A Smart Agent Permission Orchestration and Context Management System and Method for Enterprise Organization and Project Dual Domains
By establishing multi-domain relationships and dynamically managing permission scope in enterprise computer systems, the problems of inconsistent permissions and mixed data contexts in enterprise permission systems are solved, achieving project context continuity, data security, and agent skill reuse, thereby improving the stability and efficiency of enterprise-level agent management.
Patent Information
- Application Number
- CN202611367814.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-04
- Publication Date
- 2026-11-17
AI Technical Summary
Existing enterprise permission systems suffer from problems such as inconsistent permission judgments, mixed data contexts, inability to reuse agent skills, high risk of data leakage, and information lag in multi-level organizations, cross-departmental projects, and cross-departmental collaborations, making it difficult to operate stably at the enterprise scale.
By establishing a correspondence between organizational domains, project domains, work package state domains, permission scope domains, and agent execution domains within the computer system, verifiable access scopes and restricted contexts are dynamically generated. This separates departmental workspaces from project workspaces, dynamically manages permission scopes, unifies agent execution configurations, and enables versioning of permission scope tokens and cross-domain authorization.
It enables the maintenance of project context continuity during personnel transfers and project collaborations, reduces the risk of data leakage, minimizes the input of irrelevant information, improves the reusability and execution efficiency of intelligent agent skills, and ensures the accuracy and traceability of supervisor queries.
Smart Images

Figure CN123116240A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence agent management, enterprise collaborative computing, organizational relationship modeling, project workspace, access control, context orchestration, and data auditing. In particular, it relates to a system and method for enterprise-level secure execution and traceable collaboration of agents in multi-layered enterprise organizations and cross-departmental project scenarios. This is achieved through separation of organizational domain and project domain, compilation of permission scope, isolation of work package state, dynamic context loading, and versioned configuration of agents and skills. Background Technology
[0002] As large language models and tool invocation technologies are gradually applied within enterprises, employees are increasingly using AI dialogue to complete tasks such as data organization, document writing, code development, data analysis, and task planning. Enterprise managers also hope to use AI to quickly obtain the latest status of departments, teams, and projects, and further utilize AI to assist in risk identification, task assignment, cross-departmental collaboration, and performance statistics. Therefore, enterprise information systems are gradually evolving from the traditional "humans directly operating business systems" to "humans operating business systems through intelligent agents."
[0003] However, organizational relationships within a company do not always align with project collaboration relationships. Employees may administratively belong to a specific department, but in practical terms, they may simultaneously participate in multiple cross-departmental projects. Furthermore, a project may experience personnel reassignments, departures, replacements, and temporary support during its execution. If tasks, documents, AI-generated memories, and work results are all tied to individual accounts or departmental workspaces, personnel changes can easily lead to fragmented project contexts, unusable historical work, or the need for copying and migration. Conversely, simply placing all work data into a shared project space can unnecessarily expose private conversations, incomplete drafts, and sensitive information.
[0004] Existing enterprise access control systems are typically centered around roles or menu permissions, such as distinguishing between employees, supervisors, project managers, and administrators. However, "a user possessing a certain role" does not equate to "a user being able to read all subordinates' or all project data." In multi-layered organizations, dotted-line management, ad-hoc projects, and cross-departmental collaborations, the data scope of the same user may vary significantly depending on their actions. If permission checks are implemented separately for each business interface, each AI agent tool, and each search entry point, issues such as inconsistent rules, continued access to old scopes after cache expiration, and AI bypassing business interfaces to directly read data it shouldn't be able to access can easily arise.
[0005] Furthermore, typical AI chatbots often use a fixed session or a fixed knowledge base to maintain context. Providing the agent with all the data from an entire department, project, or even the entire enterprise at once will cause the context length to continuously increase, irrelevant information to occupy model input, old data to become mixed with new data, and increase the risk of unauthorized leaks. Deploying a resident agent for each department, position, or project node will introduce a large amount of resident processes, model sessions, and state synchronization overhead, making it difficult to operate stably at an enterprise scale.
[0006] Enterprises also contain numerous functional agents, such as code review agents, document generation agents, testing agents, financial analysis agents, and risk agents. If tools, prompts, and model configurations are hardcoded into each agent's implementation, the same skill cannot be reused by multiple agents; after a skill is updated, it is impossible to accurately determine which version was used in a particular historical execution; and it is difficult to uniformly restrict the range of tools and data that agents can access when security policies change.
[0007] On the other hand, managers want to directly ask AI questions like "What is the progress of a certain department this week?", "What are the risks of a certain project?", and "What is currently blocking a certain member?". If the supervisory AI directly reads employees' private chat logs, while it can obtain more information, it will violate the boundaries of employees' private workspaces and may treat unconfirmed speculations, drafts, or errors as official facts. Relying solely on human weekly reports, on the other hand, presents problems such as information lag, inconsistent formats, and difficulty in tracking evidence.
[0008] Therefore, a computer implementation mechanism is needed that can simultaneously handle administrative organizational relationships, cross-departmental project relationships, personal private work processes, shared project work results, and agent operation configuration. This mechanism would enable supervisors and project leaders to directly obtain structured and traceable work status through agents within their authorized scope, while ensuring that project context is not lost due to personnel changes and that private conversations, unsubmitted drafts, and cross-departmental sensitive data are not loaded by agents without authorization. Summary of the Invention
[0009] To overcome the shortcomings of the prior art, the present invention aims to provide an intelligent agent permission orchestration and context management system and method for enterprise organization and project dual domains. This system does not simply deploy multiple chatbots within an enterprise, but rather establishes a correspondence between organizational domain, project domain, work package status domain, permission scope domain, and intelligent agent execution domain within the computer system, and dynamically generates verifiable access scope and restricted context before each intelligent agent execution.
[0010] This invention separates departmental workspaces from project workspaces: the organizational domain expresses administrative affiliation and default management boundaries; the project domain expresses cross-departmental project members, project tasks, and shared project resources. Employees' private agent conversations remain within their private domains, and only work packages confirmed and submitted by the employee are converted into shared project objects. After submission, work packages retain their primary affiliation based on project and task identifiers, thus maintaining continuity of project knowledge and work status during personnel changes, while preventing the transfer of the original employee's private conversations to the new employee.
[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Enterprise access module (1): Used to identify the current enterprise tenant, login subject, terminal type and request link, and to pass enterprise identifier, user identifier, request identifier and target action to subsequent modules.
[0012] Organization Domain Management Module (2): Used to maintain organizational nodes at any level, such as groups, companies, business units, centers, departments, and teams, as well as organizational member relationships, direct management relationships, dotted line management relationships, and delegated management relationships. Organizational nodes preferably store the parent node identifier, hierarchical path, and hierarchical depth, enabling the system to simultaneously perform direct child node queries, organizational subtree queries, and ancestor node queries.
[0013] Project Domain Management Module (3): Used to independently maintain project workspaces, project members, project roles, project tasks, project milestones, project shared files, project work packages, and project status. Project members can come from multiple organizational nodes, and project permissions do not require changes to employees' administrative affiliations.
[0014] Work Package Status Control Module (4): This module separates employee private agent dialogues, work records extracted by the agent, work package drafts, submitted work packages, reviewed work packages, and withdrawn work packages. Work package drafts are only visible to the employee and explicitly authorized collaborators; after a work package is submitted, a versioned snapshot is generated, becoming a formal shared source of facts that can be used by the supervisor agent and the project agent.
[0015] Permission scope compilation module (5): It is used to separate the role ability of "what actions a user can perform" from the data scope of "what data the action can be applied to", and to calculate the set of organization nodes, user set, project set, resource type set and action set by integrating organizational domain relationship, project domain relationship, explicit authorization and explicit denial rules, and generate permission scope token.
[0016] Dynamic Context Loading Module (6): Used to read the permission scope token before actually calling the large model or tool, select only the content related to the current query target from the allowed data sources, and further filter according to resource status, version, update time, security level and visibility scope, and finally form a restricted context for the execution of this intelligent agent.
[0017] Intelligent Agent Execution Engine (7): Used to instantiate employee intelligent agents, supervisor intelligent agents, project intelligent agents, security intelligent agents, or other functional intelligent agents as needed based on the request scenario. The unified execution engine does not require each organizational node to have a permanent independent intelligent agent, but rather generates different logical intelligent agents by loading different execution configurations and contexts.
[0018] Skill Configuration Module (8): Used to maintain reusable skills, tools, model routes, prompt word templates and output protocols, and bind skill identifiers and skill versions to agent execution configurations, enabling enterprises to configure mature functions as skills to different agents, while also tracing the version of historical executions.
[0019] Cross-domain authorization module (9): Used to process access requests across departments, organizational nodes or temporary projects, and to generate explicit authorization records for the access subject, target project, permitted actions, resource scope, approval subject, validity period and revocation status, and to automatically expire when the authorization expires or the project ends.
[0020] Source and Evidence Module (10): Used to output associated source identifiers, evidence references, update times, submitter snapshots and optional data digest values to work packages, task statuses, project statuses and agents, enabling agent conclusions to be traced back to specific shared data objects.
[0021] Audit and Version Module (11): Used to record permission calculation, agent context loading, work package submission, cross-domain authorization, skill version, tool call and output summary, and maintain organization version, project version, authorization version and agent configuration version so that cache and permission scope expire in a timely manner after structural changes.
[0022] The statistics aggregation module (12) is used to generate statistical information such as team or project progress, blockage, risk, update frequency and completion status based on submitted work packages, task events, project milestones and review events; the statistical process does not use employee private AI dialogue as direct calculation input.
[0023] The permission scope of the organization domain and project domain is calculated by combining the allowed set and the denied set. For the request subject, target action, and query target, the system calculates the organization allowed scope, project allowed scope, and explicit authorization scope respectively. After taking the union of the above scopes, the explicit denied scope is subtracted. Then, a second trimming is performed based on resource type, target object, resource status, and version information to obtain the set of data objects that can actually be loaded in this intelligent agent execution.
[0024] The permission scope token preferably includes at least the enterprise tenant identifier, user identifier, action type, set of organization node identifiers, set of user identifiers, set of project identifiers, set of resource types, version tuple, issuance time, expiration time, and signature value; wherein, the version tuple consists of organization version, project version, authorized version, and agent configuration version. When any current version is inconsistent with the corresponding version in the token, the system stops using the original permission scope token and re-executes the scope calculation.
[0025] The submitted work package preferably includes at least a work package identifier, enterprise identifier, project identifier, task identifier, submitter snapshot, work package content, evidence reference set, visibility scope, status, version number, submission time, and update time. After the work package is submitted, the project shared view references this versioned object; when the work package is revised, a new version is generated and the version relationship is preserved, rather than overwriting all historical data.
[0026] The agent execution configuration preferably includes at least an agent configuration identifier, agent role, configuration version, model route, prompt word version, skill binding relationship, tool permissions, context template, and output structure. The skill binding relationship stores the skill identifier and skill version, enabling the same functional skill to be reused by different logical agents, while ensuring that audit events can reconstruct the current execution configuration.
[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By separating the organizational domain from the project domain, administrative organizational adjustments and project collaboration relationships are not mutually exclusive; shared project objects are primarily attributed to projects and tasks, and can be retained even after personnel transfers, departures, or replacements, thereby reducing project context breaks and data migration caused by account changes.
[0028] 2. By dividing private agent dialogues, drafts, and submitted work packages into different resource states, supervisors and project managers can obtain confirmed official work statuses without reading employees' private dialogues, reducing unnecessary data exposure at the computer access control level.
[0029] 3. By compiling permission scopes and versioning scope tokens, organizational relationships, project relationships, explicit authorization and denial rules are uniformly converted into executable data scopes, avoiding agents from directly using high-privilege database connections or repeatedly implementing inconsistent permission logic in multiple interfaces.
[0030] 4. By using dynamic context loading, only the data objects required and authorized for this request are provided to the large model or tool. This reduces irrelevant context input, avoids mixing old and new data, and promptly invalidates old tokens and caches when the scope of the organization or project changes.
[0031] 5. Through a unified intelligent agent execution engine and versioned execution configuration, different logical intelligent agents can be generated as needed, without the need to maintain a resident intelligent agent process for each department or project, and skills, tools, model routes and prompt words can be reused and upgraded in a unified manner.
[0032] 6. By associating work packages, tasks, projects, and evidence objects with the results of the intelligent agent through the source and evidence module, and synchronously outputting the update time and credibility status, the natural language query results of the supervisor can be drilled down for verification, auditing, and reproduction.
[0033] 7. By explicitly authorizing across domains, cross-departmental projects can obtain limited, expirable, and revocable access without changing the administrative organizational structure; permissions are recompiled after authorization expires, reducing the risk of data leakage caused by long-term accumulated permissions. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of the overall structure of the enterprise-level intelligent agent management system provided by the present invention.
[0035] Figure 2 This is a schematic diagram illustrating the relationship between the organization domain and the project domain provided by the present invention.
[0036] Figure 3 This is a schematic diagram of the work package state isolation and sharing process provided by the present invention.
[0037] Figure 4 This is a schematic diagram illustrating the process of pruning permissions and loading dynamic context for managerial queries provided by the present invention.
[0038] Figure 5 This is a schematic diagram of the cross-department explicit authorization and scope token generation process provided by the present invention.
[0039] Figure 6 This is a schematic diagram illustrating the continuity of project context in the event of personnel changes, as provided by the present invention.
[0040] Figure 7 This is a schematic diagram of the versioned execution configuration of intelligent agents and skills provided by the present invention; Figure 8This is a schematic diagram of the interaction architecture of the enterprise-level intelligent agent management system module provided by the present invention.
[0041] In the diagram: 1-Enterprise access module, 2-Organization domain management module, 3-Project domain management module, 4-Work package status control module, 5-Permission scope compilation module, 6-Dynamic context loading module, 7-Agent execution engine, 8-Cross-domain authorization module. Detailed Implementation
[0042] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The following embodiments are used to illustrate the principle and feasible implementation of the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the core idea of the present invention, the database type, backend language, model provider, terminal form, and specific field names can all be replaced.
[0043] like Figure 1 As shown, the system in this embodiment is deployed in an enterprise server or cloud computing environment. Employee terminals, supervisor command terminals, and enterprise management terminals access the backend service through a unified gateway. The backend isolates data by enterprise identifier and assigns a unique request identifier to each request. After a request enters the system, it first parses the logged-in user, enterprise tenant, target action, and target resource, and then enters the permission scope compilation module 5. The intelligent agent execution engine 7 can only read data through controlled data interfaces and controlled tool interfaces, and cannot bypass the permission scope token to directly read business tables using a global database account.
[0044] In a preferred embodiment, the enterprise access module 1 classifies access subjects into employees, supervisors, project managers, enterprise administrators, and system configuration administrators. These roles only define action capabilities and are not directly equivalent to data scope. For example, a supervisor role can perform supervisor query actions, but the objects they can query are still determined by the organizational domain, project domain, and explicit authorization; an enterprise administrator can configure the organizational structure and intelligent agents to execute configurations, but does not automatically gain access to read employees' private conversations.
[0045] like Figure 2 As shown, the organization domain management module 2 and the project domain management module 3 maintain independent data structures. Organization nodes in an organization domain include at least the following information: node identifier, enterprise identifier, parent node identifier, hierarchical path, hierarchical depth, responsible user identifier, status, and version number. Organizational member relationships record the affiliation between users and organization nodes, while management relationships record direct, dotted, temporary, project, or delegated management relationships. The system can quickly read direct subordinates based on the parent node identifier, read the organization subtree based on hierarchical path prefixes or closure relationships, and perform hierarchical restrictions and display based on hierarchical depth.
[0046] The project workspace within the project domain includes at least the project identifier, company identifier, project name, project status, project manager identifier, project version, and visibility policy; project member relationship records include user identifier, project role, join time, leave time, and status; task object records include task identifier, project identifier, parent task identifier, executor identifier, task status, priority, deadline, and version number; work packages record the formal working status bound to the project and tasks. Project members can come from different departments, therefore, administrative organizational relationships do not need to change due to participation in the project.
[0047] Organizational workspaces are used to store departmental policies, shared departmental knowledge, organizational task templates, and organizational management data; project workspaces are used to store specific project tasks, work packages, project documents, project decisions, and project events. The two types of workspaces are independent of each other in terms of data primary keys, permission sources, and lifecycles. When an employee moves from department A to department B, their organizational domain changes, but the project workspaces they are still authorized to participate in do not need to be migrated; when an employee leaves a project, only the project membership is removed, and the formal work packages already submitted in the project are not deleted.
[0048] like Figure 3 As shown, when employees process tasks, they first interact with their employee agents in a private workspace. Private conversations can include exploratory reasoning, unverified information, and personal notes. The system allows the employee agent to extract work records from task fields, private conversations, user additions, and uploaded evidence, and generate a draft work package. At this point, the draft is in either draft or generated draft status and is not indexed by the supervisor. Employees can continue to modify completed items, blockages, risks, and next steps, and add supplementary evidence. Only after the employee performs a confirmation submission does the system generate a versioned work package snapshot with a submitted status.
[0049] During the submission process, the work package status control module 4 verifies the existence of the project identifier, task identifier, submitter's current project member relationships, and evidence references, and then generates a new work package_version number. The system can calculate hash values for the content summary and evidence list to determine whether subsequent versions have changed. The submission state object is written to the project's shared index, while the private dialogue index remains in the private repository and is not automatically made public due to the generation of a work package.
[0050] If an employee withdraws a work package, the system will update the shared visibility status to "withdrawn" and exclude that version from the supervisor's default search set; the audit log will still retain the withdrawal event. If corporate policy requires retaining historical submissions for compliance audits, auditors can read the corresponding historical versions with dedicated audit permissions, but regular supervisor and project queries will still use the currently valid submission as the source of fact.
[0051] The permission scope compilation module 5 performs scope calculations before each sensitive read and agent query. First, it determines whether the current user has the capability to perform the target action based on the role mapping; second, it calculates the organizational scope, such as the user's own scope, direct subordinate scope, specified organizational node scope, organizational subtree scope, and enterprise scope; then, it calculates the project scope, such as the set of specified project scopes that the user can access as a project member, project leader, or explicitly authorized; finally, it reads the temporary authorization from the cross-domain authorization module 9 and applies explicit denial rules.
[0052] Taking a supervisor querying the R&D division as an example, the system allocates the R&D division and its descendant organizational nodes based on the supervisor's management relationships or organizational scope, expanding the corresponding user set; then, it determines the set of projects in which these users participate and which the supervisor has access permissions from the project domain. If the supervisor is also the head of a cross-departmental project, that project can be added as an additional project scope. Subsequently, the system generates a permission scope token based on the supervisor's query action and resource types such as work packages and evidence.
[0053] Permission scope tokens can be either server-stored, unforgeable session objects or signed structured tokens. The key is that they carry the scope and version. If the organizational tree is adjusted, project members change, temporary authorizations are revoked, or agent configurations are upgraded, the system increments the corresponding version. The dynamic context loading module 6 compares the version tuple before using the old token; if an inconsistency is found, it terminates the old context loading and recalculates, thus preventing the old administrator from continuing to access data that has been moved out of its scope before the permission cache has naturally expired.
[0054] like Figure 5 As shown, cross-departmental access is completed without directly modifying the user's global role. After the applicant selects the target project or resource, the system generates a cross-domain authorization request. The request includes at least the applicant identifier, target project identifier, resource type, action type, scope of application, reason for application, and authorization duration. Upon approval, a cross-domain authorization record is generated. This cross-domain authorization record can be limited to reading only a specific project's work package, calling only a specific functional agent, or being valid only within a specified time period. After the cross-domain authorization expires, the authorization version increments, and the permission scope tokens dependent on that authorization immediately become invalid.
[0055] After receiving the permission scope token and query text, the dynamic context loading module 6 first identifies the query target. For example, "What risks does the platform R&D center currently face?" corresponds to the organizational node target; "What obstacles remain in the authentication upgrade project?" corresponds to the project target; and "What has Zhang done recently?" corresponds to the user target. The target parsing result is only used to further narrow down the scope and cannot expand the authorization scope of the token itself.
[0056] For organization node queries, module 6 reads the valid members in the authorized organization subtree and further obtains the committed work packages of these members in the allowed project set; for project queries, module 6 directly reads the tasks, committed work packages, project milestones and authorization evidence in the project; for user queries, module 6 only reads the committed work packages of the user within the current supervisor's visibility range, and does not read their private conversations, private drafts and work outside the project.
[0057] To control context length, Module 6 can sort by update time, task priority, work package status, risk level, and query relevance, and collapse older versions. For example, it can retain only the current valid commit version for each task, while converting the most recent status changes into short event summaries. For work packages with conflicting sources, the system does not directly overwrite them, but adds a conflict marker to the context, causing the supervisor agent to explicitly indicate inconsistencies in its responses.
[0058] like Figure 4 As shown, after the supervisor inputs a natural language question, the system sequentially executes target parsing, permission-scoped compilation, state filtering, version verification, context assembly, and supervisor agent invocation. The supervisor agent's system constraints explicitly require: it can only use the input restricted context; it cannot claim to have read employees' private conversations; and its conclusions must distinguish between verified facts, source-based inferences, and information gaps. The output may include a summary, risk items, blocking items, next steps, source, credibility, and information gap fields.
[0059] Trustworthiness status can be calculated based on source coverage, evidence coverage, update time, and the number of conflicts. For example, the system considers tasks with submitted work packages within the last 72 hours and at least one valid piece of evidence as high-freshness sources; tasks that have not been updated for a set time are marked as expired; if multiple sources conflict for the same milestone status, the trustworthiness status is lowered and the supervisor is required to drill down to investigate. Trustworthiness status is not an evaluation of an employee's individual ability, but rather a machine-readable description of the data support for the agent's response.
[0060] like Figure 7 As shown, the Agent Execution Engine 7 instantiates logical agents on demand through agent execution configuration. A supervisor agent configuration can be bound to a management summary skill module, a risk aggregation skill module, a project query skill module, and a source verification skill module; an employee agent configuration can be bound to a task organization skill module, a document writing skill module, and a work package generation skill module. Each skill module is stored as an independent configuration object, storing its name, input mode, output mode, allowed tools, version, and release status.
[0061] During each execution, the agent execution engine 7 first reads the current configuration version and skill version, and then combines them with the permission scope token to generate a running instance. Even if the same user calls the same supervisor agent twice consecutively, if the enterprise administrator upgrades the risk identification skill module between the two calls, the audit log can still record the skill version used in each execution separately. If a skill module needs to call a code repository or documentation system, the tool call must still carry the current scope token or be re-authorized by the backend based on the token.
[0062] The security agent or security policy module can perform sensitive information checks before and after context loading. For example, for evidence containing customer keys, passwords, or highly sensitive fields, the system can prevent it from entering the large model context according to the resource security level, and only provide a de-identified digest to the supervisor agent; if the user's permissions only allow viewing project progress and not viewing the original files, the context loading module only provides a work package digest and source identifier, without providing the complete evidence content.
[0063] This invention also supports hierarchical task allocation and feedback. Enterprise leaders can create upper-level tasks in the organizational or project domain. After these upper-level tasks are broken down into lower-level tasks, each executor completes the work and submits a work package through a personal intelligent agent. The supervisory agent does not need to read all lower-level dialogues; instead, it summarizes the submitted work packages to form departmental or project-level work packages. When higher-level supervisors query again, they can prioritize loading the already summarized upper-level work packages and drill down to the specific source as needed, thereby reducing the context volume during high-level queries.
[0064] In one implementation, each work package contains a source work package identifier set field, used to record the lower-level work packages referenced by the summary work package. When a department head generates a "Weekly Progress Report for the R&D Center," the system writes the identifiers of multiple employee work packages into the source relationship; when a business unit head further summarizes work packages from multiple centers, a second-level reference is established. This forms a traceable directed relationship graph of work packages, enabling the high-level agent's response to trace back from the summary conclusion to the underlying submitting object.
[0065] The statistics aggregation module 12 takes work package submission events, task completion events, milestone events, review events, and risk events as inputs to generate metrics such as update timeliness rate, task completion rate, number of blocked events, number of risks, milestone achievement status, and work package coverage. If the enterprise uses these metrics for performance reference, the system can retain the source of the metrics and the calculation time; however, this embodiment does not require reading employees' private conversations, nor does it use artificial intelligence to infer private processes to directly generate formal performance results.
[0066] like Figure 6As shown, in the event of personnel changes, the project context remains independent. Assume that the original executor, A, was responsible for Task A within Project A and had submitted three work packages. Later, the original executor, A, was transferred from the project, and a new executor, B, took over. The project administrator updates the original executor A's project member relationship to a "left" status, adds the new executor B to the project, and reassigns Task A. The permission scope compilation module 5 generates a new project scope for the new executor B. The new executor B can read valid work packages, project decisions, and authorization evidence within the scope of Project A and Task A, thus continuing to work; the original executor A's chat logs and unsubmitted drafts in their private workspace remain solely the property of the original executor and are not considered project handover data.
[0067] If a project requires the retention of important knowledge from certain private processes, the original executor, A, can convert the relevant content into a new work package or project document and make an explicit commit before leaving. In this case, what is shared are the confirmed new resource objects, rather than a complete copy of the original private conversations to the new executor, B. This mechanism separates "project knowledge transfer" from "personal private work processes" at the data level.
[0068] Below is an example of a cross-departmental project. The company comprises a Research and Development Division, a Marketing Division, and a Delivery Center. Project B involves all three departments. Researcher A is responsible for interface development, Marketing Personnel B is responsible for confirming customer requirements, and Delivery Personnel C is responsible for deployment. The three have different administrative supervisors, but the project manager for Project B gains access to the submitted work packages of Project B through the project domain. When the project manager asks, "What is the biggest delivery risk for Project B this week?", the system only loads the valid work packages and project evidence for Researcher A, Marketing Personnel B, and Delivery Personnel C within the scope of Project B; it does not load their work in other projects or their private conversations.
[0069] If the R&D department head is not the person in charge of Project B, but needs temporary assistance in troubleshooting interface issues for Project B, the head can initiate a cross-departmental access request to read resources in Project B that are classified as work packages and tagged "interface," with a 24-hour authorization period. Once approved, the access token will only cover the specified resource range within Project B. The head's agent can query interface-related work packages during the authorization period, but cannot read market requirement documents or delivery deployment evidence; the token expires after the authorization period.
[0070] In another embodiment, a company leader queries "blockages in all key projects this quarter." The system first obtains the set of allowed projects based on the leader's company scope or a custom project scope, then prioritizes loading the most recent project-level summary work package and high-risk events for each project. If the update time of a project's summary work package exceeds a threshold, the system adds an information gap warning and can drill down to load its underlying committed work packages as needed. This hierarchical loading mechanism can reduce the amount of raw data that must be entered into the model for a single query, even with a large number of employees and projects.
[0071] Organizational domains and project domains can be implemented using relational databases, graph databases, or a combination of both. Organizational hierarchies can use parent node identifiers plus materialized paths, or closure tables to record ancestor-descendant relationships; project tasks can use parent task identifiers to form a task tree, or dependency edges to form a directed acyclic graph. Regardless of the specific storage method, as long as the organizational domain and project domain maintain their respective ownership and permission relationships, and are combined through unified permission scope compilation and context loading before the agent executes, they fall within the scope of this invention.
[0072] The intelligent agent execution engine 7 can access one or more large language models, and can also assign different models to different skill modules according to task type. For example, a low-cost model is used for summarizing tasks, a high-capability model is used for complex risk analysis, and a proprietary enterprise model is used for sensitive tasks. Model routing results are written to audit events for easy subsequent reproduction. Tools can include file systems, knowledge bases, task systems, code repositories, testing systems, and internal enterprise application programming interfaces, but each tool call is subject to action permissions and resource scope constraints.
[0073] To improve concurrency performance, scoped compilation results can be cached as scoped snapshots. Snapshots save the subject, action, and version tuples. When organization members, project members, or authorization relationships change, the Audit and Versioning module 11 releases a scope invalidation event, invalidating the affected snapshots. The system can recalculate only the scope for the affected subject without requiring the entire enterprise cache to be cleared.
[0074] Work packages and project events can also be published using event streams or message queues. The statistics aggregation module 12 subscribes to events such as submitted status, reviewed status, task completion, milestone changes, and risk changes, incrementally updating team and project statistics. When the supervisor agent queries, it reads the latest aggregation results and can further retrieve the corresponding work package through a set of source identifiers. In this way, statistical results and natural language responses share the same source system, avoiding situations where reports and agent responses reference completely different data sources.
[0075] In terms of security auditing, audit events should ideally record the event identifier, request identifier, enterprise identifier, operating entity identifier, operating entity type, action type, target type, target identifier, permission scope token identifier, set of installed resource identifiers, agent configuration version, skill version set, tool call record, output summary, decision result, and creation time. For important work packages, a content summary value can be further saved to determine whether subsequent data is consistent with the audit content at that time.
[0076] In summary, this invention separates the organizational domain from the project domain, making the project work context no longer dependent on a specific person's account; it separates private processes from shared facts through work package state control; it converts organizational relationships, project relationships, and explicit authorizations into the minimum data scope executable by the agent through permission scope compilation and dynamic context loading; and it achieves function reuse and traceable execution through a unified agent execution engine and skill versioning configuration, thereby forming an agent management mechanism suitable for multi-level enterprise management and cross-departmental collaboration.
Claims
1. A smart agent permission orchestration and context management system for both enterprise organization and project domains, characterized in that, include: Organization Domain Management Module (1): Used to maintain enterprise organization domains based on enterprise identifier, organization node identifier, parent node identifier, hierarchical path, hierarchical depth, organization member relationship and management relationship. The organization domain is used to determine the user's administrative affiliation, default management relationship and organization data scope. Project Domain Management Module (2): Used to maintain project workspace, project member relationships, project task relationships and project resource relationships independently of the organization domain, so that shared data objects in the project workspace use project identifier and task identifier as the primary owner identifier, rather than the current executor's account as the sole primary owner; Work package status control module (3): used to divide employee private agent dialogue, work record, work package draft and submitted work package into different resource statuses, and generate a versioned work package snapshot bound to project identifier and task identifier when the employee confirms submission; Permission scope compilation module (4): For the target action initiated by the user, it calculates the set of allowed organizational nodes, user set, project set and resource type set based on user role, organizational domain relationship, project domain relationship, explicit authorization relationship and explicit denial relationship, and generates a permission scope token containing version information and validity period; Dynamic Context Loading Module (5): Used to select authorized data objects matching the query target from the organization domain and project domain according to the permission scope token, filter data objects in private state, draft state, withdrawn state, expired version or beyond the authorization scope, and form a restricted agent context; Agent Execution Engine (6): Used to combine the restricted agent context with the versioned agent execution configuration, instantiate one or more of the following agents as needed: employee agent, supervisor agent, project agent, security agent or functional agent, and call the bound models, prompts, skills and tools to execute the target task; Source and Audit Module (7): Used to associate data source, update time and trusted status in agent output, and record permission scope, context object identifier, agent execution configuration version, tool call and output summary; Cross-domain authorization module (8): Used to handle data access requests across organizational nodes, departments, projects, task scopes, or resource types. When the access target does not belong to the default organizational scope or project member scope of the requesting subject, a cross-domain authorization application is generated or read, and the following information is recorded: requesting subject, source domain, target domain, target project or resource, allowed actions, resource type, scope constraints, authorization purpose, approving body, effective time, expiration time, transferable flag, and revocation status.
2. The system according to claim 1, characterized in that: The organization domain management module (1) maintains organization nodes that include at least the enterprise identifier, node identifier, node type, parent node identifier, hierarchical path, hierarchical depth, person in charge identifier, and organization version; the project domain management module (2) maintains project workspaces that include at least the enterprise identifier, project identifier, project status, project person in charge, project members, task set, work package set, and project version.
3. The system according to claim 1, characterized in that: When a project member leaves, is transferred, replaced, or temporarily transferred, the project domain management module (2) only updates the project member relationship and corresponding permission scope, without migrating the original member's private intelligent agent dialogue; after obtaining project permissions, the new member can load the work package, task status and authorization evidence submitted in the project, so that the project shared context remains continuous before and after personnel changes.
4. The system according to claim 1, characterized in that: The permission scope token includes at least the enterprise identifier, the main user identifier, the action identifier, the allowed set of organizational nodes, the allowed set of users, the allowed set of projects, the allowed resource types, the organization version, the project version, the authorized version, the agent configuration version, the issuance time, and the expiration time; when any of the versions is inconsistent with the current version, the dynamic context loading module (5) prohibits the continued use of the original permission scope token and triggers recompilation.
5. The system according to claim 1, characterized in that: The system also includes a cross-domain authorization module (8), which is used to receive cross-departmental project access applications, form an explicit authorization record by including the applicant, target project, target resource, allowed action, authorization scope, effective time, expiration time and approval subject, and use the explicit authorization record as input to the permission scope compilation module (4); and invalidate the corresponding permission scope when the authorization expires, the project ends, the authorization is revoked or the version changes.
6. The system according to claim 1, characterized in that: The work package status control module (3) saves at least the task identifier, project identifier, submitter snapshot, completed items, in-progress items, blocked items, risk items, next steps, evidence references, version number, submission time and visibility range for submitted work packages; the shared retrieval data source of the supervisor agent and the project agent by default only includes submitted work packages and authorized evidence, and does not include employee private agent dialogues and unsubmitted drafts.
7. The system according to claim 1, characterized in that: The agent execution configuration includes at least the agent role, configuration version, model routing parameters, prompt word version, skill identifier set, skill version set, tool action set, context template, and output mode; the agent execution engine (6) constructs agent instances as needed based on the agent execution configuration and permission scope token during each execution, instead of maintaining an independent resident agent process for each organization node or project node.
8. The system according to claim 1, characterized in that: The source and audit module (7) generates a trusted status based on one or more of the following: source quantity, source update time, work package completeness, evidence coverage, and conflict detection results. It also outputs a summary, risk, blockage, information gap, source identifier, update time, and trusted status in the supervisor query results. At the same time, it writes the query object, permission scope, loaded resource identifier, and output summary into the audit event.
9. A method for intelligent agent permission orchestration and context management in a dual domain of enterprise organization and project, characterized in that, include: Step 1: Establish independent organizational domains and project domains. The organizational domain is used to maintain administrative organizational relationships and default management scope, while the project domain is used to maintain project workspaces, project members, project tasks, and project shared resources. Step 2: Receive private agent interactions, work records, and evidence information from employees in the task context, generate a draft work package, and create a snapshot of the submitted work package bound to the project identifier and task identifier after the employee confirms the submission. Step 3: Receive user-initiated agent queries or agent execution requests, and parse the corresponding actions and target objects; Step 4: Calculate the permission scope and generate a permission scope token based on user roles, organizational domain relationships, project domain relationships, explicit authorization relationships, and explicit denial relationships; Step 5: Based on the permission scope token, filter authorized data objects from the organization domain and project domain, perform resource status filtering and version verification, and form a restricted agent context; Step 6: Based on the restricted agent context and versioned agent execution configuration, instantiate the target agent on demand, and call the bound skills and tools to obtain the execution result; Step 7: Associate the execution result with the source identifier, update time, and trusted status, and record the permission scope, context object, execution configuration version, and output summary.
10. The method according to claim 9, characterized in that: In step 4, the organizational scope, project scope, and explicit authorization scope are calculated separately. The union of the allowed scopes is calculated, and the difference of the explicit rejection scope is calculated. The scopes are then further pruned according to the target action and resource type. If no matching allowed rule exists, a rejection result is returned.
