Multi-strategy engine application method and system of intelligent agent

Abstract

The invention provides a multi-strategy engine application method and system of an intelligent agent, and the method comprises the steps: a strategy evaluation model of a strategy layer receives request feature data of a task request, input by a user, of the intelligent agent, outputs an evaluation result, and a strategy selection model selects a target strategy engine from more than two strategy engines based on the evaluation result, the target strategy engine generates a calling instruction; the execution layer receives the calling instruction, executes tool calling operation and returns an operation result to the strategy layer, and the strategy layer generates task reply information according to the operation result and returns the task reply information to the user through the interface layer. According to the method, the intelligent body can adapt to different tasks and is suitable for a proper target strategy engine, and the coordination of solving complex and simple tasks at the same time is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a multi-strategy engine application method and system of an agent. BACKGROUND

[0002] At present, various agents have emerged on the market. Most of the agents independently use a strategy engine such as function calling or reasoning-acting (Reasoning Acting, ReAct) after receiving a task request of a user, which leads to over-reasoning for simple tasks or failure of execution for complex tasks, lacks a dynamic switching mechanism, and has insufficient collaboration. In addition, there is a high possibility that a large language model (Large Language Model, LLM) in the starting process of the strategy engine does not match an interface between the tool to be called, and execution of the task is prone to lag. SUMMARY

[0003] The present application provides a multi-strategy engine application method and system of an agent, which can make the agent face different tasks, apply a suitable target strategy engine, and improve the coordination of the same agent in solving complex and simple tasks.

[0004] The first aspect of the present application provides a multi-strategy engine application method of an agent. A strategy layer includes a strategy evaluation model and a strategy selection model. The method includes that the strategy evaluation model of the strategy layer receives request feature data of a task request of a user input agent, outputs an evaluation result, the strategy selection model selects a target strategy engine in two or more strategy engines based on the evaluation result, and the target strategy engine generates a calling instruction. An execution layer receives the calling instruction, performs a tool calling operation, and returns an operation result to the strategy layer. The strategy layer generates task reply information according to the operation result and returns the user through an interface layer.

[0005] In some embodiments, the request feature data includes ambiguity of the task request, a number of tool dependencies, and historical execution time consumption. The strategy evaluation model is constructed by using a logistic regression algorithm. The score result is an evaluation score.

[0006] In some embodiments, the strategy evaluation model receives request feature data of a task request of a user input agent, and outputs an evaluation result, including: The strategy evaluation model multiplies the ambiguity of the task request of the user input agent, the number of tool dependencies, and the historical execution time consumption, respectively, and a corresponding preset feature weight. After adding a plurality of multiplication results and adding a preset bias term, a total sum is obtained. After normalization, an evaluation score is obtained, and the evaluation score is output.

[0007] In some embodiments, before the policy evaluation model receives the request feature data of the task request of the user input agent, the method further comprises: The large language model of the policy layer receives the request text of the task request of the user sent by the interface layer, calculates the edit distance between the request text and the standard API description, determines the ambiguity of the task request according to the edit distance, and sends it to the policy evaluation model; The execution layer sends the historical execution time of the available tool set corresponding to the task request to the policy evaluation model.

[0008] In some embodiments, the interface layer includes a request parsing module, a service discovery module, and an interface registration module, and before the large language model receives the request text of the task request of the user sent by the interface layer, the method further comprises: The request parsing module receives the task request of the user and parses the user intent and entity parameters; The service discovery module queries the available tool set, the tool standardized description corresponding to the tool, and the tool dependency number from the interface registration module according to the user intent, constructs the request text according to the tool standardized description and the entity parameters, and sends the tool dependency number to the policy evaluation model and the request text and the available tool set to the large language model, respectively; wherein the interface registration module defines the input and output formats of the tool according to the tool standardized description, and unifies the formats of the tools of the available tool set to formats that can be recognized by the large language model.

[0009] In some embodiments, the policy selection model selects a target policy engine from two or more policy engines based on the evaluation result, comprising: If the evaluation score is greater than a preset score threshold, the policy selection model enables the reasoning-action policy engine, and if the evaluation score is less than or equal to the preset score threshold, the policy selection model enables the function call policy engine.

[0010] In some embodiments, the execution layer further comprises a tool calling module, a model context protocol server, an asynchronous execution module, and a state tracking module, the execution layer receives a calling instruction, performs a tool calling operation, and returns an operation result to the policy layer, and the policy layer generates task reply information according to the operation result and returns it to the user through the interface layer, comprising: The tool calling module receives the calling instruction and calls the asynchronous execution module; The asynchronous execution module calls the tool indicated by the calling instruction, performs the operation corresponding to the task request, generates an operation result, and returns it to the target policy engine of the policy layer through the tool calling module; If it is judged that the operation is successful according to the operation result, the target strategy engine obtains context information corresponding to the operation result from the model context protocol server, constructs a prompt according to the context information, and sends the prompt to the large language model; wherein the context information is constructed by coordinating the task state snapshot of the state tracking module, the tool calling result of the tool calling module, and the operation history record through the model context protocol server. The large language model generates task reply information according to the prompt and returns the user through the interface layer.

[0011] In some embodiments, the method further comprises: If it is judged that the operation is abnormal according to the operation result, the target strategy engine sends an abnormal operation result to the strategy selection model to generate an enable instruction of the reasoning-action strategy engine to the large language model, the large language model corrects the entity parameters or the available tool set according to the context information corresponding to the abnormal operation result, generates a new reasoning-action sequence and sends it to the reasoning-action strategy engine, and the reasoning-action strategy engine generates a new calling instruction according to the reasoning-action sequence.

[0012] In some embodiments, the tool calling module includes a DAG scheduler, and the tool calling module receives the calling instruction, and calls the asynchronous execution module, including: The DAG scheduler receives the calling instruction, arranges the calling order of the tools based on the directed acyclic graph carried by the calling instruction, sends an interface calling request to the interface registration module, and calls the asynchronous execution module according to the returned tool access information and calling order, wherein the directed acyclic graph represents the dependency relationship and execution order between parallel sub-tasks.

[0013] The second aspect of the present application provides a multi-strategy engine application system of an intelligent agent, the system comprising a strategy layer, an execution layer and an interface layer, the strategy layer comprising a strategy evaluation model and a strategy selection model; The strategy evaluation model of the strategy layer receives request feature data of a task request of the intelligent agent input by the user, outputs an evaluation result, the strategy selection model selects a target strategy engine among two or more strategy engines based on the evaluation result, and the target strategy engine generates a calling instruction; The execution layer receives the calling instruction, performs a tool calling operation, and returns the operation result to the strategy layer, the strategy layer generates task reply information according to the operation result and returns the user through the interface layer.

[0014] It can be understood that the multi-strategy engine application method and system of the agent of the application, the strategy evaluation model of the strategy layer receives the request feature data of the task request of the user input agent, outputs the evaluation result, and then the strategy selection model selects the target strategy engine in two or more strategy engines based on the evaluation result, and the target strategy engine generates the calling instruction, so that the agent can face different tasks and apply the appropriate target strategy engine, and the coordination of the same agent in solving complex and simple tasks is improved. BRIEF DESCRIPTION OF DRAWINGS

[0015] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate embodiments consistent with the application and, together with the description, serve to explain the principles of the application.

[0016] Figure 1 A schematic diagram of a framework of the multi-strategy engine application system of the agent provided for the embodiments of the application is shown. Figure 2 A flowchart of the multi-strategy engine application method of the agent provided for the embodiments of the application is shown. Figure 3 An application scenario diagram of the multi-strategy engine application system of the agent provided for the embodiments of the application is shown. Figure 4 Another application flowchart of the multi-strategy engine application method of the agent provided for the embodiments of the application is shown.

[0017] Through the above drawings, the specific embodiments of the application have been shown, and more detailed descriptions will be given hereinafter. These drawings and textual descriptions are not intended to limit the scope of the concept of the application in any way, but to illustrate the concept of the application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0018] The exemplary embodiments will be described in detail herein with reference to the drawings. Unless otherwise specified, the same or similar components in different drawings are denoted by the same reference numerals, and the description of the same or similar components will not be repeated. The embodiments described in the following exemplary embodiments do not represent all the embodiments consistent with the application.

[0019] The terms "first", "second", and the like used in the application are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features.

[0020] Please refer to Figure 1 , Figure 1The application relates to a multi-strategy engine application system 100 for an intelligent agent, which can be integrated into the intelligent agent, so that the intelligent agent can apply a suitable strategy engine to solve a task request of a user after receiving the task request.

[0021] The technical solutions of the application and how the technical solutions solve the technical problems will be described in detail in specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described again in some embodiments.

[0022] Please refer to Figures 1 to 3 , Figure 2 A flowchart of a multi-strategy engine application method provided by the application is shown in the figure. The application method can be applied to a multi-strategy engine application system 100. As shown in Figure 2 the figure, the multi-strategy engine application method can include the following steps: Step S210: A strategy evaluation model of a strategy layer receives request feature data of a task request of a user input intelligent agent, outputs an evaluation result, a strategy selection model selects a target strategy engine from two or more strategy engines based on the evaluation result, and the target strategy engine generates a calling instruction.

[0023] Specifically, before step S210 is performed, the request feature data is obtained through steps S110 to S150. As shown in Figure 1 the figure, the interface layer 10 includes a request parsing module 11, a service discovery module 12, and an interface registration module 13.

[0024] Step S110: The request parsing module receives a task request of a user and parses the user intent and entity parameters. Specifically, the user intent is “looking for popular snacks” or “the location of snacks”. The entity parameters are the location “A place” and the time “today”.

[0025] Step S120: The service discovery module queries the available tool set, the tool standardized description corresponding to the tool, and the tool dependency number from the interface registration module according to the user intent, constructs a request text according to the tool standardized description and the entity parameters, sends the tool dependency number to the strategy evaluation model, and sends the request text and the available tool set to the large language model.

[0026] Specifically, the strategy layer 20 further includes a large language model 24 and a strategy evaluation model 25. The request text contains which tool to be called and what the specific calling parameters are. The interface registration module 13 defines the input and output formats of the tool according to the tool standardized description (MCP Schema), and unifies the formats of the tools in the available tool set to formats that can be recognized by the large language model 24.

[0027] It can be understood that in the related art, the tool call needs to customize the interface for different LLMs, resulting in high development complexity. In the embodiment, the MCP Schema under the MCP protocol can be compatible with LLMs of multiple manufacturers, reducing the tool access cost.

[0028] Step S140: The large language model of the policy layer receives the request text of the user's task request sent by the interface layer, calculates the edit distance between the request text and the standard API description, determines the ambiguity of the task request according to the edit distance, and sends it to the policy evaluation model.

[0029] Step S150: The execution layer sends the historical execution time of the available tool set corresponding to the task request to the policy evaluation model. Specifically, the state tracking module 34 of the execution layer 30 stores the historical execution time of the available tool set corresponding to the task request.

[0030] After step S150 is executed, the policy evaluation model multiplies the ambiguity of the user's task request, the number of tool dependencies, and the historical execution time by the corresponding preset feature weight respectively; adds the plurality of multiplication results to obtain a total, and adds the total to a preset bias term to obtain an evaluation score, and outputs the evaluation score. The expression of the policy evaluation model 25 is as follows:

[0031] Wherein, represents the evaluation score, represents the preset bias term, represents the ambiguity, represents the preset feature weight corresponding to the ambiguity, represents the number of tool dependencies, represents the preset feature weight corresponding to the number of tool dependencies, represents the historical execution time, represents the preset feature weight corresponding to the historical execution time. It can be understood that the request feature data includes the ambiguity of the task request, the number of tool dependencies, and the historical execution time, the policy evaluation model 25 is constructed by using a logistic regression algorithm (Logistic Regression), and the score result is the evaluation score.

[0032] In one embodiment, the target policy engine is a reasoning-action policy engine or a function call policy engine. The policy selection model selects the target policy engine in two or more policy engines based on the evaluation result, including: if the evaluation score is greater than a preset score threshold, the policy selection model enables the reasoning-action policy engine, and if the evaluation score is less than or equal to the preset score threshold, the policy selection model enables the function call policy engine.

[0033] In particular, the preset score threshold can be 0.7. The specific way in which the strategy selection model 23 enables the inference-action strategy engine 22 is that the strategy selection model 23 sends an enabling instruction for the strategy selection model 23 to generate the inference-action strategy engine 22 to the large language model 24, and the large language model 24 generates an inference-action sequence based on the available tool set, the request text and the entity parameter and sends it to the inference-action strategy engine 22. The strategy selection model 23 enables the function call strategy engine for the strategy selection model 23 to send an enabling instruction for the strategy selection model 23 to generate the function call strategy engine 21 to the large language model 24, and the large language model 24 generates a tool call sequence based on the available tool set, the request text and the entity parameter and sends it to the function call strategy engine 21.

[0034] It can be understood that the evaluation score greater than the preset score threshold indicates that the task request is relatively complex, and the inference-action strategy engine 22 is suitable for accurately performing the complex task. The evaluation score less than or equal to the preset score threshold indicates that the task request is relatively simple, and the function call strategy engine 21 is suitable for directly calling the tool to avoid excessive reasoning.

[0035] It can be understood that in the related art, some agents only independently use the function call strategy engine 21, and the function call strategy engine 21 is only suitable for explicit tasks, and parameter extraction depends on a predefined structure, making it difficult to handle ambiguous intentions or multi-step reasoning requirements. Some agents only independently use the inference-action strategy engine 22, and the inference-action strategy engine 22 needs to iterate reasoning multiple times, which is high in resource consumption and low in execution efficiency, and the stability of tool calling depends on the external system 100. In the embodiment of the present application, the corresponding strategy engine is selected according to the evaluation score, which can solve the problem of independent use of a single strategy engine by an agent in the related art.

[0036] Step 220: The execution layer receives the calling instruction, performs the tool calling operation, and returns the operation result to the strategy layer, and the strategy layer generates task reply information according to the operation result and returns it to the user through the interface layer.

[0037] In particular, the execution layer 30 further includes a tool calling module 32, a model context protocol server 31 (MCPServer), an asynchronous execution module 33 and a state tracking module 34. In an implementation, the tool calling module 32 receives the calling instruction and calls the asynchronous execution module 33. The tool calling module 32 includes a DAG scheduler (not shown) that receives the calling instruction, arranges the calling order of the tools based on the directed acyclic graph carried by the calling instruction, sends an interface calling request to the interface registration module 13, and calls the asynchronous execution module 33 according to the returned tool access information and the calling order. Figure 1 The DAG scheduler can represent the dependency relationship and execution order between parallel sub-tasks.

[0038] Next, the asynchronous execution module 33 calls the tool indicated by the call instruction to perform the operation corresponding to the task request, generates the operation result, and returns it to the target strategy engine of the strategy layer 20 through the tool call module 32.

[0039] If the operation is deemed successful based on the operation result, the target strategy engine obtains the context information corresponding to the operation result from the model context protocol server 31, constructs a prompt based on the context information, and sends it to the large language model 24. The context information is constructed by coordinating the task status snapshot of the state tracking module 34, the tool call results of the tool call module 32, and the operation history through the model context protocol server 31.

[0040] Understandably, in related technologies, function call engines lack cross-task state management capabilities, making it difficult to adapt to complex task scenarios. This application embodiment coordinates information such as task state snapshots, tool call results, and historical records of multiple execution layer 30 modules through an integrated model context protocol server 31, unifying the scheduling of tool calls and ensuring system data consistency when switching strategies, thus achieving cross-task state management.

[0041] Finally, the large language model 24 generates task response information based on the prompts and returns it to the user through interface layer 10. Figure 3 (Not shown).

[0042] In another implementation, the tool invocation module 32 receives an invocation instruction and invokes the asynchronous execution module 33. Then, the asynchronous execution module 33 invokes the tool indicated by the invocation instruction to perform the operation corresponding to the task request, generates the operation result, and returns it to the target policy engine of the policy layer 20 through the tool invocation module 32.

[0043] If the operation result indicates an anomaly, the target strategy engine generates an anomaly signal and sends it to the strategy selection model 23. The strategy selection model 23 generates an activation instruction for the reasoning-action strategy engine 22 and sends it to the large language model 24. The large language model 24 corrects the entity parameters or available toolset based on the context information corresponding to the anomaly result, generates a new reasoning-action sequence, and sends it to the reasoning-action strategy engine 22. The reasoning-action strategy engine 22 generates a new invocation instruction based on the reasoning-action sequence, so that the tool invocation module 32 receives the new invocation instruction and executes the previous implementation method. Figure 3 (Not shown).

[0044] Understandably, if the target strategy engine (i.e., the inference-action strategy engine 22 or the function call strategy engine 21) calls the tool with abnormal results, the inference-action strategy engine 22 can analyze and correct the cause of the abnormality and generate new call instructions to enhance the system's 100% fault tolerance.

[0045] Understandably, in the above technical solution, the strategy evaluation model 25 of the strategy layer 20 receives the request feature data of the task request of the intelligent agent input by the user, outputs the evaluation result, and then the strategy selection model 23 selects the target strategy engine from two or more strategy engines based on the evaluation result. The target strategy engine generates the calling instruction, which can enable the intelligent agent to apply the appropriate target strategy engine when facing different tasks, thereby improving coordination.

[0046] Combination Figure 3 and Figure 4 The following application scenario will further illustrate the application method of the multi-policy engine for intelligent agents.

[0047] Step S1: The request parsing module of the interface layer receives the user's task request and parses out the user intent and entity parameters; the service discovery module of the interface layer queries the available toolset, the tool standardization description (MCP Schema) corresponding to the tool, and the number of tool dependencies from the interface registration module according to the user intent, constructs the request text according to the tool standardization description and entity parameters, and sends the number of tool dependencies to the policy evaluation model, and sends the request text and available toolset to the large language model.

[0048] Among them, the interface registration module of the interface layer defines the input and output formats of tools according to the tool standardization description, and unifies the format of tools in the available toolset into a format that can be recognized by the large language model.

[0049] Step S2: The large language model of the strategy layer receives the request text, calculates the edit distance between the request text and the standard API description, determines the ambiguity of the task request based on the edit distance, and sends it to the strategy evaluation model.

[0050] Step S3: The execution layer sends the historical execution time of the available toolset corresponding to the task request to the strategy evaluation model.

[0051] Step S4: The strategy evaluation model receives the ambiguity of the task request, the number of tool dependencies, and the historical execution time, and outputs an evaluation score.

[0052] Step S5: If the evaluation score is greater than the preset score threshold, the strategy selection model activates the inference-action strategy engine, which generates a call instruction; if the evaluation score is less than or equal to the preset score threshold, the strategy selection model activates the function call strategy engine, which generates a call instruction.

[0053] Specifically, the strategy selection model enables the inference-action strategy engine by sending an activation command to the large language model. The large language model then generates an inference-action sequence based on the available toolset, request text, and entity parameters, and sends it to the inference-action strategy engine. Similarly, the strategy selection model enables the function call strategy engine by sending an activation command to the large language model. The large language model then generates a tool call sequence based on the available toolset, request text, and entity parameters, and sends it to the function call strategy engine.

[0054] Step S6: The tool invocation module of the execution layer receives the invocation instruction and invokes the asynchronous execution module of the execution layer.

[0055] The tool invocation module integrates a DAG scheduler. The DAG scheduler receives invocation instructions, arranges the invocation order of tools based on the directed acyclic graph carried by the invocation instructions, sends an interface invocation request to the interface registration module, and invokes the asynchronous execution module according to the returned tool access information and invocation order. The directed acyclic graph represents the dependencies and execution order between parallel subtasks.

[0056] Step S7: The asynchronous execution module calls the tool indicated by the call instruction to perform the operation corresponding to the task request, generates the operation result, and returns it to the corresponding inference-action strategy engine or function call strategy engine in the strategy layer through the tool call module.

[0057] Step S8: The reasoning-action strategy engine or function call strategy engine determines whether the operation is successful or abnormal based on the operation result. If successful, proceed to step S9; if abnormal, proceed to step S10.

[0058] Step S9: The reasoning-action strategy engine or function call strategy engine obtains the context information corresponding to the operation result from the model context protocol server of the execution layer, constructs a prompt based on the context information and sends it to the large language model. The large language model generates task response information based on the prompt and returns it to the user through the interface layer.

[0059] The context information is constructed by combining the task status snapshot of the model context protocol server coordination state tracking module, the tool call results of the tool call module, and the operation history.

[0060] Step S10: The inference-action strategy engine or function call strategy engine generates an exception signal from the abnormal operation result and sends it to the strategy selection model. The strategy selection model generates an enable instruction for the inference-action strategy engine and sends it to the large language model. The large language model corrects the entity parameters or available toolset according to the context information corresponding to the abnormal operation result, generates a new inference-action sequence and sends it to the inference-action strategy engine. The inference-action strategy engine generates a new call instruction according to the inference-action sequence and re-executes step S6.

[0061] It is understood that the method in this application scenario is similar in principle and technical effect to the technical solutions provided in any of the aforementioned method embodiments, and will not be repeated here.

[0062] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A multi-policy engine application method for an agent, characterized in that, The policy layer comprises a policy evaluation model and a policy selection model, and the method comprises, The policy evaluation model of the policy layer receives request feature data of a task request input by a user agent, outputs an evaluation result, and the policy selection model selects a target policy engine from two or more policy engines based on the evaluation result, and the target policy engine generates a calling instruction; The execution layer receives the calling instruction, executes a tool calling operation, and returns an operation result to the policy layer, and the policy layer generates task reply information according to the operation result and returns the task reply information to the user through the interface layer.

2. The method of claim 1, wherein, The request feature data comprises ambiguity of the task request, tool dependency quantity and historical execution time consumption, the policy evaluation model is constructed by using a logistic regression algorithm, and the score result is an evaluation score.

3. The method of claim 2, wherein, The policy evaluation model receives request feature data of a task request input by a user agent, and outputs an evaluation result, which comprises: The policy evaluation model multiplies the ambiguity of the task request input by the user agent, the tool dependency quantity and the historical execution time consumption respectively by corresponding preset feature weights, adds a plurality of multiplication results to obtain a total, adds a preset bias item to the total to obtain an evaluation score, and outputs the evaluation score.

4. The method of claim 2, wherein, Before the policy evaluation model receives the request feature data of the task request input by the user agent, the method further comprises: The large language model of the policy layer receives request text of the task request of the user sent by the interface layer, calculates an edit distance between the request text and a standard API description, determines the ambiguity of the task request according to the edit distance, and sends the ambiguity to the policy evaluation model; The execution layer sends historical execution time consumption of a set of available tools corresponding to the task request to the policy evaluation model.

5. The method of claim 4, wherein, The interface layer comprises a request analysis module, a service discovery module and an interface registration module, and before the large language model receives the request text of the task request of the user sent by the interface layer, the method further comprises: The request analysis module receives the task request of the user and analyzes the user intent and entity parameters; The service discovery module queries a set of available tools, tool standardized descriptions corresponding to the tools and a tool dependency quantity from the interface registration module according to the user intent, constructs the request text according to the tool standardized descriptions and the entity parameters, and sends the tool dependency quantity to the policy evaluation model and the request text and the set of available tools to the large language model respectively; wherein the interface registration module defines input and output formats of the tools according to the tool standardized descriptions, and unifies formats of the tools of the set of available tools into formats that can be recognized by the large language model.

6. The method of claim 2, wherein, The policy selection model selects a target policy engine from two or more policy engines based on the evaluation result, which comprises: If the evaluation score is greater than a preset score threshold, the policy selection model enables a reasoning-action policy engine, and if the evaluation score is less than or equal to the preset score threshold, the policy selection model enables a function calling policy engine.

7. The method of claim 1, wherein, The execution layer further comprises a tool calling module, a model context protocol server, an asynchronous execution module and a state tracking module, the execution layer receives the calling instruction, executes a tool calling operation, and returns an operation result to the policy layer, the policy layer generates task reply information according to the operation result and returns the task reply information to the user through the interface layer, which comprises: The tool calling module receives the calling instruction, and calls the asynchronous execution module; The asynchronous execution module calls the tool indicated by the calling instruction, executes the operation corresponding to the task request, generates an operation result, and returns the operation result to the target strategy engine of the strategy layer through the tool calling module; If it is judged that the operation is successful according to the operation result, the target strategy engine obtains the context information corresponding to the operation result from the model context protocol server, constructs a prompt according to the context information, and sends the prompt to the large language model; wherein the context information is constructed by the model context protocol server coordinating the task state snapshot of the state tracking module, the tool calling result of the tool calling module, and the operation history record; The large language model generates task reply information according to the prompt and returns the task reply information to the user through the interface layer.

8. The method of claim 7, wherein, The method further comprises: If it is judged that the operation is abnormal according to the operation result, the target strategy engine generates an exception signal of the abnormal operation result to the strategy selection model, the strategy selection model generates an enabling instruction of the reasoning-action strategy engine to the large language model, the large language model corrects the entity parameters or the available tool set according to the context information corresponding to the abnormal operation result, generates a new reasoning-action sequence and sends the new reasoning-action sequence to the reasoning-action strategy engine, and the reasoning-action strategy engine generates a new calling instruction according to the reasoning-action sequence.

9. The method of claim 7, wherein, The tool calling module includes a DAG scheduler, and the tool calling module receives the calling instruction, calls the asynchronous execution module, and includes: The DAG scheduler receives the calling instruction, arranges the calling order of the tool based on the directed acyclic graph carried by the calling instruction, sends an interface calling request to the interface registration module, and calls the asynchronous execution module according to the returned tool access information and the calling order, wherein the directed acyclic graph represents the dependency relationship and execution order between parallel sub-tasks.

10. A multi-policy engine application system for an agent, characterized by The system includes a strategy layer, an execution layer, and an interface layer, and the strategy layer includes a strategy evaluation model and a strategy selection model; The strategy evaluation model of the strategy layer receives the request feature data of the task request of the user input intelligent agent, outputs an evaluation result, and the strategy selection model selects a target strategy engine among two or more strategy engines based on the evaluation result, and the target strategy engine generates a calling instruction; The execution layer receives the calling instruction, executes the tool calling operation, and returns the operation result to the strategy layer, and the strategy layer generates task reply information according to the operation result and returns the task reply information to the user through the interface layer.

Images