A multi-agent task coordination method and system

Abstract

The application discloses a multi-agent task coordination method and system, the method comprising: receiving a task instruction of an enterprise user and performing task analysis to obtain task attribute information; determining a target agent set and task execution control information from a candidate agent set according to the task attribute information and a preset coordination strategy; distributing task content based on the task execution control information and performing controlled coordination on a task execution process of the target agent to obtain an execution result; and performing summary processing on the execution result, generating a task feedback result and feeding back the task feedback result to the enterprise user. Through the method, the allocation accuracy and execution controllability of complex tasks can be improved, the context access range is clear, and the integration efficiency of multi-agent execution results is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a multi-agent task coordination method and system. Background Technology

[0002] With the development of artificial intelligence technology, intelligent assistants are increasingly being applied in scenarios such as enterprise office work, R&D management, project collaboration, operational analysis, knowledge retrieval, and workflow processing. Compared to general question-and-answer systems for individual users, intelligent assistants in enterprise environments typically need to handle multiple types of business tasks and provide capabilities such as information organization, content generation, data analysis, workflow assistance, and collaborative support based on different business scenarios. To improve the efficiency of handling complex tasks, some enterprise intelligent assistant systems have begun to introduce multiple functional units or multiple intelligent processing modules to handle different types of tasks respectively.

[0003] In related technologies, enterprise intelligent assistants typically rely on a single intelligent model or pre-set task processing flow to respond to user commands. When the task type is relatively simple, this approach can meet basic question-and-answer, document generation, or information retrieval needs. However, in real-world enterprise applications, user commands often involve multiple business processes, multiple data sources, and multiple execution objects. Different tasks may also have sequential dependencies, permission restrictions, and result integration requirements. Existing solutions are prone to problems such as inaccurate task understanding, unreasonable task allocation, lack of effective control over the execution process, and difficulty in unifying and integrating multiple processing results when handling such complex tasks. Furthermore, there are usually data boundaries between departments, positions, projects, and users within an enterprise. If related systems lack reasonable data access and task execution constraints, they are prone to problems such as information misuse, unauthorized processing, exposure of sensitive data, and interference from historical context, affecting the reliability and security of enterprise intelligent assistants in complex business scenarios.

[0004] Therefore, in the complex task processing of enterprise intelligent assistants, the lack of accuracy in task understanding and allocation, the weak ability to control the execution process, the unclear data boundary constraints, and the difficulty in integrating the results of multi-task processing have become problems that urgently need to be solved. Summary of the Invention

[0005] This application provides a multi-agent task coordination method and system, aiming to solve the problems of insufficient accuracy in task understanding and allocation, weak execution process control, unclear data boundary constraints, and difficulty in integrating multi-task processing results in the complex task processing of enterprise intelligent assistants in the existing technology.

[0006] Firstly, a multi-agent task coordination method, the method comprising: Receive task instructions from enterprise users, and parse the task instructions to obtain task attribute information; Based on the task attribute information and the preset coordination strategy, a target agent set and corresponding task execution control information are determined from the candidate agent set; wherein, the preset coordination strategy is at least used to constrain the agent's task allocation, execution permissions, and context access scope; Based on the task execution control information, the task content corresponding to the task instruction is distributed to the target agents in the target agent set, and the task execution process of the target agents is controlled and coordinated to obtain the execution result; The execution results are summarized and processed to generate task feedback results, which are then sent back to the enterprise user.

[0007] Optionally, in the above scheme, receiving the task instruction from the enterprise user and parsing the task instruction to obtain task attribute information includes: Obtain the user identity information of the enterprise user and the task text information corresponding to the task instruction; The job attribute information corresponding to the enterprise user is determined based on the user identity information; The task text information is subjected to intent recognition and keyword extraction to obtain task intent information; Based on the job attribute information and the task intent information, determine the task type information and task priority information corresponding to the task instruction; The task attribute information is generated based on the job attribute information, the task type information, and the task priority information.

[0008] Optionally, in the above scheme, determining the target agent set and corresponding task execution control information from the candidate agent set based on the task attribute information and the preset coordination strategy includes: Based on the job attribute information, load the job routing rules corresponding to the enterprise user from the preset coordination strategy to obtain the target job routing rules; Based on the task type information and the target job routing rules, a set of candidate Agent identifiers matching the task instruction is determined in a preset routing mapping table; Based on the set of candidate Agent identifiers, query the capability description information of the corresponding Agent to obtain the capability information of the candidate Agent; Based on the task priority information and the candidate agent capability information, the target agent set is determined from the candidate agent set; The task execution control information is generated based on the target Agent set and the target job routing rules.

[0009] Optionally, in the above scheme, determining the target agent set from the candidate agent set based on the task priority information and the candidate agent capability information includes: Based on the candidate agent capability information, the candidate agents in the candidate agent set are matched to obtain the agent matching result; Based on the task priority information, the candidate Agents in the Agent matching results are sorted by execution priority to obtain the Agent sorting result; Based on the Agent sorting results and the preset task complexity conditions, determine the single Agent execution mode or the multi Agent collaborative execution mode; The target agent set is determined from the candidate agent set based on the single agent execution method or the multi-agent collaborative execution method.

[0010] Optionally, in the above scheme, determining the target agent set and corresponding task execution control information from the candidate agent set based on the task attribute information and the preset coordination strategy further includes: The operation type information corresponding to the task instruction is determined based on the task attribute information; Based on the operation type information and the permission control rules in the preset coordination strategy, the permission of the task instruction is determined to obtain the execution permission level; Based on the execution permission level, determine one of the following execution processing methods: autonomous execution, execution after confirmation, or execution prohibited; The context access scope is determined based on the execution processing method and the data access permissions corresponding to the enterprise user; Based on the execution processing method and the context access scope, the task execution control information is generated.

[0011] Optionally, in the above scheme, the step of distributing the task content corresponding to the task instruction to the target agents in the target agent set based on the task execution control information, and controlling and coordinating the task execution process of the target agents to obtain the execution result includes: When the execution processing mode is autonomous execution, authorization context information is generated based on the context access scope; When the execution processing method is to execute after confirmation, a task confirmation request is generated, and after receiving the confirmation pass information, authorization context information is generated according to the context access scope; If the execution processing mode is set to prohibit execution, a prohibition feedback message is generated, and the task execution process corresponding to the task instruction is terminated. When the authorization context information is generated, the task content and the authorization context information are sent to the target agents in the target agent set to obtain the task distribution result; Based on the task distribution results, monitor the execution status of the target Agent to obtain execution status information; If the execution status information meets the preset abnormal conditions, an abnormal coordination instruction is generated, and the task execution process of the target Agent is retried, switched, or terminated according to the abnormal coordination instruction to obtain the abnormal coordination result. The execution result is obtained based on the abnormal coordination result and / or the subtask result returned by the target Agent.

[0012] Optionally, in the above scheme, the step of summarizing the execution results to generate task feedback results includes: Obtain the subtask results returned by each target agent in the target agent set to obtain the subtask result set; The subtask result set is format-standardized to obtain a standardized result set; The standardized result set is subjected to duplicate content merging and conflict content identification to obtain result organization information; Based on the results, organize the information and preset feedback templates to generate unified feedback content; The task feedback result is generated based on the unified feedback content.

[0013] Optionally, in the above scheme, the method further includes: Based on the user identity information of the enterprise user, an independent running environment is allocated to the Agent instance corresponding to the enterprise user to obtain a sandbox running environment; An independent memory storage space is established in the sandbox operating environment to obtain a personal memory space corresponding to the enterprise user; The personal memory space is divided into a short-term memory area for storing the current session context, a long-term memory area for storing historical interaction information, and a vector memory bank for storing semantic knowledge fragments. Based on the context access scope, memory content matching the task instruction is extracted from the personal memory space to obtain the isolated memory context; The isolated memory context is provided as the authorization context information to the corresponding target Agent.

[0014] Optionally, in the above scheme, the method further includes: Based on the organizational level to which the enterprise user belongs, at least two levels of memory banks are determined from the enterprise-level memory bank, department-level memory bank, team-level memory bank, and personal memory space to obtain a set of memory banks to be inherited; According to the preset memory inheritance rules, the memory content in the set of memory banks to be inherited is inherited and overwritten to obtain a shared memory context; In cases where the task instruction involves cross-user, cross-team, or cross-departmental collaboration, an Agent collaboration request is generated based on the shared memory context. The Agent's collaboration request is encapsulated according to a preset A2A communication protocol to obtain a collaboration message. The collaboration message is used to indicate task delegation, context sharing, result request, or state synchronization. The context content in the collaboration message is subjected to permission verification and desensitization processing to obtain a shareable collaboration message; Send the shareable collaboration message to the collaboration agent and receive the collaboration processing result returned by the collaboration agent; The execution result is updated based on the collaborative processing result.

[0015] Secondly, a multi-agent task coordination system, the system comprising: The task parsing module is used to receive task instructions from enterprise users and parse the task instructions to obtain task attribute information. The coordination decision module is used to determine the target agent set and the corresponding task execution control information from the candidate agent set based on the task attribute information and the preset coordination strategy. The preset coordination strategy is used to constrain the agent's task allocation, execution permissions and context access scope. The execution coordination module is used to distribute the task content corresponding to the task instruction to the target agents in the target agent set based on the task execution control information, and to coordinate the task execution process of the target agents in a controlled manner to obtain the execution result; The result aggregation module is used to aggregate the execution results, generate task feedback results, and send the task feedback results back to the enterprise user. The memory management module is used to allocate an independent operating environment and independent memory storage space to the Agent instance corresponding to the enterprise user, and to determine the authorization context information provided to the target Agent according to the context access scope; The Agent collaboration module is used to transmit collaboration messages between different Agents according to a preset A2A communication protocol when the task instruction involves cross-user, cross-team, or cross-department collaboration.

[0016] Compared with the prior art, this application has at least the following beneficial effects: Based on further analysis and research of existing technical problems, this application recognizes that existing technologies suffer from insufficient accuracy in task understanding and allocation, weak execution process control, unclear data boundary constraints, and difficulty in integrating multi-task processing results in complex task processing of enterprise intelligent assistants. This application addresses these issues by receiving task instructions from enterprise users and parsing them, converting natural language or business-related tasks into task attribute information that can be used for coordinated processing. This allows the system to first clarify the task type, object, and processing requirements. Furthermore, based on the task attribute information and preset coordination strategies, it determines the target agent set and task execution control information, enabling tasks to be assigned to agents matching the task requirements and ensuring unified constraints on agent task allocation, execution permissions, and context access scope. After task distribution, the execution process of the target agents is controlled and coordinated through the task execution control information, preventing execution deviations or unauthorized processing caused by a lack of control during task execution. Finally, the execution results are summarized and processed to generate unified task feedback results, allowing the processing results from multiple agents to be integrated into feedback content directly usable by enterprise users. Therefore, this application can improve the accuracy of enterprise intelligent assistants in understanding and allocating complex tasks, enhance the controllability of task execution process, clarify the data access boundaries when agents process tasks, and improve the problem of difficulty in unifying and integrating the results of multi-task or multi-agent execution. This solves the problems in the background technology of insufficient accuracy in task understanding and allocation, weak execution process control capability, unclear data boundary constraints, and difficulty in integrating the results of multi-task processing in enterprise intelligent assistants in handling complex tasks. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a multi-Agent task coordination method provided in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] In one embodiment, such as Figure 1 As shown, a multi-agent task coordination method is provided, including the following steps: Receive task instructions from enterprise users, and parse the task instructions to obtain task attribute information; Based on the task attribute information and the preset coordination strategy, a target agent set and corresponding task execution control information are determined from the candidate agent set; wherein, the preset coordination strategy is at least used to constrain the agent's task allocation, execution permissions, and context access scope; Based on the task execution control information, the task content corresponding to the task instruction is distributed to the target agents in the target agent set, and the task execution process of the target agents is controlled and coordinated to obtain the execution result; The execution results are summarized and processed to generate task feedback results, which are then sent back to the enterprise user.

[0020] In one optional implementation, the multi-Agent task coordination method can be applied to the coordination terminal of an enterprise intelligent assistant. This coordination terminal serves as a unified entry point between the enterprise user and multiple agents, receiving task instructions input by the enterprise user and determining the agents to be invoked, their corresponding execution permissions, and the applicable context based on the task instructions. The enterprise user can be a technical staff member, product manager, operations staff member, manager, or other personnel with an enterprise account. The task instructions can be natural language instructions or structured or semi-structured instructions containing a task description, task objective, output requirements, associated business objects, or priority markers.

[0021] After receiving task instructions from enterprise users, the coordinating end can parse the instructions to obtain task attribute information. Specifically, it can perform word segmentation, keyword extraction, intent recognition, entity recognition, and task target recognition on the task instructions to determine information such as the task category, business scenario, urgency, involved parties, and output format. For example, task categories may include technical research, architecture planning, code review, product planning, requirements analysis, competitor analysis, operational promotion, data analysis, user feedback, scheduling, meeting management, or team coordination. The task attribute information can serve as the basis for subsequently determining the target agent set and task execution control information.

[0022] After obtaining the task attribute information, the coordinator can determine the target agent set and corresponding task execution control information from the candidate agent set based on the task attribute information and preset coordination strategies. The candidate agent set can include multiple pre-registered professional agents, such as technology research agents, architecture planning agents, technical problem-solving agents, product planning agents, requirements analysis agents, competitor analysis agents, operation and promotion agents, data analysis agents, user feedback agents, schedule management agents, meeting management agents, and team coordination agents. Preset coordination strategies can include job task routing relationships, agent capability descriptions, the scope of tools that agents can call, task priority rules, execution permission rules, and context access rules. The coordinator can determine whether the task should be executed by a single agent or by multiple agents collaboratively, based on the matching relationship between the task attribute information and the preset coordination strategies.

[0023] After determining the target agent set, the coordinator can distribute the task content corresponding to the task instruction to the target agents in the target agent set based on task execution control information, and coordinate the task execution process of the target agents in a controlled manner. Task execution control information may include task splitting method, execution order, parallel execution strategy, timeout control rules, exception handling rules, execution permission requirements, accessible context scope, and result return format. The coordinator can send task parameters, authorization context, and output requirements to the target agents according to the above task execution control information, and monitor their execution status during the target agent's execution, such as received, executing, awaiting confirmation, execution exception, and execution completed. When timeouts, failures, missing results, or permission exceptions occur, the coordinator can trigger coordination control operations such as retrying, switching to a backup agent, requesting confirmation from the enterprise user, or terminating execution, thereby obtaining the corresponding execution result.

[0024] After receiving the execution results, the coordinator can summarize and process them, generate task feedback results, and then send these results back to the enterprise users. Execution results can include the results returned by a single target agent, or the results of sub-tasks returned by multiple target agents. The coordinator can standardize the format of multiple sub-task results according to unified result summarization rules, merge duplicate content, identify conflicting content, remove abnormal results, generate result summaries, and populate output templates, forming unified feedback content for enterprise users. Task feedback results can be output in the form of text, tables, task lists, document drafts, meeting schedule information, analysis reports, or system operation suggestions.

[0025] It should be noted that the preset coordination policy is used to constrain the task allocation, execution permissions, and context access scope of the Agent. Task allocation determines which Agents should handle a task; execution permissions restrict whether an Agent can execute tasks independently, whether enterprise user confirmation is required, or whether execution is prohibited; context access scope restricts the range of data that an Agent can read or use when executing tasks, thereby preventing the mixing of information between different enterprise users, different positions, different teams, or different departments.

[0026] This embodiment introduces a coordinator in the enterprise intelligent assistant to uniformly process task parsing, agent selection, execution control, and result aggregation. This enables task instructions to be assigned to appropriate target agents based on task attribute information and preset coordination strategies, and constrains the execution process through execution permissions and context access scope. Therefore, it can improve the accuracy of understanding and assigning complex tasks, enhance the controllability of the task execution process, reduce the risk of information misuse caused by unclear data boundaries, and improve the problem of difficulty in unifying and integrating the execution results of multiple agents.

[0027] In this embodiment, receiving task instructions from enterprise users and parsing the task instructions to obtain task attribute information includes: Obtain the user identity information of the enterprise user and the task text information corresponding to the task instruction; The job attribute information corresponding to the enterprise user is determined based on the user identity information; The task text information is subjected to intent recognition and keyword extraction to obtain task intent information; Based on the job attribute information and the task intent information, determine the task type information and task priority information corresponding to the task instruction; The task attribute information is generated based on the job attribute information, the task type information, and the task priority information.

[0028] In one optional implementation, when parsing task instructions, the coordinating end can first obtain the user's identity information and the task text information corresponding to the task instructions. User identity information may include enterprise account, employee identifier, job identifier, department identifier, team identifier, role identifier, permission level, or organizational hierarchy information, etc. Task text information may include a natural language task description input by the enterprise user, or it may include a task title, task content, task object, and task attachment summary transmitted from the front-end system, business system, or enterprise office platform.

[0029] The coordinating end can determine the job attribute information corresponding to enterprise users based on user identity information. Job attribute information can be used to indicate the job type to which an enterprise user belongs, such as a technical position, product position, operations position, management position, or other business position. Specifically, the coordinating end can determine the job attribute information of enterprise users through the enterprise organizational structure table, user permission table, job configuration table, or enterprise identity authentication system. For example, when user identity information corresponds to the technical department and the job identifier is R&D engineer, the job attribute information can be determined to be a technical position; when user identity information corresponds to the product department and the job identifier is product manager, the job attribute information can be determined to be a product position.

[0030] After obtaining the task text information, the coordinating end can perform intent recognition and keyword extraction to obtain task intent information. Specifically, methods such as rule matching, classification models, semantic vector matching, keyword dictionaries, or large language model inference can be used to identify the task objectives, action words, business objects, and output requirements in the task text information. For example, for the task text "Help me investigate whether a certain technology is suitable for the current system architecture," keywords such as "investigate," "technology," and "system architecture" can be extracted, and the task intent of a technology investigation or architecture evaluation type can be identified; for the task text "Organize competitor functions and output requirement suggestions," the task intent of a competitor analysis or requirement analysis type can be identified.

[0031] Subsequently, the coordinator can determine the task type and priority information corresponding to the task instruction based on job attribute information and task intent information. Task type information indicates the business category to which the task belongs, while task priority information indicates the urgency or importance of the task. Task priority information can include levels such as P0, P1, and P2, or can be converted to high, medium, and low priorities according to preset rules within the enterprise. Specifically, the coordinator can determine task priority information based on urgent words, deadlines, task origins, job attributes, and business impact scope in the task text. For example, tasks involving online failures, customer deliveries, or management instructions can be assigned a higher priority.

[0032] After determining the job attribute information, task type information, and task priority information, the coordinator can generate task attribute information based on this information. This task attribute information can then serve as input data for subsequent tasks such as loading job routing rules, matching candidate agents, determining execution permissions, and defining the context access scope.

[0033] This embodiment can incorporate enterprise user identity, job attributes, and task text semantics into the task parsing process, avoiding the problem of one-sided task understanding caused by relying solely on task text keywords. Since subsequent agent selection and execution control are based on task attribute information, it can improve the accuracy of task type identification, priority judgment, and task allocation in enterprise multi-job scenarios.

[0034] In this embodiment, determining the target agent set and corresponding task execution control information from the candidate agent set based on the task attribute information and the preset coordination strategy includes: Based on the job attribute information, load the job routing rules corresponding to the enterprise user from the preset coordination strategy to obtain the target job routing rules; Based on the task type information and the target job routing rules, a set of candidate Agent identifiers matching the task instruction is determined in a preset routing mapping table; Based on the set of candidate Agent identifiers, query the capability description information of the corresponding Agent to obtain the capability information of the candidate Agent; Based on the task priority information and the candidate agent capability information, the target agent set is determined from the candidate agent set; The task execution control information is generated based on the target Agent set and the target job routing rules.

[0035] In one optional implementation, the coordinator can load the corresponding job routing rules from a preset coordination strategy based on job attribute information to obtain the target job routing rules. Different jobs have different task processing priorities, so job routing rules can be pre-configured for technical, product, operations, and management positions. For example, the job routing rules for technical positions can prioritize matching technical research agents, architecture planning agents, and technical problem-solving agents; the job routing rules for product positions can prioritize matching product planning agents, requirements analysis agents, and competitor analysis agents; the job routing rules for operations positions can prioritize matching operations promotion agents, data analysis agents, and user feedback agents; and the job routing rules for management positions can prioritize matching schedule management agents, meeting management agents, and team coordination agents.

[0036] After obtaining the routing rules for the target position, the coordinator can determine a set of candidate Agent identifiers matching the task instruction from a pre-defined routing mapping table based on the task type information and the target position routing rules. The pre-defined routing mapping table may include fields such as task type, position, Agent identifier, Agent capability tag, execution method, permission requirements, priority rule, and backup Agent. Through the pre-defined routing mapping table, different task types under different positions can be mapped to corresponding professional Agents. For example, for the same "data analysis" task, an operations position can be preferentially routed to an operations data analysis Agent, while a management position can be preferentially routed to a business analysis or team coordination Agent.

[0037] After determining the set of candidate Agent identifiers, the coordinator can query the capability description information of the corresponding Agent based on the candidate Agent identifier set to obtain the candidate Agent capability information. Agent capability description information may include the types of tasks the Agent is good at, the data sources it can invoke, the tools it can invoke, the supported output formats, historical execution performance, the priority range it can handle, and its current load status. By querying the candidate Agent capability information, the coordinator can further determine whether the candidate Agent is suitable for handling the current task.

[0038] Subsequently, the coordinator can determine the target agent set from the candidate agent set based on task priority information and candidate agent capability information. For example, for high-priority tasks, agents with high execution stability, fast response speed, and low current load can be selected first; for complex tasks, multiple agents with complementary capabilities can be selected to execute collaboratively. The coordinator further generates task execution control information based on the target agent set and target job routing rules. The task execution control information may include the execution order of target agents, task parameter allocation method, whether parallel execution is required, whether user confirmation is required, the scope of accessible context, and result return requirements, etc.

[0039] This embodiment determines the target agent set by combining job routing rules, preset routing mapping tables, and agent capability description information. This makes agent selection no longer dependent on fixed hard-coded rules, but can adapt to the enterprise task processing needs of different jobs, different task types, and different priorities. Therefore, it can improve the task routing accuracy and agent invocation flexibility of enterprise intelligent assistants in multi-job scenarios.

[0040] In this embodiment, determining the target agent set from the candidate agent set based on the task priority information and the candidate agent capability information includes: Based on the candidate agent capability information, the candidate agents in the candidate agent set are matched to obtain the agent matching result; Based on the task priority information, the candidate Agents in the Agent matching results are sorted by execution priority to obtain the Agent sorting result; Based on the Agent sorting results and the preset task complexity conditions, determine the single Agent execution mode or the multi Agent collaborative execution mode; The target agent set is determined from the candidate agent set based on the single agent execution method or the multi-agent collaborative execution method.

[0041] In one optional implementation, when determining the target agent set based on candidate agent capability information, the coordinator can first perform capability matching on the candidate agents in the candidate agent set based on the candidate agent capability information to obtain the agent matching result. Capability matching can be performed based on task type, task objective, required tools, required data source, output format, agent capability tags, and historical execution results. For example, when the task requires outputting a technical research report, the coordinator can use agents with technical retrieval, technical summary, and technical feasibility analysis capabilities as highly matched agents; when the task requires generating meeting schedules, agents with schedule reading, meeting creation, and participant coordination capabilities can be used as highly matched agents.

[0042] After obtaining the Agent matching results, the coordinator can prioritize the candidate Agents based on task priority information, resulting in an Agent ranking. This ranking can be based on factors such as task urgency, candidate Agent matching accuracy, current agent load, historical success rate, response latency, and enterprise preferences. For P0 level tasks, agents with high availability, fast processing speed, and sufficient permissions are prioritized; for general tasks, agents with lower cost or lower load are prioritized.

[0043] Subsequently, the coordinator can determine whether to use a single-agent execution method or a multi-agent collaborative execution method based on the Agent sorting results and preset task complexity conditions. Preset task complexity conditions can be determined based on factors such as the number of task steps, the number of business domains involved, the number of data sources involved, whether it crosses departments, whether there are multiple output targets, and the complexity of the task text. For example, if the task involves only a single domain and the output target is clear, it can be determined to use a single-agent execution method; if the task involves multiple domains such as technology, product, and operations, or requires multiple sub-results to combine to form the final result, it can be determined to use a multi-agent collaborative execution method.

[0044] After determining whether to use a single-agent execution method or a multi-agent collaborative execution method, the coordinator can select the target agent set from the candidate agent set. For the single-agent execution method, the agent with the highest ranking and that meets the permission requirements can be selected as the target agent; for the multi-agent collaborative execution method, multiple agents can be selected according to the task sub-domain or sub-step, and the task boundaries and result return requirements of each target agent can be determined.

[0045] This embodiment can dynamically determine whether to execute a single agent or multiple agents collaboratively based on the matching of candidate agent capabilities, combined with task priority and task complexity. This avoids resource waste caused by simply calling multiple agents in parallel, and also avoids insufficient professionalism caused by handling complex tasks by only a single agent. Therefore, it can improve the rationality of agent resource allocation and the quality of complex task processing.

[0046] In this embodiment, determining the target agent set and corresponding task execution control information from the candidate agent set based on the task attribute information and the preset coordination strategy further includes: The operation type information corresponding to the task instruction is determined based on the task attribute information; Based on the operation type information and the permission control rules in the preset coordination strategy, the permission of the task instruction is determined to obtain the execution permission level; Based on the execution permission level, determine one of the following execution processing methods: autonomous execution, execution after confirmation, or execution prohibited; The context access scope is determined based on the execution processing method and the data access permissions corresponding to the enterprise user; Based on the execution processing method and the context access scope, the task execution control information is generated.

[0047] In one optional implementation, the coordinating end can also determine the operation type information corresponding to the task instruction based on the task attribute information. The operation type information can represent the nature of the operation involved in the task instruction, such as reading information, generating a draft, organizing data, creating a document, creating a meeting, assigning tasks, sending information externally, accessing sensitive data, or making decisions on behalf of the user. Different operation types correspond to different levels of risk, therefore further permission verification is required.

[0048] The coordinating end can determine the execution permission level of task instructions based on operation type information and the permission control rules in the preset coordination strategy. Permission control rules can be configured according to enterprise security requirements, job permissions, data sensitivity levels, and operational risk levels. Execution permission levels can include self-execution, execution after confirmation, and execution prohibited. Self-execution is suitable for low-risk operations such as reading, organizing, and generating drafts; execution after confirmation is suitable for medium-risk operations such as creating meetings, sending external messages, assigning tasks, and operations involving other personnel; execution prohibited is suitable for high-risk operations such as making critical decisions on behalf of users, accessing confidential data, bypassing approval processes, or performing other sensitive operations.

[0049] After obtaining the execution permission level, the coordinator can determine one of the following execution methods based on the permission level: autonomous execution, execution after confirmation, or execution prohibition. For autonomous execution, the coordinator can directly distribute the task to the target Agent; for execution after confirmation, the coordinator can first send a confirmation request to the enterprise user or approver, and continue execution after confirmation; for execution prohibition, the coordinator can stop the execution process and return the reason for prohibition.

[0050] The coordinator can also determine the contextual access scope based on the execution processing method and the corresponding data access permissions of enterprise users. Data access permissions can be determined based on the RBAC model or the enterprise's existing permission system, and permission dimensions can include personal, team, department, enterprise, and confidential levels. The contextual access scope can be used to limit the session context, history, enterprise knowledge, departmental information, team project information, or personal memory content that the target agent can read. For sensitive data, control can also be achieved through methods such as encrypted storage, access auditing, and anonymized output.

[0051] Finally, the coordinator can generate task execution control information based on the execution processing method and context access scope, so that subsequent task distribution and execution coordination are constrained by permission boundaries and data boundaries.

[0052] This embodiment further introduces operation type identification, permission judgment, and context access scope determination after determining the target agent, so that the execution behavior of the target agent is constrained by the enterprise permission rules and data access permissions; therefore, it can reduce the risks of unauthorized access, accidental execution of sensitive operations, and mixed use of enterprise data, and improve the security and controllability of enterprise intelligent assistant task execution.

[0053] In this embodiment, the step of distributing the task content corresponding to the task instruction to the target agents in the target agent set based on the task execution control information, and controlling and coordinating the task execution process of the target agents to obtain the execution result includes: When the execution processing mode is autonomous execution, authorization context information is generated based on the context access scope; When the execution processing method is to execute after confirmation, a task confirmation request is generated, and after receiving the confirmation pass information, authorization context information is generated according to the context access scope; If the execution processing mode is set to prohibit execution, a prohibition feedback message is generated, and the task execution process corresponding to the task instruction is terminated. When the authorization context information is generated, the task content and the authorization context information are sent to the target agents in the target agent set to obtain the task distribution result; Based on the task distribution results, monitor the execution status of the target Agent to obtain execution status information; If the execution status information meets the preset abnormal conditions, an abnormal coordination instruction is generated, and the task execution process of the target Agent is retried, switched, or terminated according to the abnormal coordination instruction to obtain the abnormal coordination result. The execution result is obtained based on the abnormal coordination result and / or the subtask result returned by the target Agent.

[0054] In one optional implementation, when the execution mode is autonomous, the coordinator can generate authorization context information based on the access scope. The authorization context information may include context content required and permitted for the target agent to execute the task, such as current session summary, task-related document summary, historical task results, fragments of company policies, fragments of project materials, or user preference information. The coordinator may only provide the target agent with content within the access scope of the context; unauthorized content will not be sent to the target agent.

[0055] When the execution method is "execute after confirmation," the coordinator can generate a task confirmation request. Upon receiving confirmation, it generates authorization context information based on the access scope. The task confirmation request may include the task name, the proposed calling agent, the proposed operation, the personnel involved, the scope of data involved, potential external impacts, and confirmation options. Confirmation can be returned by the enterprise user, an authorized approver with relevant permissions, or the enterprise approval system. Only after confirmation does the coordinator continue task distribution.

[0056] When the execution mode is set to "prohibit execution," the coordinator can generate a prohibition feedback message and terminate the task execution process corresponding to the task instruction. The prohibition feedback message may include the reason for the prohibition, the relevant permission rules, suggested alternative operations, or the type of permission required. This method prevents the Agent from continuing execution when it lacks the necessary permissions or when the task risk is too high.

[0057] With authorization context information generated, the coordinator can send the task content and authorization context information to the target agents in the target agent set, obtaining the task distribution result. The task distribution result can indicate whether the task was successfully sent, whether the target agent has received the task, and whether task execution has started. The task content can include the original task instructions, parsed task parameters, task output requirements, execution time limit, and subtasks that the target agent needs to process.

[0058] The coordinator can monitor the execution status of the target agent based on the task distribution results to obtain execution status information. This information can include whether the target agent is online, whether it has received the task, its current execution progress, whether it is waiting for external input, whether any exceptions have occurred, whether execution has been completed, and whether the returned result is complete. When the execution status information meets preset exception conditions, such as the target agent not responding after a timeout, returning an error, failing permission verification, having empty output, conflicting execution results, or failing execution, the coordinator can generate an exception coordination instruction. Based on this instruction, the coordinator can retry, switch, or terminate the task execution process of the target agent, resulting in an exception coordination outcome. Retry handling can be used for temporary failure scenarios; switching handling can call a backup agent; and termination handling can be used for scenarios involving permission exceptions, sensitive risks, or multiple failures.

[0059] Ultimately, the coordinator can obtain the execution result based on the abnormal coordination result and / or the subtask result returned by the target agent. For example, if the task is completed normally, the execution result can be formed by the subtask result returned by the target agent; if the task encounters an error and is completed after retrying or switching, the execution result can include the subtask result after retrying or switching; if the task is terminated, the execution result can include the termination status and reason explanation.

[0060] This embodiment sets up corresponding processing paths for three types of execution methods: autonomous execution, execution after confirmation, and execution prohibited. After task distribution, it continuously monitors the execution status of the target Agent and performs retry, switching, or termination processing in abnormal situations. Therefore, it can improve the closed-loop and reliability of the enterprise intelligent assistant task execution process and avoid task loss of control due to incomplete permission branches or execution abnormalities and lack of coordination.

[0061] In this embodiment, the step of summarizing the execution results to generate task feedback results includes: Obtain the subtask results returned by each target agent in the target agent set to obtain the subtask result set; The subtask result set is format-standardized to obtain a standardized result set; The standardized result set is subjected to duplicate content merging and conflict content identification to obtain result organization information; Based on the results, organize the information and preset feedback templates to generate unified feedback content; The task feedback result is generated based on the unified feedback content.

[0062] In one optional implementation, the coordinator can obtain the subtask results returned by each target agent in the target agent set, thus obtaining a subtask result set. Subtask results may include textual conclusions, analytical data, candidate solutions, task execution status, tool call results, document drafts, meeting arrangement results, or information returned by the business system, etc. For multi-agent collaborative task execution, different target agents can return subtask results of different types or from different perspectives.

[0063] After obtaining the subtask result set, the coordinator can perform format standardization on the subtask result set to obtain a standardized result set. Format standardization can include standardizing field names, time formats, data units, paragraph structures, result labels, confidence level expressions, and citation or source identifiers. For example, the technical conclusions returned by the technology research agent, the product suggestions returned by the product planning agent, and the promotion plans returned by the operations and promotion agent can be converted into standardized results with the same result field structure.

[0064] Subsequently, the coordinator can merge duplicate content and identify conflicting content in the standardized result set to obtain result summary information. Duplicate content merging can be used to eliminate repeated facts, suggestions, or conclusions in the results returned by multiple agents; conflicting content identification can be used to identify inconsistencies between different agents in terms of task conclusions, timelines, priority suggestions, data definitions, or risk assessments. For conflicting content, the coordinator can mark the source of the conflict, retain multiple viewpoints, determine the priority of results based on agent weights or result confidence levels, or request further confirmation from enterprise users.

[0065] After receiving and compiling the results, the coordinating team can generate standardized feedback content based on the compiled results and pre-set feedback templates. These pre-set feedback templates can be configured according to task type or company role, such as technical research report templates, requirements analysis templates, competitor analysis templates, operational promotion plan templates, meeting minutes templates, and task allocation list templates. The standardized feedback content can include a task summary, processing steps, core conclusions, items requiring confirmation, risk warnings, follow-up suggestions, and attachments or references.

[0066] Finally, the coordinator can generate task feedback results based on the unified feedback content and send them to enterprise users. These results can be displayed in the enterprise intelligent assistant interface or exported to the enterprise office system, document system, task management system, or message notification system.

[0067] This embodiment can standardize, deduplicate, identify conflicts, and summarize the sub-task results returned by multiple target agents, enabling enterprise users to receive unified, clear, and executable task feedback results. Therefore, it can solve the problems of scattered execution results, inconsistent formats, and difficulty in integrating conclusions from multiple agents, and improve the efficiency of enterprise users in using the results of multi-agent task processing.

[0068] In this embodiment, the method further includes: Based on the user identity information of the enterprise user, an independent running environment is allocated to the Agent instance corresponding to the enterprise user to obtain a sandbox running environment; An independent memory storage space is established in the sandbox operating environment to obtain a personal memory space corresponding to the enterprise user; The personal memory space is divided into a short-term memory area for storing the current session context, a long-term memory area for storing historical interaction information, and a vector memory bank for storing semantic knowledge fragments. Based on the context access scope, memory content matching the task instruction is extracted from the personal memory space to obtain the isolated memory context; The isolated memory context is provided as the authorization context information to the corresponding target Agent.

[0069] In one optional implementation, the coordinator can allocate an independent runtime environment to the Agent instance corresponding to an enterprise user based on the user's identity information, thus obtaining a sandbox runtime environment. The sandbox runtime environment can be used to isolate the runtime resources and data boundaries of Agent instances corresponding to different enterprise users. Specifically, the sandbox runtime environment may include independent runtime containers, independent process spaces, independent configuration files, independent access tokens, independent log spaces, and independent memory access interfaces, thereby preventing memory sharing, configuration sharing, or context sharing between Agent instances of different enterprise users.

[0070] In a sandbox environment, an independent memory storage space can be established, creating a personal memory space corresponding to each enterprise user. This personal memory space stores the user's personal context, preferences, historical tasks, and semantic knowledge fragments generated while using the enterprise intelligent assistant. The personal memory space can be bound to the enterprise user's identity information, and access control rules can restrict direct access by other users or other Agent instances.

[0071] Personal memory space can be divided into short-term memory, long-term memory, and a vector memory. Short-term memory stores the current session context, such as a summary of the current task, recent interactions, and the current task state. Long-term memory stores historical interaction information, such as user preferences, historical decisions, common output formats, work habits, historical task results, and common business rules. The vector memory stores semantic knowledge fragments and supports similarity retrieval, enabling rapid retrieval of relevant historical content when processing similar tasks.

[0072] In practical implementation, the personal memory space can also support memory snapshots and recovery. Memory snapshots can be generated before task execution, memory updates, or critical operations, for subsequent auditing, rollback, or anomaly recovery. Sensitive memory data can be stored using encryption methods, such as AES-256 encryption. For memory read and write operations, access audit logs can be recorded, and the audit content can include the access subject, access time, access type, access object, access result, and task identifier, etc.

[0073] When a target agent needs to use memory content to perform a task, the coordinator can extract memory content matching the task instructions from the individual memory space based on the context access scope, thus obtaining an isolated memory context. Specifically, search criteria can be generated based on the task text, task type information, user job information, and task objectives, and authorized access content can be retrieved from the short-term memory area, long-term memory area, and vector memory. Memory content that does not fall within the context access scope is not extracted, nor is it provided to the target agent as an authorized context.

[0074] Finally, the coordinator can provide the isolated memory context as authorization context information to the corresponding target agent, so that the target agent can only use personal memory content within the authorized scope when performing tasks.

[0075] This embodiment isolates the Agent instances corresponding to enterprise users through a sandbox operating environment and independent memory storage space, and provides isolated memory context to the target Agent within the authorized scope; therefore, it can reduce the risks of memory pollution, unauthorized access and privacy leakage between Agent instances of different users, while retaining the ability to improve task processing quality based on personal historical context.

[0076] In this embodiment, the method further includes: Based on the organizational level to which the enterprise user belongs, at least two levels of memory banks are determined from the enterprise-level memory bank, department-level memory bank, team-level memory bank, and personal memory space to obtain a set of memory banks to be inherited; According to the preset memory inheritance rules, the memory content in the set of memory banks to be inherited is inherited and overwritten to obtain a shared memory context; In cases where the task instruction involves cross-user, cross-team, or cross-departmental collaboration, an Agent collaboration request is generated based on the shared memory context. The Agent's collaboration request is encapsulated according to a preset A2A communication protocol to obtain a collaboration message. The collaboration message is used to indicate task delegation, context sharing, result request, or state synchronization. The context content in the collaboration message is subjected to permission verification and desensitization processing to obtain a shareable collaboration message; Send the shareable collaboration message to the collaboration agent and receive the collaboration processing result returned by the collaboration agent; The execution result is updated based on the collaborative processing result.

[0077] In one optional implementation, the coordinating end can determine at least two levels of memory banks—enterprise-level memory bank, department-level memory bank, team-level memory bank, and personal memory space—based on the organizational hierarchy of the enterprise user, thus obtaining a set of memory banks to be inherited. The enterprise-level memory bank can be used to store corporate culture, rules and regulations, general business processes, standardized templates, and common knowledge; the department-level memory bank can be used to store departmental standards, departmental documents, job processes, and professional knowledge; the team-level memory bank can be used to store project experience, collaboration standards, task reviews, and team materials; and the personal memory space can be used to store personal preferences, work habits, historical tasks, and personal experience.

[0078] After determining the set of memories to be inherited, the coordinator can inherit and overwrite the memory content in the set of memories according to preset memory inheritance rules to obtain a shared memory context. Preset memory inheritance rules can include top-down inheritance rules and lower-level overriding-upper-level rules. Specifically, the Agent instance corresponding to an enterprise user can inherit relevant content from the enterprise-level memory, department-level memory, team-level memory, and personal memory space. When lower-level memory content conflicts with higher-level memory content, a general rule of lower-level memory overriding higher-level memory can be used according to preset priorities. For example, personal preferences can override departmental general output formats, and team project specifications can override some content in enterprise general templates.

[0079] In practical implementation, a memory contribution mechanism and a memory version management mechanism can also be set up. Enterprise users can submit personal experiences, task review results, or effective processing procedures to team-level, department-level, or enterprise-level memory repositories. After review by team leaders, department leaders, or managers, these repositories are shared with users within the corresponding scope. The memory version management mechanism can record the creation, modification, review, publication, and rollback information of memory content to support enterprise knowledge evolution auditing.

[0080] Regarding access permissions, different read / write permissions for different memory banks can be configured according to organizational levels. For example, an enterprise-level memory bank can be configured to be readable by all employees and writable by administrators; a department-level memory bank can be configured to be readable within the department and writable by the department head; a team-level memory bank can be configured to be readable within the team and writable by the team leader; and personal memory spaces can be configured to be readable and writable only by the user, while allowing enterprise users to authorize others to access them. For confidential memories, stricter access approval and encrypted storage policies can be set.

[0081] When task instructions involve cross-user, cross-team, or cross-departmental collaboration, the coordinator can generate Agent collaboration requests based on the shared memory context. Agent collaboration requests can include collaboration goals, task content, a summary of the context to be shared, collaborating parties, permission requirements, expected results, and validity period. Cross-user collaboration can be used for agents from different employees to share project context; cross-team collaboration can be used for agents from different teams to coordinate task dependencies; and cross-departmental collaboration can be used for agents from different departments such as technology, product, operations, or management to jointly handle complex tasks.

[0082] The coordinator can encapsulate Agent collaboration requests according to a preset A2A communication protocol to obtain collaboration messages. These messages can be used to indicate task delegation, context sharing, result requests, or state synchronization. Specifically, collaboration messages may include sender Agent identifier, receiver Agent identifier, message type, task identifier, context summary, permission level, de-identification flag, version number, validity period, callback address, and audit flag. The preset A2A communication protocol supports both synchronous request-response and asynchronous event notification modes to adapt to different collaboration scenarios.

[0083] Before sending a collaborative message, the coordinator can perform permission verification and anonymization on the context content of the message to obtain a shareable collaborative message. Permission verification can be used to determine whether the sender's agent, receiver's agent, enterprise user, team, or department has the permission to share the corresponding context; anonymization can be used to automatically filter salary, performance, personal privacy, confidential document content, or other sensitive information. Collaborative messages can also have a context version number and expiration date set, supporting incremental updates and automatic cleanup.

[0084] Regarding collaboration access control, enterprise users can configure a whitelist of other employees or other agents that can collaborate with their Agent; cross-departmental collaboration can be configured to require approval from the heads of both departments; all A2A communications can be logged with audit logs, which can include sender, receiver, message type, shared content summary, permission verification result, approval record, and processing time; enterprise users can also terminate the collaboration relationship with a specific Agent, and after termination, the corresponding collaboration context will no longer be shared.

[0085] Finally, the coordinator can send shareable collaboration messages to the collaboration agent and receive the collaboration processing results returned by the collaboration agent. The collaboration processing results may include the execution result of the delegated task, a requested context summary, status synchronization information, or result feedback information. The coordinator can update the execution results based on the collaboration processing results, enabling information generated by cross-user, cross-team, or cross-department agent collaboration to be incorporated into the final task processing.

[0086] This embodiment further introduces a hierarchical memory inheritance system at the enterprise, department, team, and individual levels, building upon memory isolation. It achieves enterprise knowledge accumulation and reuse through memory contribution, review, version management, and access control. Simultaneously, it utilizes a pre-defined A2A communication protocol to enable task delegation, context sharing, result requests, and status synchronization between different agents. Furthermore, it enhances the security of cross-organizational collaboration by incorporating access verification, data masking, validity periods, approval, auditing, and collaboration termination mechanisms. Therefore, it can simultaneously address the needs of enterprise knowledge sharing, personal data isolation, and cross-agent collaborative processing of complex tasks.

[0087] In one embodiment, a multi-agent task coordination system is provided, comprising: The task parsing module is used to receive task instructions from enterprise users and parse the task instructions to obtain task attribute information. The coordination decision module is used to determine the target agent set and the corresponding task execution control information from the candidate agent set based on the task attribute information and the preset coordination strategy. The preset coordination strategy is used to constrain the agent's task allocation, execution permissions and context access scope. The execution coordination module is used to distribute the task content corresponding to the task instruction to the target agents in the target agent set based on the task execution control information, and to coordinate the task execution process of the target agents in a controlled manner to obtain the execution result; The result aggregation module is used to aggregate the execution results, generate task feedback results, and send the task feedback results back to the enterprise user. The memory management module is used to allocate an independent operating environment and independent memory storage space to the Agent instance corresponding to the enterprise user, and to determine the authorization context information provided to the target Agent according to the context access scope; The Agent collaboration module is used to transmit collaboration messages between different Agents according to a preset A2A communication protocol when the task instruction involves cross-user, cross-team, or cross-department collaboration.

[0088] The specific implementation details of each module can be found in the above description of the multi-agent task coordination method, and will not be repeated here.

[0089] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A multi-agent task coordination method, characterized in that, The method includes: Receive task instructions from enterprise users, and parse the task instructions to obtain task attribute information; Based on the task attribute information and a preset coordination strategy, a target agent set and corresponding task execution control information are determined from the candidate agent set. The preset coordination strategy at least constrains the agent's task allocation, execution permissions, and context access scope. Determining the target agent set and corresponding task execution control information based on the task attribute information and the preset coordination strategy includes: determining the operation type information corresponding to the task instruction based on the task attribute information; performing permission judgment on the task instruction based on the operation type information and the permission control rules in the preset coordination strategy to obtain the execution permission level; determining one of the following execution processing methods based on the execution permission level: autonomous execution, execution after confirmation, or prohibited execution; determining the context access scope based on the execution processing method and the data access permissions corresponding to the enterprise user; and generating the task execution control information based on the execution processing method and the context access scope. When the execution processing mode is autonomous execution, authorization context information is generated based on the context access scope; When the execution processing method is to execute after confirmation, a task confirmation request is generated, and after receiving the confirmation pass information, authorization context information is generated according to the context access scope; If the execution processing mode is set to prohibit execution, a prohibition feedback message is generated, and the task execution process corresponding to the task instruction is terminated. The step of generating authorization context information based on the context access scope includes: Based on the user identity information of the enterprise user, an independent running environment is allocated to the Agent instance corresponding to the enterprise user to obtain a sandbox running environment; An independent memory storage space is established in the sandbox operating environment to obtain a personal memory space corresponding to the enterprise user; The personal memory space is divided into a short-term memory area for storing the current session context, a long-term memory area for storing historical interaction information, and a vector memory bank for storing semantic knowledge fragments. Based on the context access scope, memory content matching the task instruction is extracted from the personal memory space to obtain the isolated memory context; The isolated memory context is used as the authorization context information; When the authorization context information is generated, based on the task execution control information, the task content corresponding to the task instruction and the authorization context information are distributed to the target agents in the target agent set to obtain the task distribution result; according to the task distribution result, the task execution process of the target agent is controlled and coordinated to obtain the execution result. The execution results are summarized and processed to generate task feedback results, which are then sent back to the enterprise user. The process of receiving task instructions from enterprise users and parsing these instructions to obtain task attribute information includes: Obtain the user identity information of the enterprise user and the task text information corresponding to the task instruction; The job attribute information corresponding to the enterprise user is determined based on the user identity information; The task text information is subjected to intent recognition and keyword extraction to obtain task intent information; Based on the job attribute information and the task intent information, determine the task type information and task priority information corresponding to the task instruction; The task attribute information is generated based on the job attribute information, the task type information, and the task priority information.

2. The multi-agent task coordination method according to claim 1, characterized in that, The step of determining the target agent set and corresponding task execution control information from the candidate agent set based on the task attribute information and preset coordination strategy includes: Based on the job attribute information, load the job routing rules corresponding to the enterprise user from the preset coordination strategy to obtain the target job routing rules; Based on the task type information and the target job routing rules, a set of candidate Agent identifiers matching the task instruction is determined in a preset routing mapping table; Based on the set of candidate Agent identifiers, query the capability description information of the corresponding Agent to obtain the capability information of the candidate Agent; Based on the task priority information and the candidate agent capability information, the target agent set is determined from the candidate agent set; The task execution control information is generated based on the target Agent set and the target job routing rules.

3. The multi-agent task coordination method according to claim 2, characterized in that, The step of determining the target agent set from the candidate agent set based on the task priority information and the candidate agent capability information includes: Based on the candidate agent capability information, the candidate agents in the candidate agent set are matched to obtain the agent matching result; Based on the task priority information, the candidate Agents in the Agent matching results are sorted by execution priority to obtain the Agent sorting result; Based on the Agent sorting results and the preset task complexity conditions, determine the single Agent execution mode or the multi Agent collaborative execution mode; The target agent set is determined from the candidate agent set based on the single agent execution method or the multi-agent collaborative execution method.

4. The multi-agent task coordination method according to claim 1, characterized in that, The controlled coordination of the task execution process of the target Agent based on the task distribution result to obtain the execution result includes: Based on the task distribution results, monitor the execution status of the target Agent to obtain execution status information; If the execution status information meets the preset abnormal conditions, an abnormal coordination instruction is generated, and the task execution process of the target Agent is retried, switched, or terminated according to the abnormal coordination instruction to obtain the abnormal coordination result. The execution result is obtained based on the abnormal coordination result and / or the subtask result returned by the target Agent.

5. The multi-agent task coordination method according to claim 4, characterized in that, The process of summarizing the execution results to generate task feedback results includes: Obtain the subtask results returned by each target agent in the target agent set to obtain the subtask result set; The subtask result set is format-standardized to obtain a standardized result set; The standardized result set is subjected to duplicate content merging and conflict content identification to obtain result organization information; Based on the results, organize the information and preset feedback templates to generate unified feedback content; The task feedback result is generated based on the unified feedback content.

6. The multi-agent task coordination method according to claim 1, characterized in that, The method further includes: Based on the organizational level to which the enterprise user belongs, at least two levels of memory banks are determined from the enterprise-level memory bank, department-level memory bank, team-level memory bank, and personal memory space to obtain a set of memory banks to be inherited; According to the preset memory inheritance rules, the memory content in the set of memory banks to be inherited is inherited and overwritten to obtain a shared memory context; In cases where the task instruction involves cross-user, cross-team, or cross-departmental collaboration, an Agent collaboration request is generated based on the shared memory context. The Agent's collaboration request is encapsulated according to a preset A2A communication protocol to obtain a collaboration message. The collaboration message is used to indicate task delegation, context sharing, result request, or state synchronization. The context content in the collaboration message is subjected to permission verification and desensitization processing to obtain a shareable collaboration message; Send the shareable collaboration message to the collaboration agent and receive the collaboration processing result returned by the collaboration agent; The execution result is updated based on the collaborative processing result.

7. A multi-agent task coordination system, characterized in that, include: The task parsing module is used to receive task instructions from enterprise users and parse the task instructions to obtain task attribute information. The coordination and decision-making module is used to determine a target agent set and corresponding task execution control information from a candidate agent set based on the task attribute information and a preset coordination strategy. The preset coordination strategy at least constrains the agent's task allocation, execution permissions, and context access scope. Determining the target agent set and corresponding task execution control information from the candidate agent set based on the task attribute information and the preset coordination strategy includes: determining the operation type information corresponding to the task instruction based on the task attribute information; performing permission judgment on the task instruction based on the operation type information and the permission control rules in the preset coordination strategy to obtain the execution permission level; determining one of the following execution processing methods based on the execution permission level: autonomous execution, execution after confirmation, or prohibited execution; determining the context access scope based on the execution processing method and the data access permissions corresponding to the enterprise user; generating the task execution control information based on the execution processing method and the context access scope; and, if the execution processing method is autonomous execution, determining the context access scope based on the context access scope. The process involves: generating authorization context information based on the access scope; generating a task confirmation request when the execution processing method is confirmation-based, and generating authorization context information based on the access scope after receiving confirmation information; generating prohibition feedback information and terminating the task execution process corresponding to the task instruction when the execution processing method is prohibition-based; wherein, generating authorization context information based on the access scope includes: allocating an independent running environment to the Agent instance corresponding to the enterprise user based on the user identity information of the enterprise user, obtaining a sandbox running environment; establishing an independent memory storage space in the sandbox running environment, obtaining a personal memory space corresponding to the enterprise user; dividing the personal memory space into a short-term memory area for storing the current session context, a long-term memory area for storing historical interaction information, and a vector memory for storing semantic knowledge fragments; extracting memory content matching the task instruction from the personal memory space based on the access scope, obtaining an isolated memory context; and using the isolated memory context as the authorization context information. The execution coordination module is used to, upon generating the authorization context information, distribute the task content corresponding to the task instruction and the authorization context information to the target agents in the target agent set based on the task execution control information, thereby obtaining a task distribution result; and to perform controlled coordination of the task execution process of the target agents according to the task distribution result, thereby obtaining an execution result. The result aggregation module is used to aggregate the execution results, generate task feedback results, and send the task feedback results back to the enterprise user. The memory management module is used to allocate an independent operating environment and independent memory storage space to the Agent instance corresponding to the enterprise user, and to determine the authorization context information provided to the target Agent according to the context access scope; The Agent collaboration module is used to transmit collaboration messages between different Agents according to a preset A2A communication protocol when the task instruction involves cross-user, cross-team or cross-department collaboration. The process of receiving task instructions from enterprise users and parsing these instructions to obtain task attribute information includes: Obtain the user identity information of the enterprise user and the task text information corresponding to the task instruction; The job attribute information corresponding to the enterprise user is determined based on the user identity information; The task text information is subjected to intent recognition and keyword extraction to obtain task intent information; Based on the job attribute information and the task intent information, determine the task type information and task priority information corresponding to the task instruction; The task attribute information is generated based on the job attribute information, the task type information, and the task priority information.

Images