Method and apparatus for controlled execution of an agent based on a managed terminal and a proxy gateway

Abstract

The application provides an agent controlled execution method and device based on a managed terminal and a proxy gateway, and a controlled execution method and platform for unified control of model calling and business capability calling through the proxy gateway. The application deploys agent task understanding and arrangement capabilities on the managed terminal side, converges model calling, atomic capability calling, permission verification, risk control and log auditing to the proxy gateway side, and realizes rapid landing and organization-level controllable execution of the agent platform without relying on a heavy centralized sandbox platform.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence application platforms, enterprise digital system integration, intelligent agent operation control and information security governance, and in particular to a method and device for controlled execution of intelligent agents based on managed terminals and proxy gateways. Background Technology

[0002] With the development of large-scale models and intelligent agent technology, more and more organizations hope to use intelligent agents to automate internal tasks such as expense reimbursement applications, meeting room reservations, procurement requests, information retrieval, draft generation, and approval assistance. Existing implementation paths typically fall into two categories.

[0003] The first type is the centralized intelligent agent platform solution. This type of solution typically builds a unified intelligent agent operating environment, tool invocation layer, model access layer, permission management module, and log auditing module on the server side. It can achieve good unified governance results, but it usually requires the additional construction of a sandbox environment, container isolation mechanism, task scheduling environment, and supporting computing resources. The construction cycle is long, and the implementation and operation and maintenance costs are high, which is not conducive to the rapid deployment of intelligent agent platforms in the short term.

[0004] The second category involves employees directly executing automated tasks locally using general-purpose intelligent agent tools. While this approach is simple to deploy and responds quickly, it has significant drawbacks: First, different employees may use different model exits, skill configurations, and plugin sources, making environment standardization difficult. Second, employee terminals may directly connect to external model interfaces, posing a risk of internal data leakage. Third, the calling methods of internal APIs, CLI tools, MCP tools, OA, ERP, CRM, financial systems, and form actions vary considerably, making unified encapsulation and stable invocation difficult. Fourth, the lack of unified authentication, risk control, logging, and audit playback mechanisms for the invocation process makes it difficult to meet organizational governance and compliance requirements.

[0005] Furthermore, traditional BPM or RPA platforms tend to favor fixed process orchestration, typically relying on predefined process nodes and rule configurations. This makes them ill-suited for agent execution models driven by natural language, involving dynamic task decomposition, dynamic capability selection, and context-sensitive decision-making. Therefore, existing technologies struggle to simultaneously address agent flexibility, low-cost rapid deployment capabilities, and organizational-level security governance requirements. Summary of the Invention

[0006] The purpose of this invention is to provide a method and device for controlled execution of intelligent agents based on managed terminals and proxy gateways.

[0007] To address the above problems, this invention provides a method for controlled execution of an intelligent agent based on a managed terminal and a proxy gateway, comprising:

[0008] The managed terminal obtains a trusted verification result and a list of user permissions based on the natural language task request input by the user.

[0009] Based on the trusted verification results and the user permission list, the managed terminal outputs the locally available skill set and the standardized atomic capability list;

[0010] The managed terminal generates model inference request messages and atomic capability invocation request messages based on the locally available skill set and standardized atomic capability list;

[0011] The managed terminal outputs the model execution result based on the model inference request message and the trusted verification result;

[0012] The atomic capability gateway outputs the atomic capability execution result based on the atomic capability invocation request sent by the managed terminal;

[0013] The model proxy gateway and the atomic capability gateway perform control operations on the model inference request message and the atomic capability call request message respectively, and output the control execution results.

[0014] The strategy control and audit module outputs an audit log dataset based on the model execution results, atomic capability execution results, and control execution results.

[0015] The managed terminal returns task execution results to the user based on the model execution results, atomic capability execution results, and control execution results.

[0016] Furthermore, in the above method, the managed terminal obtains a trusted verification result and a user permission list based on the user's input natural language task request, including:

[0017] The managed terminal receives a natural language task request input by the user, calls the preset organization identity authentication system, obtains and verifies the user identity information, role permission data, device trust status and session context information in the natural language task request, and outputs the trust verification result and user permission list.

[0018] Furthermore, in the above method, the managed terminal, based on the trusted verification result and the user permission list, outputs a locally available skill set and a standardized atomic capability list, including:

[0019] The managed terminal uses the root trust verification result and user permission list, combined with the task scenario and session context information corresponding to the natural language task request, to dynamically load the skill packs, task templates, and atomic capability descriptions allowed by the user from a preset organization policy server. The managed terminal parses and filters the skill packs and task templates to generate a locally available skill set. The managed terminal then standardizes and encapsulates the atomic capability descriptions to generate a standardized atomic capability list. Finally, the locally available skill set and the standardized atomic capability list are output.

[0020] Furthermore, in the above method, the managed terminal generates model inference request messages and atomic capability invocation request messages based on the locally available skill set and the standardized atomic capability list, including:

[0021] Based on the locally available skill set and standardized atomic capability list, the managed terminal performs intent recognition, task decomposition, context management and execution path planning for natural language tasks, and generates model inference request messages and atomic capability invocation request messages.

[0022] Furthermore, in the above method, the managed terminal outputs the model execution result based on the model inference request message and the trust verification result, including:

[0023] The managed terminal sends the model inference request message to the model proxy gateway. Based on the trust verification result, the model proxy gateway verifies the triple trust status of the user, device, and session. After the verification is successful, under the conditions of device trust, user trust, and session trust, it issues a short-term access token that binds the user, device, session, and time window. According to the preset organizational model scheduling policy, it forwards the model inference request message to the internal model service or the controlled access external model service, obtains the model inference result, and returns it to the managed terminal. The managed terminal outputs the model inference result as the model execution result.

[0024] Furthermore, in the above method, the atomic capability gateway outputs the atomic capability execution result based on the atomic capability invocation request sent by the managed terminal, including:

[0025] The managed terminal sends an atomic capability call request message to the atomic capability gateway. Based on the standardized atomic capability list, the atomic capability gateway maps internal API interfaces, CLI tools, MCP tools, form actions, and business system actions to standardized atomic capabilities and executes them, outputting the atomic capability execution results.

[0026] Furthermore, in the above method, the model proxy gateway and the atomic capability gateway perform control operations on the model inference request message and the atomic capability invocation request message, respectively, and output the control execution results, including:

[0027] Based on organizational policies, the model proxy gateway and atomic capability gateway simultaneously perform parameter validity verification, user role and permission verification, operation risk level identification, sensitive data anonymization, interface rate limiting and circuit breaking, and high-risk action approval interception or secondary confirmation on model inference request messages and atomic capability call request messages, respectively, and output the control execution results.

[0028] Furthermore, in the above method, the policy control and audit module outputs an audit log dataset based on the model execution results, atomic capability execution results, and control execution results, including:

[0029] Based on the model execution results, atomic capability execution results, and control execution results, the strategy control and audit module performs full-field logging, audit tag marking, and persistent storage of execution traces for model calls, atomic capability calls, user confirmation operations, approval operations, and execution results, and outputs an audit log dataset.

[0030] Furthermore, in the above method, the managed terminal returns task execution results to the user based on the model execution results, atomic capability execution results, and control execution results, including:

[0031] The managed terminal receives the model execution results, atomic capability execution results, and control execution results, integrates them, and returns the task execution results to the user. The task execution results include three execution states: task completion, failure interruption, and risk blocking. For the three states of task completion, failure interruption, or risk blocking, corresponding actions such as result display, failure recovery, or risk warning are triggered, and execution replay, responsibility attribution, and process optimization are supported based on the audit log dataset.

[0032] According to another aspect of the present invention, a computer-readable storage medium is also provided, having stored thereon computer-executable instructions, wherein when executed by a processor, the computer-executable instructions cause the processor to perform the method as described in any of the preceding claims.

[0033] This invention relates to a controlled execution method and platform for running intelligent agents within an organization via managed terminals and uniformly controlling model invocation and business capability invocation through a proxy gateway. By deploying intelligent agent task understanding and orchestration capabilities on the managed terminal side, while converging model invocation, atomic capability invocation, permission verification, risk control, and log auditing to the proxy gateway side, this invention achieves rapid deployment and organization-level controllable execution of the intelligent agent platform without relying on a heavy-duty centralized sandbox platform.

[0034] This invention can solve the following problems existing in the prior art:

[0035] 1. Centralized intelligent agent platforms have high construction costs and long implementation cycles, which are not conducive to rapid deployment in the short term.

[0036] 2. In the local intelligent agent execution mode, the model exit is uncontrollable, the business capability call is uncontrollable, and the log audit is uncontrollable, which poses data security and governance risks.

[0037] 3. The lack of a unified abstraction for heterogeneous business capabilities within the organization makes it difficult for intelligent agents to reliably invoke real business actions in different systems.

[0038] 4. Existing process systems are unable to meet the execution requirements of natural language-driven, dynamic programming, and risk boundary control.

[0039] Compared with the prior art, the present invention has at least the following beneficial effects:

[0040] 1. This invention deploys the intelligent agent task understanding and orchestration capabilities on the terminal side, and converges the high-value resource access and security governance capabilities on the gateway side. It enables the rapid deployment of the intelligent agent platform without relying on a heavy centralized sandbox platform, thereby reducing the overall construction cost and deployment cycle.

[0041] 2. This invention effectively restricts terminal agents to complete inference requests only through model paths permitted by the organization by combining a mechanism of "managed terminal + short-term token + model proxy gateway + network egress whitelist + device trust verification", thereby reducing the risk of data outflow and bypassing governance.

[0042] 3. This invention unifies and abstracts internal APIs, CLIs, MCPs, form actions, and business system actions into standardized atomic capabilities, improving the feasibility and reusability of heterogeneous business capabilities being stably invoked by intelligent agents.

[0043] 4. This invention encapsulates capability call rules, parameter templates, and exception handling experience in a skill-based manner, enabling personal experience to be accumulated into auditable, publishable, and reusable process assets.

[0044] 5. This invention enables intelligent agents to have dynamic planning capabilities while still meeting the organization's boundary control requirements for critical write operations through risk level classification, secondary confirmation, and approval interception mechanisms.

[0045] 6. This invention establishes a full-link log and audit playback system from terminals, models, capabilities to business systems, transforming the intelligent agent execution process from a black-box question-and-answer process into a traceable, explainable, replayable, and accountable business execution chain. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of a controlled execution method and platform for intelligent agents based on managed terminals and proxy gateways according to an embodiment of the present invention.

[0047] Figure 2 This is a flowchart of a method for controlled execution of an intelligent agent based on a managed terminal and a proxy gateway, according to an embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of a model exit control mechanism according to an embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram of an embodiment of the present invention. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] like Figures 1 to 4 As shown, the present invention provides a method for controlled execution of an intelligent agent based on a managed terminal and a proxy gateway, comprising:

[0052] Step S1: The managed terminal obtains a trusted verification result and a list of user permissions based on the natural language task request input by the user.

[0053] Here, task reception and identity trust verification: The managed terminal receives the natural language task request input by the user, calls the preset organization identity authentication system, obtains and verifies the user identity information, role permission data, device trust status and session context information in the natural language task request, and outputs the trust verification result and user permission list.

[0054] Step S2: Based on the trusted verification result and the user permission list, the managed terminal outputs the locally available skill set and the standardized atomic capability list.

[0055] Here, skills and capabilities are dynamically loaded: Based on the trusted verification result and user permission list output in step 1, and combined with the task scenario and session context information corresponding to the natural language task request, the managed terminal dynamically loads the skill packages, task templates, and atomic capability descriptions allowed by the user from a preset organization policy server. The managed terminal parses and filters the skill packages and task templates to generate a locally available skill set. The managed terminal standardizes and encapsulates the atomic capability descriptions to generate a standardized atomic capability list. Finally, the locally available skill set and the standardized atomic capability list are output.

[0056] Step S3: The managed terminal generates a model inference request message and an atomic capability invocation request message based on the locally available skill set and the standardized atomic capability list.

[0057] Here, task understanding and execution orchestration: Based on the locally available skill set and standardized atomic capability list output in step 2, the managed terminal performs intent recognition, task decomposition, context management and execution path planning for the natural language task, and generates model inference request messages and atomic capability call request messages;

[0058] Step S4: The managed terminal outputs the model execution result based on the model inference request message and the trusted verification result;

[0059] Here, the model invokes proxy execution: The managed terminal sends the model inference request message output in step 3 to the model proxy gateway. Based on the trust verification result in step 1, the model proxy gateway verifies the triple trust status of the user, device, and session. After the verification is passed, under the conditions of device trust, user trust, and session trust, it issues a short-term access token that binds the user, device, session, and time window. According to the preset organizational model scheduling strategy, it forwards the model inference request message to the internal model service or the controlled access external model service, obtains the model inference result, and returns it to the managed terminal. The managed terminal outputs the model inference result as the model execution result.

[0060] Step S5: The atomic capability gateway outputs the atomic capability execution result based on the atomic capability invocation request sent by the managed terminal.

[0061] Here, business capabilities are executed uniformly: the managed terminal sends the atomic capability call request message output in step 3 to the atomic capability gateway. The atomic capability gateway maps internal API interfaces, CLI tools, MCP tools, form actions and business system actions to standardized atomic capabilities based on the standardized atomic capability list in step 2 and executes them, outputting the atomic capability execution results.

[0062] Step S6: The model proxy gateway and the atomic capability gateway perform control operations on the model inference request message and the atomic capability call request message from step 3, respectively, and output the control execution results.

[0063] Here, the entire process of risk and access control is implemented: Based on organizational policies, the model proxy gateway and atomic capability gateway simultaneously perform parameter validity verification, user role and access control verification, operation risk level identification, sensitive data desensitization, interface rate limiting and circuit breaking, and high-risk action approval interception or secondary confirmation on the model inference request message and atomic capability call request message in step 3, and output the control execution results.

[0064] Here, at the model proxy gateway and atomic capability gateway, request execution parameter verification, permission verification, risk level identification, data anonymization, rate limiting and circuit breaking, approval interception or secondary confirmation are performed.

[0065] Step S7: Based on the model execution results, atomic capability execution results, and control execution results, the policy control and audit module outputs an audit log dataset.

[0066] Here, the full-link audit log is recorded: Based on the model execution results of step 4, the atomic capability execution results of step 5, and the control execution results of step 6, the policy control and audit module performs full-field log recording, audit tag marking, and persistent storage of execution trajectory for model calls, atomic capability calls, user confirmation operations, approval operations, and execution results, and outputs an audit log dataset;

[0067] In step S8, the managed terminal returns the task execution result to the user based on the model execution result, the atomic capability execution result, and the control execution result.

[0068] Here, the result return and execution loop is completed: the managed terminal receives the model execution result from step 4, the atomic capability execution result from step 5, and the control execution result from step 6, integrates them, and returns the task execution result to the user; the task execution result includes three execution states: task completion, failure interruption, and risk blocking; for the three states of task completion, failure interruption, or risk blocking, corresponding result display, failure recovery, or risk warning operations are triggered, and execution replay, responsibility attribution, and process optimization are supported based on the audit log dataset from step 7.

[0069] Furthermore, the managed terminal is a controlled operating environment uniformly distributed by the organization, and has the following characteristics:

[0070] Only skills, templates, plugins, and atomic capability descriptions approved by the organization are allowed to be loaded; loading unauthorized external components is prohibited.

[0071] Model configuration, permission rules, gateway address, and network policy are all uniformly distributed by the organization and cannot be manually modified locally.

[0072] The managed terminal memory only temporarily caches session-level short-term tokens and does not store model service keys or high-privilege business credentials for long periods of time.

[0073] Managed terminals only handle lightweight logic for task understanding, process orchestration, and result display, and do not directly access sensitive model services and core business systems.

[0074] Furthermore, the model proxy gateway is used to unify the convergence of model outputs and has the following characteristics:

[0075] Managed endpoints are only allowed to access the model proxy gateway address specified by the organization;

[0076] The underlying internal model service and external model service are only exposed to the model proxy gateway and are not directly exposed to the managed terminals;

[0077] Short-term access tokens are used instead of long-term keys, and the token's validity period does not exceed the duration of the current session.

[0078] The token is strongly bound to the user ID, device unique identifier, operator instance ID, session ID, and time window to prevent token theft.

[0079] By combining network egress whitelist, device trust verification, and session state verification, managed terminals are blocked from bypassing the model proxy gateway and directly calling authorized model services.

[0080] Furthermore, the atomic capability gateway is used to unify and abstract heterogeneous service capabilities, and each standardized atomic capability contains the following metadata:

[0081] Capability name, capability description, input parameter schema, and output result schema;

[0082] Required role permissions, operation risk level, whether it is a read-only operation, and whether it involves a real commit;

[0083] Whether approval or secondary confirmation is required, whether the process can be rolled back, and audit labels.

[0084] Furthermore, the locally available skill set is a standardized encapsulation of atomic capability invocation rules and process experience, including the following:

[0085] Atomic capability retrieval matching rules, parameter input examples, and validation rules;

[0086] Applicable business process templates, exception handling strategies, and risk boundary descriptions;

[0087] Perform audit requirements, log fields, and compliance verification rules.

[0088] Furthermore, the method constructs an independent execution context for each task instance, which includes at least: an independent session identifier, an independent task state, an independent temporary variable, an independent cache space, an independent access token, and an independent capability call record, in order to avoid context pollution across tasks, users, and skills.

[0089] Furthermore, the method sets automation boundaries based on the risk level of atomic capabilities. Query-type capabilities and draft generation capabilities can be executed automatically, while high-risk capabilities involving approval submission, contract sending, financial writing, or other status changes require confirmation or approval, thereby achieving a combination of dynamic planning of the agent and controlled execution of key actions.

[0090] The present invention also provides a controlled execution platform for intelligent agents based on managed terminals and proxy gateways, the platform comprising:

[0091] 1. Managed terminal intelligent agent operator, used to receive user tasks, dynamically load skill and atomic capability descriptions, execute task understanding and local orchestration;

[0092] 2. Identity and Device Trust Module, used to identify user identity, device trust status, operating environment status, and access policy;

[0093] 3. Model proxy gateway, used to uniformly proxy internal model services and controlled access to external model services;

[0094] 4. Atomic capability gateway, used for unified encapsulation and execution of heterogeneous business capabilities within an organization;

[0095] 5. The policy control and audit module is used for access control, risk classification, approval interception, logging, execution replay, and responsibility attribution.

[0096] Figure 2 This is a flowchart of a method for controlled execution of an intelligent agent based on a managed terminal and a proxy gateway, according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the model exit control mechanism according to an embodiment of the present invention; Figure 4 This is a schematic diagram of an embodiment of the present invention.

[0097] The present invention will be further described below with reference to specific embodiments, but the present invention is not limited to the following embodiments.

[0098] Example 1: Draft generation and controlled submission of expense reports:

[0099] The user enters "Create an expense report for me based on last week's business trip records and save a draft" on the managed terminal. The managed terminal first verifies that the user is logged into the organization's identity system and that the current terminal is a managed device through the identity and device trust module. Subsequently, based on the user's authentication results and the task context, the managed terminal dynamically loads descriptions of expense reimbursement-related skills, travel inquiry capabilities, invoice organization capabilities, expense report creation capabilities, and draft saving capabilities.

[0100] The managed terminal understands the user's task and breaks it down into four steps: "querying travel records," "organizing invoice fields," "creating a draft expense report," and "saving the draft." It then sends the relevant inference requests to the model proxy gateway. After verifying the trustworthiness of the user, device, and session, the model proxy gateway issues a short-term token and forwards the inference request to the organization-approved model service. Based on the steps and parameters returned by the model, the managed terminal then sequentially initiates business requests to the atomic capability gateway. The atomic capability gateway maps the travel system interface, invoice processing logic, and expense report form actions to standardized atomic capability execution.

[0101] If the user further enters "Submit for Approval," the atomic capability gateway identifies "Approval Submission" as a high-risk capability, triggering secondary confirmation or approval interception. The submission action is only executed after the user confirms or the approval conditions are met. Throughout the process, model call records, capability call parameters, confirmation behavior, and final results are all recorded by the policy control and audit module for subsequent playback and traceability.

[0102] Example 2: Meeting room reservation and outgoing notification control:

[0103] The user enters "Reserve an 8-person meeting room for tomorrow afternoon from 2 PM to 3 PM, and invite members of R&D Team 2" on the managed terminal. The managed terminal dynamically loads descriptions of meeting room query capabilities, schedule conflict checking capabilities, meeting creation capabilities, and notification draft capabilities. The system first calls the model proxy gateway to obtain the task breakdown results, and then calls the atomic capability gateway to execute the meeting room query, idle time period judgment, and meeting creation actions.

[0104] When the system needs to send a notification to participants, the strategy control module determines its risk level based on rules such as the scope of the notification recipients, whether it crosses departments, and whether it involves external transmission. If it is a normal internal notification, a draft notification can be automatically generated and displayed to the user for confirmation; if it involves a higher-risk transmission scenario, the user is required to confirm again or an approval process is triggered. This embodiment illustrates that the present invention can not only perform query and creation actions, but also control the subsequent transmission actions based on risk boundaries.

[0105] Example 3: Automatic generation of purchase requisitions:

[0106] The user enters "Initiate a standard office equipment procurement request for the new employee" on the managed terminal. Based on the user's department, role permissions, and task context, the managed terminal loads the procurement template skills, standard material capability descriptions, budget verification capability descriptions, and procurement request creation capability descriptions. The terminal completes the procurement task planning steps through the model proxy gateway, and then queries the standard office equipment template, completes the procurement fields, and generates a procurement draft through the atomic capability gateway.

[0107] If the purchase amount, purchase category, or budget status triggers preset risk rules, the strategy control and audit module can require the user to provide supplementary information, execute confirmation, or enter the approval process. The entire procurement process, including task inputs, model versions, skill versions, atomic capability call results, and approval status, is recorded throughout the entire process.

[0108] Example 4: Task-level independent context execution:

[0109] In another embodiment, the present invention creates independent execution contexts for different users, different tasks, and different skill instances. Each execution context has an independent session identifier, independent cache, independent access token, and independent call record. Even if multiple users execute similar tasks simultaneously on different terminals, temporary variables, intermediate results of capability calls, or high-privilege credentials will not be shared between tasks, thereby avoiding context pollution and unauthorized propagation.

[0110] Example 5: Controlled execution of model exit:

[0111] In another implementation, the network only allows managed terminals to access designated model proxy gateway addresses and blocks terminals from directly accessing unauthorized model domain names or IPs. The model proxy gateway issues short-term tokens only if the device, user, and session are trusted. Even if a terminal copies the configuration file, it cannot directly access the underlying model services if the device is untrusted, the token has expired, or the session state is abnormal. This implementation elevates the model exit control of the terminal agent from configuration constraints to technical control through the combined action of gateways, tokens, and network policies.

[0112] This invention relates to a controlled execution method and platform for running intelligent agents within an organization via managed terminals and uniformly controlling model invocation and business capability invocation through a proxy gateway. By deploying intelligent agent task understanding and orchestration capabilities on the managed terminal side, while converging model invocation, atomic capability invocation, permission verification, risk control, and log auditing to the proxy gateway side, this invention achieves rapid deployment and organization-level controllable execution of the intelligent agent platform without relying on a heavy-duty centralized sandbox platform.

[0113] This invention can solve the following problems existing in the prior art:

[0114] 1. Centralized intelligent agent platforms have high construction costs and long implementation cycles, which are not conducive to rapid deployment in the short term.

[0115] 2. In the local intelligent agent execution mode, the model exit is uncontrollable, the business capability call is uncontrollable, and the log audit is uncontrollable, which poses data security and governance risks.

[0116] 3. The lack of a unified abstraction for heterogeneous business capabilities within the organization makes it difficult for intelligent agents to reliably invoke real business actions in different systems.

[0117] 4. Existing process systems are unable to meet the execution requirements of natural language-driven, dynamic programming, and risk boundary control.

[0118] Compared with the prior art, the present invention has at least the following beneficial effects:

[0119] 1. This invention deploys the intelligent agent task understanding and orchestration capabilities on the terminal side, and converges the high-value resource access and security governance capabilities on the gateway side. It enables the rapid deployment of the intelligent agent platform without relying on a heavy centralized sandbox platform, thereby reducing the overall construction cost and deployment cycle.

[0120] 2. This invention effectively restricts terminal agents to complete inference requests only through model paths permitted by the organization by combining a mechanism of "managed terminal + short-term token + model proxy gateway + network egress whitelist + device trust verification", thereby reducing the risk of data outflow and bypassing governance.

[0121] 3. This invention unifies and abstracts internal APIs, CLIs, MCPs, form actions, and business system actions into standardized atomic capabilities, improving the feasibility and reusability of heterogeneous business capabilities being stably invoked by intelligent agents.

[0122] 4. This invention encapsulates capability call rules, parameter templates, and exception handling experience in a skill-based manner, enabling personal experience to be accumulated into auditable, publishable, and reusable process assets.

[0123] 5. This invention enables intelligent agents to have dynamic planning capabilities while still meeting the organization's boundary control requirements for critical write operations through risk level classification, secondary confirmation, and approval interception mechanisms.

[0124] 6. This invention establishes a full-link log and audit playback system from terminals, models, capabilities to business systems, transforming the intelligent agent execution process from a black-box question-and-answer process into a traceable, explainable, replayable, and accountable business execution chain.

[0125] Those skilled in the art will understand that, without departing from the core concept of this invention, modifications or replacements can be made to the managed terminal form, model gateway deployment method, atomic capability description format, approval strategy, and audit field, all of which should fall within the protection scope of this invention.

[0126] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0127] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0128] Obviously, those skilled in the art can make various modifications and variations to the invention without departing from the spirit and scope of the invention. Therefore, if these modifications and variations fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A method for controlled execution of an intelligent agent based on a managed terminal and a proxy gateway, characterized in that, include: The managed terminal obtains a trusted verification result and a list of user permissions based on the natural language task request input by the user. Based on the trusted verification results and the user permission list, the managed terminal outputs the locally available skill set and the standardized atomic capability list; The managed terminal generates model inference request messages and atomic capability invocation request messages based on the locally available skill set and standardized atomic capability list; The managed terminal outputs the model execution result based on the model inference request message and the trusted verification result; The atomic capability gateway outputs the atomic capability execution result based on the atomic capability invocation request sent by the managed terminal; The model proxy gateway and the atomic capability gateway perform control operations on the model inference request message and the atomic capability call request message respectively, and output the control execution results. The strategy control and audit module outputs an audit log dataset based on the model execution results, atomic capability execution results, and control execution results. The managed terminal returns task execution results to the user based on the model execution results, atomic capability execution results, and control execution results.

2. The controlled execution method for intelligent agents based on managed terminals and proxy gateways as described in claim 1, characterized in that, Based on the user's input of a natural language task request, the managed terminal obtains a trusted verification result and a list of user permissions, including: The managed terminal receives a natural language task request input by the user, calls the preset organization identity authentication system, obtains and verifies the user identity information, role permission data, device trust status and session context information in the natural language task request, and outputs the trust verification result and user permission list.

3. The controlled execution method for intelligent agents based on managed terminals and proxy gateways as described in claim 1, characterized in that, Based on the trusted verification results and the user permission list, the managed terminal outputs a locally available skill set and a standardized atomic capability list, including: The managed terminal uses the root trust verification result and user permission list, combined with the task scenario and session context information corresponding to the natural language task request, to dynamically load the skill packs, task templates, and atomic capability descriptions allowed by the user from a preset organization policy server. The managed terminal parses and filters the skill packs and task templates to generate a locally available skill set. The managed terminal then standardizes and encapsulates the atomic capability descriptions to generate a standardized atomic capability list. Finally, the locally available skill set and the standardized atomic capability list are output.

4. The controlled execution method for intelligent agents based on managed terminals and proxy gateways as described in claim 1, characterized in that, The managed terminal generates model inference request messages and atomic capability invocation request messages based on the locally available skill set and standardized atomic capability list, including: Based on the locally available skill set and standardized atomic capability list, the managed terminal performs intent recognition, task decomposition, context management and execution path planning for natural language tasks, and generates model inference request messages and atomic capability invocation request messages.

5. The controlled execution method for intelligent agents based on managed terminals and proxy gateways as described in claim 1, characterized in that, Based on the model inference request message and the trust verification result, the managed terminal outputs the model execution result, including: The managed terminal sends the model inference request message to the model proxy gateway. Based on the trust verification result, the model proxy gateway verifies the triple trust status of the user, device, and session. After the verification is successful, under the conditions of device trust, user trust, and session trust, it issues a short-term access token that binds the user, device, session, and time window. According to the preset organizational model scheduling policy, it forwards the model inference request message to the internal model service or the controlled access external model service, obtains the model inference result, and returns it to the managed terminal. The managed terminal outputs the model inference result as the model execution result.

6. The controlled execution method for intelligent agents based on managed terminals and proxy gateways as described in claim 1, characterized in that, The atomic capability gateway outputs the atomic capability execution results based on the atomic capability invocation requests sent by the managed terminals, including: The managed terminal sends an atomic capability call request message to the atomic capability gateway. Based on the standardized atomic capability list, the atomic capability gateway maps internal API interfaces, CLI tools, MCP tools, form actions, and business system actions to standardized atomic capabilities and executes them, outputting the atomic capability execution results.

7. The controlled execution method for intelligent agents based on managed terminals and proxy gateways as described in claim 1, characterized in that, The model proxy gateway and atomic capability gateway perform control operations on model inference request messages and atomic capability invocation request messages, respectively, and output the control execution results, including: Based on organizational policies, the model proxy gateway and atomic capability gateway simultaneously perform parameter validity verification, user role and permission verification, operation risk level identification, sensitive data anonymization, interface rate limiting and circuit breaking, and high-risk action approval interception or secondary confirmation on model inference request messages and atomic capability call request messages, respectively, and output the control execution results.

8. The controlled execution method for intelligent agents based on managed terminals and proxy gateways as described in claim 1, characterized in that, The strategy control and audit module outputs an audit log dataset based on the model execution results, atomic capability execution results, and control execution results, including: Based on the model execution results, atomic capability execution results, and control execution results, the strategy control and audit module performs full-field logging, audit tag marking, and persistent storage of execution traces for model calls, atomic capability calls, user confirmation operations, approval operations, and execution results, and outputs an audit log dataset.

9. The controlled execution method for intelligent agents based on managed terminals and proxy gateways as described in claim 1, characterized in that, Based on the model execution results, atomic capability execution results, and control execution results, the managed terminal returns task execution results to the user, including: The managed terminal receives the model execution results, atomic capability execution results, and control execution results, integrates them, and returns the task execution results to the user. The task execution results include three execution states: task completion, failure interruption, and risk blocking. For the three states of task completion, failure interruption, or risk blocking, corresponding actions such as result display, failure recovery, or risk warning are triggered, and execution replay, responsibility attribution, and process optimization are supported based on the audit log dataset.

10. A computer-readable storage medium having stored thereon computer-executable instructions, wherein, When the computer-executable instructions are executed by the processor, the processor causes the processor to perform the method as described in any one of claims 1 to 9.

Images