Multi-agent management system, management method, and storage medium

Abstract

This invention discloses a multi-agent management system, management method, and storage medium. The system includes an agent metadata manager, a task sandbox scheduler, a policy compliance verifier, and a full-link audit and circuit breaker controller. The agent metadata manager manages the metadata of the multiple agents. The task sandbox scheduler creates resource-isolated sandbox execution environments for each task and activates the required business agents within the sandbox as needed. The policy compliance verifier performs permission checks and policy checks on operation requests, with the policy checks verifying whether the operation content matches a preset security policy library. The full-link audit and circuit breaker controller records the process logs of each task and includes a circuit breaker API. This invention improves the granularity of multi-agent control, makes the decision-making process transparent, eliminates concurrent state pollution, and avoids erroneous execution.

Description

Technical Field

[0001] This invention relates to the field of multi-agent management technology, specifically to a multi-agent management system, management method, and storage medium. Background Technology

[0002] Currently, in complex business scenarios (such as cybersecurity operations and financial risk control), a single large model or program often struggles to balance professionalism, flexibility, and efficiency. A common approach is to break down the overall task into multiple sub-tasks and deploy multiple "agents" with specific functions to handle them separately. Through their collaboration, comprehensive tasks that are difficult for a single program to process can be accomplished. These agents are typically designed to have autonomous decision-making capabilities, enabling them to perceive the environment, process information, and take action.

[0003] In actual system architectures, these agents may run as microservices, containerized processes, or independent processes. Their collaboration primarily follows two patterns: 1. Direct Invocation Mode: After completing its own task, an agent directly invokes the service interface of the next agent through a network request, forming a linear or mesh-like task chain.

[0004] 2. Message-driven mode: Agents publish tasks or results to a shared message queue, and other interested agents subscribe to the corresponding topics and process them, forming a loosely coupled asynchronous collaboration.

[0005] Regardless of the collaboration model employed, the core idea of ​​the system architecture is the decomposition and combination of functions. Designers focus on how to efficiently orchestrate the workflows of these agents to achieve end-to-end task objectives. The entire system operates on an implicit but crucial assumption: that each agent is trustworthy and its behavior predictable, its internal decision-making logic sound, and its output reliable.

[0006] Against this backdrop, current system-level designs almost entirely neglect proactive governance of the runtime behavior of multiple agents. There is no independent, high-privilege component to enforce constraints on communication methods between agents, isolate their execution environments, verify the security of their operational instructions, or fully record their decision-making processes. This "execution-heavy, governance-light" design philosophy may be feasible in laboratory or low-risk scenarios, but it poses a significant hidden danger in production environments that demand high reliability, high security, and strong auditing capabilities.

[0007] In summary, while the current architecture can achieve basic functional collaboration, it exposes serious governance deficiencies when facing production environments with high reliability and security requirements, specifically in the following four aspects: 1. Inability to precisely stop: When an agent erroneously behaves due to model illusion or input contamination, the lack of a unified control plane prevents administrators from precisely pausing, terminating, or rolling back the specific task flow being executed by the agent at runtime. Often, the only option is to resort to the crude method of restarting the entire service, which affects the overall availability of the system.

[0008] 2. Untraceable Decision-Making: The collaborative process between intelligent agents is dynamic and distributed. Once an error occurs in the final output, it is difficult to fully reconstruct the complete causal chain that led to the result, including the data, internal states, and interaction order of each intelligent agent involved in the decision-making process at that time. This creates a "black box effect," which brings great difficulties to fault diagnosis and responsibility determination.

[0009] 3. State Contamination: In high-concurrency scenarios, multiple user requests may simultaneously trigger the same type of agent. If the agent uses non-thread-safe global variables or shared caches to store the context when processing requests, the processing states of different requests will interfere with each other, leading to distorted output or data leakage.

[0010] 4. Stubbornly Adhering to Errors: When multiple agents reach an erroneous consensus due to shared incorrect input (such as adversarial prompts), existing architectures lack a built-in "circuit breaker" mechanism. The system faithfully translates this erroneous consensus into actual operational instructions and executes them to the end, potentially amplifying risks and causing serious consequences due to automated cascading reactions. Summary of the Invention

[0011] To address the aforementioned problems, this invention provides a multi-agent management system, management method, and storage medium, aiming to systematically solve the core issues existing in current multi-agent management processes, such as insufficient control granularity, opaque decision-making processes, concurrent state contamination, and lack of security fallback mechanisms.

[0012] According to a first aspect of the present disclosure, a multi-agent management system is provided, the system comprising an agent metadata manager, a task sandbox scheduler, a policy compliance verifier, and a full-link audit and circuit breaker controller, wherein... The agent metadata manager is used to manage the metadata of multiple agents. The metadata includes any one or more of the following: agent ID, role definition, capability declaration, permission boundary, and interface specification. The task sandbox scheduler creates a sandbox execution environment isolated from resources for each task and activates the required business intelligence agents within the sandbox as needed. The sandbox execution environment includes independent memory space, file system, network stack and task ID. The policy compliance verifier performs permission checks and policy checks on operation requests. The permission check is used to check whether the agent initiating the operation request is within the registered permission boundaries; the policy check is used to check whether the operation content matches the preset security policy library. The end-to-end audit and circuit breaker controller is used to record the process logs of each task and has a circuit breaker API.

[0013] A further technical solution of the present invention is as follows: after receiving a task request, the task sandbox scheduler parses the task objective and determines the combination of intelligent agents required to complete the task based on the information in the intelligent agent metadata manager.

[0014] A further technical solution of the present invention is as follows: the end-to-end audit and circuit breaker controller records the process logs of each task, including at least one of the following (1) to (4): (1) The original input of the task; (2) Each agent involved and its state snapshot at the start of the task; (3) The message content transmitted between the agents; (4) All intercepted or executed operation instructions and their verification results.

[0015] According to a second aspect of the present disclosure, a multi-agent management method is provided, comprising the following steps: The metadata of multiple agents is managed using an agent metadata manager. The metadata includes any one or more of the following: agent ID, role definition, capability declaration, permission boundary, and interface specification. The task sandbox scheduler creates a sandbox execution environment isolated from resources for each task and activates the required business intelligence agents on demand within the sandbox. The sandbox execution environment includes independent memory space, file system, network stack and task ID. The policy compliance verifier performs permission checks and policy checks on operation requests. The permission check is used to check whether the agent initiating the operation request is within the registered permission boundaries; the policy check is used to check whether the operation content matches the preset security policy library. The system utilizes end-to-end auditing and circuit breaker controllers to record process logs for each task and includes a circuit breaker API.

[0016] A further technical solution of the present invention is as follows: the method further includes receiving a task request using a task sandbox scheduler, parsing the task objective, and determining the combination of intelligent agents required to complete the task based on the information in the intelligent agent metadata manager.

[0017] A further technical solution of the present invention is as follows: the method further includes recording the process logs of each task using a full-link audit and circuit breaker controller, including at least one of the following (1) to (4): (1) The original input of the task; (2) Each agent involved and its state snapshot at the start of the task; (3) The message content transmitted between the agents; (4) All intercepted or executed operation instructions and their verification results.

[0018] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the multi-agent management method described above.

[0019] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, the storage medium storing computer instructions that, when executed by a processor, implement the steps of the multi-agent management method described above.

[0020] The multi-agent management system, management method, and storage medium provided in this disclosure have the following advantages: 1. Achieved precise halting: Through TaskID and the circuit breaker API, any task can be intervened at the second level, with control granularity accurate to a single task.

[0021] 2. Ensures traceability of decisions: The end-to-end audit log indexed by TaskID provides a complete and tamper-proof snapshot of decisions, achieving true "white-box" governance.

[0022] 3. Eliminates state contamination: The task-based sandbox isolation mechanism ensures that even if the same agent is reused by multiple tasks, its internal state remains absolutely pure and does not interfere with each other.

[0023] 4. Avoids erroneous execution: Pre-policy verification forms the first line of defense, and the circuit breaker controller provides a second means of manual intervention. This double insurance ensures that the system will not blindly execute erroneous instructions.

[0024] In summary, this invention provides a solid and reliable governance foundation for the large-scale and secure implementation of multi-agent systems.

[0025] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0027] Figure 1This is a structural diagram of the multi-agent management system in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the implementation of the multi-agent management system in this embodiment of the invention. Figure 3 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0028] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present invention are shown in the drawings, not the entire structure.

[0029] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. The process can be terminated when its operation is complete, but may also have additional steps not included in the figures. The process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0030] At the heart of this invention is a standalone system component called the Agent Governance Hub (AGH). The AGH is not an ordinary business agent, but rather serves as the sole trusted execution path and the highest-authority manager for the entire multi-agent system. All business agents (regardless of their number, type, or architecture) must register, communicate, and execute operations through the AGH.

[0031] The embodiments relate to a multi-agent management system, such as Figure 1 As shown, system 100 includes an intelligent agent metadata manager 110, a task sandbox scheduler 120, a policy compliance verifier 130, and a full-link audit and circuit breaker controller 140, wherein, The agent metadata manager 110 is used to manage the metadata of multiple agents, which includes any one or more of agent ID, role definition, capability declaration, permission boundary, and interface specification; The task sandbox scheduler 120 creates a sandbox execution environment isolated from resources for each task and activates the required business intelligence agents within the sandbox as needed. The sandbox execution environment includes independent memory space, file system, network stack and task ID. The policy compliance verifier 130 performs permission checks and policy checks on operation requests. The permission check is used to check whether the agent that initiates the operation request is within the registered permission boundaries; the policy check is used to check whether the operation content matches the preset security policy library. The end-to-end audit and circuit breaker controller 140 is used to record the process logs of each task and has a circuit breaker API.

[0032] After receiving a task request, the task sandbox scheduler 120 parses the task objective and determines the combination of agents required to complete the task based on the information in the agent metadata manager.

[0033] The end-to-end audit and circuit breaker controller 140 records the process logs of each task, including at least one of the following (1) to (4): (1) The original input of the task; (2) Each agent involved and its state snapshot at the start of the task; (3) The message content transmitted between the agents; (4) All intercepted or executed operation instructions and their verification results.

[0034] Specifically, such as Figure 2 As shown, taking two intelligent agents A and B as an example, the system workflow diagram is illustrated. The workflow of this invention strictly follows the closed-loop principle of "registration, scheduling, execution, verification, and auditing," as detailed below: Step S1: Forced registration and metadata binding of intelligent agents All business agents must initiate a registration request to AGH's agent metadata manager upon startup.

[0035] Registration information includes: a unique agent ID, role definition (such as "intelligence analyst" or "executor"), capability declaration (such as "accessible database X"), permission boundaries (such as "read-only, not writable"), and interface specifications.

[0036] AGH assigns an internal credential to each successfully registered agent and persists its information. Unregistered agents are completely isolated by the system and cannot participate in any tasks. The internal credential is used for agent operation request authentication / signing. Operation requests must include at least AgentID, TaskID, and the credential / signature. The verifier verifies the signature first and then performs permission boundary and policy matching. Step S2: Task Submission and Sandbox Environment Creation When an external system or user needs to perform a task, submit a standardized task request to AGH.

[0037] After receiving a request, the AGH task sandbox scheduler parses the task objective and determines the combination of agents required to complete the task based on the information in the metadata manager.

[0038] The scheduler creates a completely new, resource-isolated sandbox execution environment for this task. This sandbox has its own memory space, file system, and network stack, and is assigned a globally unique task ID (TaskID).

[0039] Step S3: Agent activation and restricted communication within the sandbox The scheduler activates the required business agents on demand within the sandbox. Each activated agent instance is injected with the current TaskID, making all its internal logs and states strongly associated with that ID.

[0040] Key constraints: Direct communication between agents within the sandbox, and between agents and the outside world, is prohibited in any form. All information exchange (including message passing between agents, calls to external APIs, and reading and writing to the database) must be closed. In specific implementation, outbound requests from agents must pass through AGH's forwarding / proxy components (such as gateway / proxy services); all messages pass through the "message forwarding module / bus" and carry the TaskID.

[0041] Step S4: Pre-implementation and execution of operation instructions When an agent in the sandbox generates an operation request (e.g., "Execute SQL: DELETE FROM users"), the request is first sent to AGH's policy compliance validator.

[0042] The verifier performs a double check: Permission check: Verify whether the agent that initiated the request is within its registered permission boundaries.

[0043] Policy check: Matches the operation content against the preset security policy library (e.g., "prohibit the execution of any DELETE statement").

[0044] Only operation requests that pass both checks simultaneously will be executed by AGH on behalf of the agent in the real environment. If either check fails, the operation will be immediately intercepted, and an alarm log containing the complete context will be generated. Specifically, the specific fields of the alarm log with the complete context include TaskID, AgentID, operation content, hit rule / failure reason, timestamp, etc.

[0045] In some specific implementation processes, the process by which the scheduler / verifier obtains metadata is as follows: agent registration → metadata manager writes to "metadata storage" (registration center / configuration center) → scheduler / verifier queries as needed. For example, when "parsing tasks and determining agent combinations", a query is performed using TaskID + required capabilities / roles to obtain fields such as agent version and capabilities. Then, each time an "operation request" enters the verifier, the fields are queried for compliance using agent_id + task_id.

[0046] Step S5: End-to-end audit logging and real-time circuit breaker AGH's end-to-end auditing and circuit breaker controller are implemented throughout all the above steps. It mandates the logging of the following information, forming an immutable audit log indexed by TaskID: The original input of the task; A snapshot of all the agents involved and their states at the start of the task; The message content transmitted between all intelligent agents; All intercepted or executed operation commands and their verification results.

[0047] Simultaneously, the controller exposes a circuit breaker API. Administrators can pause, resume, or terminate running tasks at any time using the TaskID. Termination immediately destroys the corresponding sandbox and releases all resources.

[0048] In practice, a task state machine is maintained for each task (TaskID). Task states include at least: RUNNING, PAUSED, and TERMINATED. The circuit breaker API includes Pause(TaskID), Resume(TaskID), and Terminate(TaskID). Pause(TaskID) blocks subsequent operation requests for that TaskID from entering the real environment execution channel and suspends the task's scheduling / execution. Resume(TaskID) resumes the task's scheduling / execution; operation requests still need to pass permission and policy checks. Terminate(TaskID) terminates the task, destroys the corresponding sandbox execution environment, and releases resources. The circuit breaker controller writes each state change to an audit log, including at least the TaskID, operation type, time, operator identifier, and the states before and after the change.

[0049] In one specific embodiment, a dynamic risk-driven multi-stage threat response is involved. For example, in a network security protection scenario, the system deploys an intelligence analysis agent and an attack chain reconstruction agent.

[0050] Scenario: An attacker launches an attack using a new type of zero-day vulnerability. An intelligence analysis agent captures an unverified false intelligence from the dark web, claiming that a legitimate business domain name is a C2 server.

[0051] Traditional system behavior: The attack chain reveals that the agent directly adopted the intelligence and automatically issued instructions to isolate critical business servers, causing business interruption.

[0052] The system of this invention performs the above implementation examples as follows: Registration: Both the intelligence analysis agent and the attack chain reconstruction agent have been registered with AGH, with permissions of "can read intelligence sources" and "can generate isolation suggestions, but cannot execute them directly," respectively.

[0053] Task Initiation: The security operations platform submits the task "Analyze the threat level of host 192.168.1.100" to AGH.

[0054] Sandbox creation: AGH creates a sandbox TaskID=789 and activates two agents within it.

[0055] Restricted Communication: The intelligence analysis agent discovers a suspicious domain and sends a message to the attack chain reconstruction agent via AGH.

[0056] Policy Verification: Attack Chain Reconstruction: The Agent generates the command "Isolate host 192.168.1.100". This command is sent to the policy compliance verifier. The verifier finds that the command's confidence level comes from a single, low-reputation source, violating the policy of "high-risk operations require multi-source verification", and the command is blocked.

[0057] Audit and Circuit Breaker: AGH recorded the entire decision-making chain. The security analyst received an alert, saw the details of TaskID=789 in the console, and could manually freeze the intelligence source via the circuit breaker controller. The entire process had no impact on business operations.

[0058] In addition to the modules described above, the multi-agent management system 100 may also include other components; however, since these components are not relevant to the content of this disclosure, their illustrations and descriptions are omitted here.

[0059] Another embodiment provides a multi-agent management method, including the following steps: The metadata of multiple agents is managed using an agent metadata manager. The metadata includes any one or more of the following: agent ID, role definition, capability declaration, permission boundary, and interface specification. The task sandbox scheduler creates a sandbox execution environment isolated from resources for each task and activates the required business intelligence agents on demand within the sandbox. The sandbox execution environment includes independent memory space, file system, network stack and task ID. The policy compliance validator performs permission checks and policy checks on operation requests. The permission check is used to check whether the agent initiating the operation request is within the registered permission boundaries; the policy check is used to check whether the operation content matches the preset security policy library. The system utilizes end-to-end auditing and circuit breaker controllers to record process logs for each task and includes a circuit breaker API.

[0060] The method also includes using the task sandbox scheduler to receive a task request, parse the task objective, and determine the combination of agents required to complete the task based on the information in the agent metadata manager.

[0061] The method also includes using end-to-end auditing and circuit breaker controllers to record process logs for each task, including at least one of the following (1) to (4): (1) The original input of the task; (2) Each agent involved and its state snapshot at the start of the task; (3) The message content transmitted between the agents; (4) All intercepted or executed operation instructions and their verification results.

[0062] Other specific working processes of the multi-agent management method are described in the above-described embodiment of the multi-agent management system, and will not be repeated here.

[0063] Another embodiment illustrating that the system of the present invention can also be achieved by means of... Figure 3 The architecture of the computing device shown is used to implement this. Figure 3 The architecture of the computing device is shown. For example... Figure 3 As shown, the computer system 310 includes a system bus 330, one or more CPUs 340, input / output 320, and memory 350. The memory 350 can store various data or files used by the computer for processing and / or communication, as well as program instructions executed by the CPU, including multi-agent management methods. Figure 3 The architecture shown is merely exemplary and should be adjusted according to actual needs when implementing different devices. Figure 3 One or more components are included. The memory 350, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the multi-agent management method described above in this embodiment of the invention. One or more CPUs 340 execute various functional applications and data processing of the system of the present invention by running the software programs, instructions, and modules stored in the memory 350. Of course, the processor of the server provided in the embodiments of the present invention is not limited to performing the method operations described above, but can also perform related operations in the multi-agent management method provided in any embodiment of the present invention.

[0064] The memory 350 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on terminal usage. Furthermore, the memory 350 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 350 may further include memory remotely configured relative to one or more CPUs 340, these remote memories being connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0065] Input / output 320 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Input / output 320 may also include a display device such as a display screen.

[0066] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements the multi-agent management method described in the above embodiments. The computer-readable storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0067] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0068] The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0069] Furthermore, other specific operating procedures of a non-transitory computer-readable storage medium are described in the above-described embodiments of the multi-agent management system, and will not be repeated here.

[0070] In this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a step or method that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such a step or method.

[0071] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A multi-agent management system, characterized in that, The system includes an intelligent agent metadata manager, a task sandbox scheduler, a policy compliance verifier, and a full-link audit and circuit breaker controller. The agent metadata manager is used to manage the metadata of multiple agents. When each agent starts up, it sends a registration request to the agent metadata manager. The agent metadata manager assigns internal credentials to each successfully registered agent. The metadata includes any one or more of the following: agent ID, role definition, capability declaration, permission boundary, and interface specification. The task sandbox scheduler is used to parse tasks and create a sandbox execution environment isolated from resources for each task based on the information in the metadata manager. It also activates the required business intelligence agents within the sandbox as needed. The sandbox execution environment includes an independent memory space, file system, network stack, and task ID. The activated business intelligence agents are all injected with the current task ID. Direct communication between intelligence agents within the sandbox, as well as between intelligence agents and the outside world, is prohibited. The policy compliance verifier performs permission checks and policy checks on operation requests. The permission check is used to check whether the agent initiating the operation request is within the registered permission boundaries. The policy check is used to check whether the operation content matches the preset security policy library. Only operation requests that pass both checks can be executed in the real environment. If any check fails, the operation will be intercepted and an alarm log containing the complete context will be generated. The end-to-end audit and circuit breaker controller is used to record the process logs of each task and has a circuit breaker API. The circuit breaker API allows administrators to pause, resume or terminate running tasks by task ID. If the operation is terminated, the end-to-end audit and circuit breaker controller will destroy the corresponding sandbox and release all resources.

2. The multi-agent management system according to claim 1, characterized in that, The end-to-end audit and circuit breaker controller records the process logs of each task, including at least one of the following (1) to (4): (1) The original input of the task; (2) Each agent involved and its state snapshot at the start of the task; (3) The message content transmitted between the agents; (4) All intercepted or executed operation instructions and their verification results.

3. A multi-agent management method, characterized in that, Includes the following steps: The metadata of multiple agents is managed by an agent metadata manager. When each agent starts up, it sends a registration request to the agent metadata manager. The agent metadata manager assigns internal credentials to each successfully registered agent. The metadata includes any one or more of the following: agent ID, role definition, capability declaration, permission boundary, and interface specification. The task sandbox scheduler parses tasks and creates resource-isolated sandbox execution environments for each task based on information in the metadata manager. Within the sandbox, the required business intelligence agents are activated as needed. The sandbox execution environment includes independent memory space, file system, network stack, and task ID. The activated business intelligence agents are all injected with the current task ID. Direct communication between intelligence agents within the sandbox, as well as between intelligence agents and the outside world, is prohibited. The policy compliance validator performs permission checks and policy checks on operation requests. The permission check is used to check whether the agent initiating the operation request is within the registered permission boundary. The policy check is used to check whether the operation content matches the preset security policy library. Only operation requests that pass both checks can be executed in the real environment. If any check fails, the operation will be intercepted and an alarm log containing the complete context will be generated. The system utilizes end-to-end auditing and circuit breaker controllers to record the process logs of each task and includes a circuit breaker API. The circuit breaker API allows administrators to pause, resume, or terminate running tasks using the task ID. If the operation is terminated, the end-to-end auditing and circuit breaker controllers will destroy the corresponding sandbox and release all resources.

4. The multi-agent management method according to claim 3, characterized in that, The method also includes recording process logs for each task using end-to-end auditing and circuit breaker controllers, including at least one of the following (1) to (4): (1) The original input of the task; (2) Each agent involved and its state snapshot at the start of the task; (3) The message content transmitted between the agents; (4) All intercepted or executed operation instructions and their verification results.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the multi-agent management method as described in any one of claims 3 and 4.

6. A non-transitory computer-readable storage medium, wherein computer instructions are stored on the storage medium, characterized in that, When the instructions are executed by the processor, they implement the steps of the multi-agent management method as described in any one of claims 3 and 4.

Images