The higher the value, the more confident the agent is in the result R it generates. , is a weight coefficient used to adjust the relative importance of "data support" and "model accuracy" in the final confidence assessment. The weight can be pre-set or dynamically adjusted according to the task type. For example: for data-driven tasks such as anomaly detection, you can set > , trust the data itself more; for model reasoning tasks, such as text generation, you can set > , more dependent on the capabilities of the model itself; Data support, which is the objective data basis for confidence assessment; The total amount of data used to process the task; for example, when processing sensor data over a period of time It could be the total number of all data points in the time period; The subset of data that supports the current conclusion R. The agent will use its internal logic or rules to determine which data supports the final conclusion. For example: in classification tasks, the number of samples correctly classified by the model with high prediction probability; in prediction tasks, the number of historical data points that meet the predicted trend; in anomaly detection, the number of data points marked as anomalous and significantly deviating from the mean. The model's inherent accuracy serves as the prior performance basis for confidence assessment. Acc refers to the performance metrics of the algorithmic model the agent relies on to perform the current task, such as accuracy, F1-score, and precision, as evaluated on an offline test set. This value is a predefined constant stored in the agent's self-description structure and represents the agent's average performance level under ideal conditions.

[0061] When the confidence assessment result is less than the preset value, the current agent is automatically blocked and the agent corresponding to the current subtask is selected again; when the confidence assessment of the execution results of all agents corresponding to the current subtask is less than the preset value, the operation is terminated and a notification of the termination is pushed to the client, triggering the exception handling process, such as manual intervention or reallocation after adjusting the task parameters.

[0062] This confidence assessment mechanism is both objective and a priori, taking into account both the specific data of the task and the general ability level of the executor. It is quantifiable and explainable, calculated through mathematical formulas, and the result is a clear numerical value rather than a vague "high / medium / low", which facilitates automated processing. It is also flexible and configurable, and can adapt to the trust preferences of different task types by adjusting the weights.

[0063] Those skilled in the art will appreciate that, in addition to implementing the system, device, and various modules provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, and the like by logically programming the method steps. Therefore, the system, device, and various modules provided by the present invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; the modules for implementing various functions can also be considered both software programs for implementing the method and structures within the hardware component.

[0064] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A collaborative method for enhancing interoperability of AI agents, characterized in that: include: Step S1: Describe the agent through structured metadata, including the agent's own capabilities, task processing scope, and resource conditions; Step S2: Decompose the target task to generate a task dependency graph G = (V, E); where V represents the set of all subtasks; E represents the dependency relationship between all subtasks; Step S3: Extract the required capability types of all subtasks based on all subtasks, and broadcast the required capability types of all subtasks. The agents are matched with the required capability types of all subtasks based on the structured metadata, and the agents that meet the required capability types of the subtasks are screened. Step S4: Comprehensively score the multiple agents corresponding to each subtask obtained by screening, and select the agent with the highest score as the executor of the current subtask; Step S5: Based on all the agents corresponding to all the screened subtasks, a minimum candidate agent subset is selected so that the combined capabilities of the minimum candidate agent subset can fully cover the total requirements of the target task; Step S6: Based on the current minimum subset of candidate agents, execute and complete the target task according to the dependencies between all subtasks.

2. The collaborative method for enhancing interoperability of AI agents according to claim 1, characterized in that: The structured metadata includes: static metadata and dynamic status information; The static metadata is the functional attributes and interface specifications of the agent; The dynamic status information is the real-time updated running status and availability of the agent.

3. The collaborative method for enhancing interoperability of AI agents according to claim 1, characterized in that: The step S4 comprises: in, represents the comprehensive score; 、 is the weight coefficient and satisfies ; Indicates the semantic matching degree; Indicates subtask capability requirements; Represents structured metadata of an agent; Representing an agent The real-time load rate; ϵ represents the smoothing constant; in, represents the agent with the highest score; Assign subtasks to the corresponding highest-scoring agent in a unicast manner .

4. The collaborative method for enhancing interoperability of AI agents according to claim 1, characterized in that: The method further comprises: performing confidence evaluation on the execution results of the agents in the minimum candidate agent subset; in, Indicates the confidence level of the final calculated result; , is the weight coefficient; is the data support; The total amount of data used to process the task; A data subset that supports the current conclusion R; is the inherent accuracy of the model; The performance indicators of the algorithm model that the agent relies on to perform the current task are evaluated on the offline test set; When the confidence assessment result is less than the preset value, the current agent is automatically blocked and steps S4 to S6 are repeatedly triggered; when the confidence assessment of the execution results of all agents corresponding to the current subtask is less than the preset value, the operation is terminated and a notification of termination is pushed to the client.

5. The collaborative method for enhancing interoperability of AI agents according to claim 1, characterized in that: The step S6 includes: allocating subtasks through autonomous negotiation through a consensus protocol according to the dependencies between all subtasks and the current minimum subset of candidate agents, and agreeing on communication rules.

6. A collaborative system for enhancing interoperability of AI agents, characterized in that: include: Module 1: Describe the agent through structured metadata, including the agent's own capabilities, task processing scope, and resource status; The second module: Decompose the target task to generate a task dependency graph G = (V, E); where V represents the set of all subtasks; E represents the dependency relationship between all subtasks; The third module extracts the required capability types of all subtasks based on all subtasks and broadcasts them. The agents match the required capability types of all subtasks based on structured metadata and select agents that meet the required capability types of the subtasks. The fourth module: comprehensively scores the multiple agents corresponding to each subtask obtained by screening, and selects the agent with the highest score as the executor of the current subtask; The fifth module: Based on all the agents corresponding to all the screened subtasks, the minimum candidate agent subset is selected so that the combined capabilities of the minimum candidate agent subset can fully cover the total requirements of the target task; Module 6: Based on the current minimum subset of candidate agents, complete the target task according to the dependencies between all subtasks.

7. The collaborative system for enhancing interoperability of AI agents according to claim 6, characterized in that: The structured metadata includes: static metadata and dynamic status information; The static metadata is the functional attributes and interface specifications of the agent; The dynamic status information is the real-time updated running status and availability of the agent.

8. The collaborative system for enhancing interoperability of AI agents according to claim 6, characterized in that: The fourth module includes: in, represents the comprehensive score; 、 is the weight coefficient and satisfies ; Indicates the semantic matching degree; Indicates subtask capability requirements; Represents structured metadata of an agent; Representing an agent The real-time load rate; ϵ represents the smoothing constant; in, represents the agent with the highest score; Assign subtasks to the corresponding highest-scoring agent in a unicast manner .

9. The collaborative system for enhancing interoperability of AI agents according to claim 6, characterized in that: The system further comprises: performing confidence evaluation on the execution results of the agents in the minimum candidate agent subset; in, Indicates the confidence level of the final calculated result; , is the weight coefficient; is the data support; The total amount of data used to process the task; A data subset that supports the current conclusion R; is the inherent accuracy of the model; The performance indicators of the algorithm model that the agent relies on to perform the current task are evaluated on the offline test set; When the confidence assessment result is less than the preset value, the current agent is automatically blocked and the first to sixth modules are repeatedly triggered; when the confidence assessment of the execution results of all agents corresponding to the current subtask is less than the preset value, the operation is terminated and a notification of termination is pushed to the client.

10. The collaborative system for enhancing interoperability of AI agents according to claim 6, characterized in that: The sixth module includes: autonomously negotiating and allocating subtasks through a consensus protocol based on the dependencies between all subtasks and the current minimum subset of candidate agents, and agreeing on communication rules.

Images