A station-based AI agent scheduling method and system

Abstract

This invention discloses a workstation-based AI agent scheduling method and system, aiming to solve the problems of tight coupling between tasks and agents and suboptimal scheduling. The method includes: determining the "workstation" corresponding to the task, which is structurally defined with skills, input / output schemas, and service quality requirements; evaluating the "capability declarations" of each candidate AI agent, calculating a comprehensive score from dimensions such as skill matching degree, schema compatibility, and service quality; and selecting the agent with the highest score to execute the task. This invention also includes capturing an execution state snapshot and automatically scheduling a substitute agent for task hot migration when an agent malfunctions. This invention achieves decoupling through workstation abstraction, achieves optimal scheduling through multi-dimensional evaluation, and improves robustness through task hot migration, effectively improving system efficiency and availability.

Description

Technical Field

[0001] This invention relates to the field of multi-agent system technology, specifically to a workstation-based AI agent scheduling method and system. Background Technology

[0002] With the development of artificial intelligence technology, multi-agent systems are playing an increasingly important role in handling complex tasks. When building large-scale AI applications, how to efficiently and reliably orchestrate and schedule multiple AI agents with different functions is the key to determining the overall system performance.

[0003] However, existing agent scheduling schemes generally suffer from several drawbacks. Firstly, there is the problem of tight coupling, where task logic is often directly bound to specific AI agent instances. This hard-coded design leads to a lack of system resilience; if the bound agent becomes unavailable due to upgrades, failures, or excessive load, the related tasks will be interrupted, requiring manual intervention to modify the code or configuration. This results in poor system fault tolerance and robustness.

[0004] Secondly, existing technologies for describing and matching the capabilities of AI agents are relatively rudimentary. Agent capabilities are typically described using unstructured natural language, making it difficult for machines to understand and match them programmatically, thus hindering the scheduling system's ability to accurately select capabilities. Furthermore, the input and output data formats of different AI agents (especially those from different vendors) vary, lacking a unified standard. When multiple agents need to collaborate on a task, developers are forced to write extensive "glue code" (adapter code or interface adaptation layer code) for format conversion, severely impacting development and collaboration efficiency.

[0005] Finally, current scheduling strategies often rely on a single dimension, such as making decisions solely based on whether an agent is online. This approach ignores several important factors, including skill matching, historical service quality, and execution costs. It's like a project manager assigning tasks based solely on employee attendance, without considering their professional skills, work efficiency, and cost. This often results in suboptimal scheduling decisions, failing to achieve overall system efficiency and cost optimization. Therefore, existing technologies struggle to meet the high availability, high efficiency, and high scalability requirements of modern complex AI applications. Summary of the Invention

[0006] To address the technical problems of existing AI agent scheduling schemes, such as tight coupling between tasks and agents, inaccurate capability matching, and single scheduling dimension, this invention provides a workstation-based AI agent scheduling method and system, aiming to decouple tasks from execution capabilities, improve the intelligence level of scheduling decisions, and enhance the overall robustness of the system.

[0007] This invention provides a workstation-based AI agent scheduling method, comprising: acquiring a task to be executed and determining the workstation corresponding to the task; the workstation comprising: structured skill requirements, which are a list, wherein each skill requirement includes a skill name, a minimum required level, and an evaluation weight, and each skill corresponds to a functional sub-module of the workstation; an input / output schema (a specification for defining data structures) for defining the data interaction format; and service quality requirements, which define at least one performance indicator including response time, execution cost, or availability. Subsequently, multiple candidate AI agents are evaluated, wherein each candidate AI agent is associated with a capability declaration, the capability declaration including: a declared skill corresponding to the skill requirement format, the input / output schemas supported by the AI ​​agent, and performance data characterizing the historical performance of the AI ​​agent. The evaluation steps include: calculating a comprehensive score for each candidate AI agent based on the workstation and the capability declarations of each candidate AI agent from at least two preset dimensions selected from skill matching degree, input / output schema compatibility, service quality satisfaction, and cost efficiency. Finally, based on the comprehensive score, at least one target AI agent is selected from the multiple candidate AI agents to perform the task. By introducing "workstation" as a standardized description of task requirements and matching it with the AI ​​agent's "capability declaration" in multiple dimensions, this method decouples the task from the executor, making scheduling decisions more comprehensive and intelligent.

[0008] Furthermore, in the evaluation step, if the input / output schema supported by a candidate AI agent is incompatible with the input / output schema in the workstation, the overall score of the candidate AI agent is adjusted based on the estimated format conversion complexity. This quantifies the cost of interface adaptation and incorporates it into the decision-making process, enabling the system to evaluate and select the execution plan with the lowest total cost of ownership.

[0009] Optionally, the selection step includes selecting the candidate AI agent with the highest overall score as the target AI agent. This ensures that the theoretically optimal executor is selected under the current conditions for each task execution, thereby maximizing the overall efficiency and quality of the system.

[0010] In one implementation, the performance data in the capability declaration is automatically generated and updated by a system through continuous monitoring of the AI ​​agent's historical task execution, and includes cost rates. This ensures that the performance data used for evaluation is objective, real-time, and reliable, avoiding evaluation bias caused by inaccurate self-declarations or capability changes in the agent, and further improving the accuracy of scheduling decisions.

[0011] Furthermore, this method also includes: monitoring the operational status of the target AI agent during task execution, and capturing and saving a snapshot of the execution status related to the task when a preset abnormal event is detected. This step provides a foundation for the system's fault tolerance and self-healing capabilities, enabling timely solidification of the task status when an anomaly occurs, preventing information loss.

[0012] Optionally, the preset abnormal events include: the target AI agent going offline or crashing, or its real-time service quality indicators falling below the threshold defined in the service quality requirements. This specifically defines the conditions for triggering the fault recovery mechanism, making monitoring and response more explicit and efficient.

[0013] In one implementation, the execution state snapshot is a dialogue state context, which includes: a dialogue ID, historical turn records, a context summary, extracted key variables, and a checkpoint vector for semantic retrieval. For stateful, complex interactive tasks (such as multi-turn dialogues), this structured snapshot can completely preserve task progress and key information, enabling smooth recovery.

[0014] Furthermore, this method also includes: after capturing the execution state snapshot, excluding the AI ​​agent that triggered the abnormal event from the candidate AI agents, and re-executing the evaluation and selection steps to determine a substitute AI agent; and loading the execution state snapshot onto the substitute AI agent to enable the substitute AI agent to resume execution from the task interruption point. This series of steps constitutes a complete task hot migration process, realizing application-level automatic fault recovery and significantly improving the robustness and availability of the system.

[0015] Preferably, the loading step includes: using the checkpoint vector to perform semantic retrieval in the knowledge base of the substitute AI agent to recover background knowledge related to the task. Through semantic vector matching, the substitute AI agent can quickly and accurately obtain context and knowledge related to the interrupted task, making task handover smoother and more efficient.

[0016] This invention also provides a workstation-based AI agent scheduling system, comprising: a determination module for acquiring a task to be executed and determining the workstation corresponding to the task; an evaluation module for evaluating multiple candidate AI agents and calculating a comprehensive score for each candidate AI agent; and a selection module for selecting at least one target AI agent from the multiple candidate AI agents to execute the task based on the comprehensive score. The workstation, capability declaration, and evaluation dimensions are consistent with those described in the aforementioned method. This scheduling system can achieve the technical effects of the above methods.

[0017] In summary, this invention, by introducing a "workstation" abstraction layer, completely decouples task requirements from the specific implementation of workstation-based AI agents, resulting in significant system flexibility. Agents can be added, deleted, replaced, or updated at any time without altering business logic. Through a standardized input / output schema, heterogeneous agents from different sources can collaborate seamlessly, greatly improving development and integration efficiency. A multi-dimensional intelligent scheduling algorithm ensures that each task is executed by the most suitable agent, optimizing costs while maintaining quality. More importantly, the dynamic rescheduling and task hot migration mechanism based on state snapshots endows the system with second-level fault self-healing capabilities. When the primary executing agent fails, the system can automatically and smoothly migrate tasks to a backup AI agent, ensuring business continuity and minimizing the impact on user experience. Attached Figure Description

[0018] Figure 1 This is a system architecture diagram of the AI ​​agent scheduling system provided in an embodiment of the present invention.

[0019] Figure 2 A flowchart of the AI ​​agent scheduling method provided in an embodiment of the present invention.

[0020] Figure 3 The flowchart for dynamic rescheduling and task hot migration provided in the embodiments of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0022] Current multi-agent systems commonly suffer from technical problems such as tight coupling between tasks and specific agent instances, vague descriptions of agent capabilities, inconsistent data interaction interfaces, and a single dimension for scheduling decisions. These issues lead to poor system scalability, insufficient robustness, and low operational efficiency. For example, when a task is hard-coded and bound to an AI agent, any failure of that agent will cause the task to be interrupted, requiring manual intervention for repair. Furthermore, unstructured capability descriptions and diverse data formats make automated and precise scheduling and collaboration extremely difficult, often forcing systems to make suboptimal decisions based on a single dimension such as whether the agent is online.

[0023] This invention proposes an innovative workstation-based AI agent scheduling method and system. Its core idea lies in introducing a "workstation" as a core abstraction layer, completely decoupling task requirements from the AI ​​agent's execution capabilities. This idea can be summarized as follows: the "workstation" is responsible for defining "what needs to be done" and "what standards need to be met," while the AI ​​agent describes "what it can do" through standardized "capability declarations." The system then plays the role of intelligent matching and scheduling between the two. In this way, this invention aims to solve the technical problems of tight coupling, ambiguous capability descriptions, inconsistent interfaces, and a single scheduling dimension in existing technologies, thereby constructing a highly available, efficient, and easily scalable agent collaboration architecture.

[0024] In the context of this invention, some core terms have specific meanings. "Workstation" refers to a standardized, structured job description that defines the specific requirements of the task, including structured skill requirements, an input / output schema for defining data interaction formats, and explicit service quality requirements. The skill requirements are a list, with each item corresponding to a functional sub-module of the workstation. Each functional sub-module has an independent minimum capability level and evaluation weight. This design allows a workstation to encompass the definition of multiple functional sub-modules. During the subsequent matching process, each functional sub-module must be independently matched with the corresponding capability of the candidate AI agent. If any sub-module fails to meet the requirements, the entire workstation is considered unsuitable for the AI ​​agent.

[0025] A "capability declaration" is a standardized capability description document submitted by an AI agent when registering with the system. It includes the agent's declared skills, its natively supported input / output schemas, performance data recorded by the system representing its historical performance, and its service cost rate. Furthermore, "input / output schema" specifically refers to a specification based on the JSON Schema standard for describing data structures and formats, ensuring unambiguity in data interaction between different entities. An "execution state snapshot" is a consistent snapshot created during task execution to handle potential anomalies, used to save the complete execution state of the task at a specific moment. For stateful dialogue tasks, a specific execution state snapshot is called a "dialogue state context," which records in detail the dialogue's identifier, history, context summary, key variables, and a checkpoint vector for quickly recovering the knowledge background. These terms collectively form the basis of the technical solution of this invention.

[0026] The architecture design of this invention brings significant benefits. First, by decoupling "workstations" from "capability declarations," significant system resilience is achieved. Intelligent agents can be added, deleted, replaced, or updated at any time without altering any business logic, thus minimizing the impact of single-point failures and increasing availability from the traditional 99% to 99.99%. Second, the unified input / output schema specification eliminates collaboration barriers between heterogeneous intelligent agents, significantly improving development and integration efficiency. Third, the multi-dimensional intelligent scheduling algorithm ensures that each task is matched with the most suitable intelligent agent, achieving overall system cost reduction and efficiency improvement. Finally, the dynamic rescheduling and task hot migration mechanism based on execution state snapshots endows the system with second-level fault self-healing capabilities, ensuring the continuity of user experience.

[0027] Please see Figure 1 This document illustrates the system architecture of an AI agent scheduling system 10 according to one embodiment of the present invention. This scheduling system can be deployed on one or more servers and run as a software system. When an external task to be executed, i.e., task request 20, enters the system, the scheduling system 10 begins operation. The scheduling system mainly includes a determination module 11, an evaluation module 12, and a selection module 13. The system also includes a workstation registration center 30 for storing and managing all defined workstations; an agent registration center 40 for storing and managing the capability declarations of all registered AI agents; a monitoring module 60 for real-time monitoring of the running status of AI agents executing tasks; and a status management module 70 for storing and retrieving execution status snapshots when needed. The entire system revolves around a resource pool containing multiple candidate AI agents 50 for scheduling operations.

[0028] like Figure 2As shown, the overall flow of the AI ​​agent scheduling method provided by this invention is as follows. First, in step S100, the system acquires a task to be executed. Next, in step S110, the system determines the workstation corresponding to the task based on the task type or metadata. Subsequently, in step S120, the system filters all candidate AI agents that are in normal and online status from the agent registration center. After entering the core evaluation stage, in step S130, the system performs a multi-dimensional evaluation for each candidate AI agent. This evaluation process can be specifically decomposed into: step S132, calculating skill matching degree; step S134, calculating input / output schema compatibility; step S136, calculating service quality satisfaction; and step S138, calculating cost efficiency. In step S140, the system weighted sums the scores of each dimension to obtain a comprehensive score for each candidate AI agent. Finally, in step S150, the system selects a target AI agent based on the comprehensive score, and in step S160, dispatches the task to the target AI agent for execution.

[0029] To illustrate the scheduling method of this invention more specifically, a "customer complaint handling" scenario is used as an example. First, the system's maintenance personnel need to create a corresponding workstation. This workstation is assigned a globally unique identifier, such as "workstation-customer-service-text". Its workstation definition includes structured skill requirements, which is a list where each item corresponds to a functional sub-module of the workstation, defining the required skill name, minimum ability level, and evaluation weight for that sub-module. For example, the minimum requirement for the "natural language understanding" skill is 80 points, with an evaluation weight of 0.30; the minimum requirement for the "sentiment analysis" skill is 65 points, with an evaluation weight of 0.20; the minimum requirement for the "complaint resolution" skill is 75 points, with an evaluation weight of 0.30; and the minimum requirement for the "multi-turn dialogue" skill is 70 points, with an evaluation weight of 0.20. This "customer complaint handling" workstation also defines the input / output schema for the data interaction format. Specifically, its input schema is defined as "schema: / / registry / customer-complaint-input-v2", and its output schema is defined as "schema: / / registry / customer-complaint-output-v2". This ensures that any agent scheduled to perform this task must be able to understand and generate data that conforms to both specifications. Furthermore, the workstation defines quality of service requirements, such as requiring 95% of request response times (P95 latency) to be less than 3000 milliseconds, a minimum accuracy of at least 92% for task execution, and a maximum token consumption of no more than 4000 per task. These structured definitions make the task requirements clear and quantifiable, laying the foundation for subsequent automated matching.

[0030] When a new AI agent, such as "agent-ServicePro-3.5-sonnet-v1," needs to connect to the system, it must submit a capability declaration. This declaration includes declared skills corresponding to the job skill requirements format. For example, the agent declares its "Natural Language Understanding" capability as 88 points, "Sentiment Analysis" as 82 points, "Complaint Resolution" as 79 points, and "Multi-turn Dialogue" as 85 points. It also declares its natively supported input / output schemas, such as input as "schema: / / registry / ServicePro-native-input-v1" and output as "schema: / / registry / ServicePro-native-output-v1". The capability declaration also includes performance data continuously monitored and updated by the system, such as a historical average accuracy of 94%, P95 latency of 2200 milliseconds, and Token utilization efficiency of 0.85. Finally, the declaration includes its cost rate, such as a billing unit of 0.003 per thousand Tokens. This standardized capability statement, like a "resume" for an AI agent, allows its capabilities to be understood and evaluated in a systematic and programmatic manner.

[0031] When a "customer complaint handling" task arrives, the evaluation module 12 of the scheduling system 10 begins evaluating multiple candidate AI agents, including "agent-ServicePro-3.5-sonnet-v1" (denoted as agent A). The evaluation process begins with a hard screening: for each functional sub-module in the job definition (i.e., each item in the skill constraint list), it checks whether the candidate AI agent's declared skill level in the corresponding skill meets the minimum level required by that sub-module. If any sub-module does not meet the minimum level requirement, the candidate AI agent is eliminated. Candidates that pass all sub-modules proceed to the subsequent multi-dimensional weighted scoring. Assuming agents A, B, and C all pass this screening, they will then be scored multi-dimensionally to calculate a comprehensive score. For agent A, the skill matching degree is calculated as follows: For each functional sub-module of the job, evaluation module 12 compares the candidate AI agent's declared skill in that sub-module with the minimum level required by the job, and performs a weighted sum based on the evaluation weight corresponding to that sub-module. Since agent A's declared skill levels in all four skills are above the minimum required for the job, its score for each skill is 1.0. Its skill matching score S_skill is calculated as follows: S_skill = 0.30 × min(88, 80) / 80 + 0.20 × min(82, 65) / 65 + 0.30 × min(79, 75) / 75 + 0.20 × min(85, 70) / 70 = 1.0. This calculation method ensures that only agents who fully meet all skill thresholds can obtain a high score.

[0032] Next, input / output schema compatibility is calculated. Since the input / output schema supported by agent A is incompatible with the schema required by its workstation, the system needs to perform format conversion. According to the conversion mapping table provided in agent A's capability declaration, the format conversion complexity from the workstation's input schema to its native input schema is 15, and the format conversion complexity from its native output schema to the workstation's output schema is 20. The maximum allowed format conversion complexity defined for the workstation is 50. Therefore, the system adjusts the score based on the estimated format conversion complexity. The formula for calculating its input / output schema compatibility score S_interface is: S_interface = 0.5 × (1.0 - 15 / 50) + 0.5 × (1.0 - 20 / 50) = 0.65. Here, 0.5 is the default weight for the input and output directions. This design quantifies the overhead caused by interface mismatch, allowing scheduling decisions to take integration costs into account.

[0033] Then, the service quality satisfaction score is calculated. Evaluation module 12 assesses the probability that agent A meets the workstation service quality requirements based on the performance data in its capability declaration. Its historical P95 latency is 2200 milliseconds, far lower than the workstation requirement of 3000 milliseconds, resulting in a latency satisfaction probability of 0.95; its historical average accuracy is 94%, higher than the workstation requirement of 92%, resulting in an accuracy satisfaction probability of 0.97; its token efficiency is 0.85, resulting in a cost satisfaction probability of 0.90. Assuming the evaluation weights for these three service quality indicators are 0.4, 0.4, and 0.2 respectively, the service quality satisfaction score S_slo is calculated as follows: S_slo = 0.4 × 0.95 + 0.4 × 0.97 + 0.2 × 0.90 = 0.942. This score reflects the degree of matching between the agent's historical performance and future task requirements.

[0034] Finally, cost efficiency is calculated. Based on agent A's token efficiency of 0.85 and the maximum token consumption of 4000 defined for the workstation, the system estimates that the actual cost of completing the task is approximately 3000 tokens. The cost efficiency score S_cost is calculated as: S_cost = 1 - 3000 / 4000 = 0.25. This score gives agents with lower costs an advantage in the evaluation, contributing to the optimization of the overall system operating costs.

[0035] After calculating the scores for all dimensions, the evaluation module 12 calculates the comprehensive score of agent A based on preset dimension weights (e.g., skill matching weight 0.35, schema compatibility weight 0.30, service quality satisfaction weight 0.25, cost efficiency weight 0.10). The formula for calculating its comprehensive score S_total(A) is: S_total(A) = 0.35 × 1.0 + 0.30 × 0.65 + 0.25 × 0.942 + 0.10 × 0.25 = 0.806. The evaluation module 12 performs the same calculation process for all other candidate AI agents (such as agents B and C). After evaluating all candidate AI agents, the selection module 13 intervenes. In a preferred embodiment, the selection module 13 selects the candidate AI agent with the highest comprehensive score as the target AI agent. For example, if agent A's comprehensive score of 0.806 is the highest among all candidates, the task will be assigned to agent A for execution. This selection mechanism, based on multi-dimensional comprehensive evaluation, ensures that each task scheduling is an intelligent decision that achieves the best balance between skills, interfaces, quality, and cost.

[0036] To enable the system to adapt, this invention also includes a performance data update mechanism. After the target AI agent completes its task, the monitoring module 60 collects the actual performance data of the task execution, such as the actual response time, the final accuracy evaluation result, and the actual token consumption. The system then uses this new data to update the performance data of the AI ​​agent stored in the agent registry 40. For example, it updates metrics such as average accuracy and P95 latency using moving averages or other statistical methods. This closed-loop feedback mechanism allows the performance data in the capability declaration to dynamically reflect the agent's true performance, thereby making subsequent scheduling decisions more accurate.

[0037] This invention not only focuses on optimal selection but also provides a robust fault self-healing mechanism. For example... Figure 3 As shown, this mechanism is used to achieve dynamic rescheduling and task hot migration. During the execution of a task by the target AI agent, the monitoring module 60 continuously monitors its running status (step S200). When a preset abnormal event is detected (step S210), for example, if the API interface (application interface) of agent A times out multiple times consecutively, its real-time service quality index is lower than the threshold defined in the workstation service quality requirements, or the monitoring module 60 directly detects that the agent is offline or has crashed, the system will immediately trigger the abnormal handling process. After detecting the abnormality, the system immediately captures and saves a snapshot of the execution status related to the current task in step S220. For an ongoing multi-turn dialogue task, this snapshot is specifically represented as a dialogue state context, which is stored by the state management module 70. This dialogue state context contains all the necessary information to ensure that the task can be smoothly recovered. For example, it contains a unique dialogue ID, such as "conv-20260606-001"; records of all historical rounds before the task was interrupted; a contextual summary of the current dialogue progress, such as "The customer reported that the electricity bill was abnormally high, and their electricity address and billing cycle have been confirmed..."; it also contains structured key variables extracted from the dialogue, such as user ID, complaint type, etc.; and a checkpoint vector for quickly locating relevant background information in the knowledge base, such as a 768-dimensional floating-point vector.

[0038] After successfully capturing the execution state snapshot, the scheduling system 10 immediately performs rescheduling. In step S230, the system temporarily excludes the AI ​​agent (Agent A) that triggered the abnormal event from the current pool of candidate AI agents. Then, in step S240, the evaluation module 12 and the selection module 13 immediately re-execute the aforementioned multi-dimensional evaluation and selection steps to determine a substitute AI agent (e.g., Agent B) with the highest overall score from the remaining candidates. After determining the substitute AI agent, the system enters the task loading and recovery phase. In step S250, the system loads the execution state snapshot previously saved in the state management module 70 into the substitute AI agent. Specifically, the loading steps are as follows: the system constructs an informative prompt from the context summary and key variables in the dialogue state context and injects it into the input of the substitute AI agent, enabling it to quickly understand the current progress of the task.

[0039] As a preferred implementation, the loading step also utilizes checkpoint vectors. In step S260, the system uses the checkpoint vectors in the dialogue state context as a query to perform efficient semantic retrieval in the knowledge base of the substitute AI agent. This can quickly recall the processing flow, product documents, or historical cases most relevant to the current dialogue interruption point, thereby restoring deep background knowledge related to the task. In this way, the substitute AI agent can resume execution from the task interruption point almost smoothly. For the end user, the entire failover process may only manifest as a brief delay, without having to start the dialogue from scratch. This mechanism reduces the anomaly recovery time from minutes, traditionally required by manual intervention, to seconds, greatly improving the robustness of the system and the user experience.

[0040] Corresponding to the above method, the present invention also provides an AI intelligent agent scheduling system 10, the structure of which is as follows: Figure 1As shown, the scheduling system includes a determination module 11, an evaluation module 12, and a selection module 13. These modules can be implemented through software programming and deployed on computing devices such as servers. Specifically, the determination module 11 is configured to acquire a task to be executed and determine the corresponding workstation from the workstation registration center 30 based on the task information. As mentioned earlier, the workstation defines structured skill requirements, input / output schema, and service quality requirements. The role of the determination module 11 is to establish a clear and standardized "target" for the subsequent matching process. The evaluation module 12 is configured to evaluate multiple candidate AI agents 50. It acquires the capability declarations of each candidate AI agent from the agent registration center 40, and then, based on the workstation requirements and the agent's declaration, calculates a comprehensive score for each candidate AI agent from multiple preset dimensions such as skill matching degree, input / output schema compatibility, service quality satisfaction, and cost efficiency by performing the aforementioned scoring and weighted calculations. The evaluation module 12 is the core of realizing intelligent decision-making. Selection module 13 is configured to select at least one target AI agent from multiple candidate AI agents to perform the task, based on the comprehensive score calculated by evaluation module 12. In the most common implementation, selection module 13 selects the agent with the highest comprehensive score. In actual operation, determination module 11, evaluation module 12, and selection module 13 work together. They interact through internal application programming interface (API) calls or a shared data bus, efficiently completing the entire scheduling process from receiving task request 20 to finally selecting the target AI agent. This scheduling system, through its modular design, clearly defines responsibilities, making the entire scheduling system easy to implement, maintain, and extend.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A workstation-based AI agent scheduling method, characterized in that, include: Obtain a task to be executed and determine the workstation corresponding to the task. The workstation includes: structured skill requirements, which are a list, where each item contains a skill name, a minimum required level, and an evaluation weight; and an input / output schema for defining the data interaction format. And service quality requirements, which define at least one performance metric including response time, execution cost, or availability; Evaluation is conducted on multiple candidate AI agents, each of which is associated with a capability declaration, the capability declaration including: declared skills corresponding to the skill requirement format, the input / output schemas supported by the AI ​​agent, and performance data characterizing the historical performance of the AI ​​agent; The evaluation steps include: based on the workstation and the capability declarations of each candidate AI agent, calculating a comprehensive score for each candidate AI agent from at least two preset dimensions selected from skill matching degree, input / output schema compatibility, service quality satisfaction, and cost efficiency; Based on the comprehensive score, at least one target AI agent is selected from the plurality of candidate AI agents to perform the task.

2. The method according to claim 1, characterized in that, In the evaluation step, if the input / output schema supported by a candidate AI agent is incompatible with the input / output schema in the workstation, the overall score of the candidate AI agent is adjusted according to the estimated format conversion complexity.

3. The method according to claim 1, characterized in that, The selection step includes: selecting the candidate AI agent with the highest comprehensive score as the target AI agent.

4. The method according to claim 1, characterized in that, Also includes: During the execution of the target AI agent's task, its operating status is monitored, and when a preset abnormal event is detected, an execution status snapshot related to the task is captured and saved.

5. The method according to claim 4, characterized in that, The preset abnormal events include: the target AI agent going offline or crashing, or its real-time service quality index falling below the threshold defined in the service quality requirements.

6. The method according to claim 4, characterized in that, The execution state snapshot is a dialogue state context, which includes: dialogue ID, historical round record, context summary, extracted key variables, and checkpoint vector for semantic retrieval.

7. The method according to claim 4, characterized in that, Also includes: After capturing the execution state snapshot, the AI ​​agent that triggered the abnormal event is excluded from the candidate AI agents, and the evaluation and selection steps are re-executed to determine a substitute AI agent; and the execution state snapshot is loaded into the substitute AI agent so that the substitute AI agent resumes execution from the task interruption point.

8. The method according to claim 7, characterized in that, The loading step includes: using the checkpoint vector to perform semantic retrieval in the knowledge base of the substitute AI agent to recover background knowledge related to the task.

9. The method according to claim 1, characterized in that, The performance data in the capability statement is automatically generated and updated by a system through continuous monitoring of the historical task execution of the AI ​​agent, and includes cost rates.

10. A workstation-based AI intelligent agent scheduling system, characterized in that, include: A determination module is used to acquire a task to be executed and determine the workstation corresponding to the task. The workstation includes: structured skill requirements, which are a list, where each item includes a skill name, a minimum required level, and an evaluation weight; and an input / output schema for defining the data interaction format. The evaluation module includes service quality requirements, which define at least one performance metric, including response time, execution cost, or availability; an evaluation module for evaluating multiple candidate AI agents, each of which is associated with a capability declaration, the capability declaration including: declared skills corresponding to the skill requirement format, the input / output schemas supported by the AI ​​agent, and performance data characterizing the historical performance of the AI ​​agent; the evaluation module is also used to calculate a comprehensive score for each candidate AI agent based on the workstation and the capability declarations of each candidate AI agent, from at least two preset dimensions selected from skill matching degree, input / output schema compatibility, service quality satisfaction, and cost efficiency. A selection module is used to select at least one target AI agent from the plurality of candidate AI agents to perform the task based on the comprehensive score.

Images