A multi-agent cooperative communication system and method

Abstract

This invention relates to the field of artificial intelligence technology and provides a multi-agent collaborative communication system and method. The system includes: multiple personal AI agent modules; a communication agent gateway; a shared context storage module; and a distributed permission management module. The communication agent gateway is configured to directly route and send messages generated by the personal AI agent modules of the first user to the personal AI agent modules of the second user according to the communication format of "first user → first user's personal AI agent module → second user's personal AI agent module → second user". This invention enables the personal AI agent modules of the first user and the second user to directly establish a communication link, realizing the direct transfer of work content, the sharing and continuation of context, and solving the interoperability problem across platforms and large models, while ensuring permission and data security during communication.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a multi-agent collaborative communication system and method. Background Technology

[0002] With the rapid development of artificial intelligence technology, personal AI agents have gradually become important tools for users to handle daily tasks and assist in decision-making. Analysis reveals that existing technologies (i.e., the current model of personal AI being used independently and relying on third-party communication software for interpersonal communication) have the following objective defects and shortcomings:

[0003] 1. Lack of direct communication between personal AIs: Each user's personal AI can only interact with that user and cannot proactively convey work content or task instructions directly to another user's personal AI. This forces workflows to be interrupted, requiring human users to act as "messengers" to manually relay and interpret messages.

[0004] Second, reliance on third-party communication software leads to information redundancy and distortion: Users need to send the results generated by their personal AI to other users through third-party communication software, and then the receiving users read and input their personal AI. In this process, the information undergoes multiple transformations: "AI → Sender User → Communication Software → Receiving User → Receiving AI", which can easily cause information omissions, semantic misinterpretations, or formatting errors, thus resulting in information errors.

[0005] Third, context cannot be shared and continued among AIs: Because each user's personal AI is isolated from each other, they are unaware of each other's task history, decision-making basis, intermediate data, and other contextual information. When multiple users collaborate based on the output of their respective AIs, the results are often based on inconsistent or incomplete background information, leading to severe information asymmetry. This requires a large amount of manual correction and communication coordination, resulting in low team collaboration efficiency.

[0006] IV. Information Security Risks and Access Control Issues When Sharing Personal AI: Directly sharing a personal AI with another user could lead to issues such as unauthorized access, data leaks, and disorganized operation logs. Personal AIs typically store the owner's private data, preferences, and some work secrets; simple sharing cannot meet enterprise-level security and compliance requirements.

[0007] Fifth, large models from different companies cannot communicate across models: Existing large AI models are developed by different companies, each using its own closed interfaces, data formats, and interaction logic. Even if two users both use AI, if they come from different vendors (e.g., one uses model A, and the other uses model B), their individual AIs cannot establish any form of direct communication. This technological barrier fundamentally prevents cross-platform and cross-ecosystem collaboration of intelligent agents.

[0008] In summary, these shortcomings of existing technologies lead to a core consequence: the inability to truly achieve AI-based team collaboration. Members still rely on inefficient human intermediaries and coordination, with AI existing only as a standalone tool and failing to integrate into multi-person collaborative workflows. This is a significant deficiency in the current AI technology architecture. Summary of the Invention

[0009] The purpose of this invention is to provide a multi-agent cooperative communication system and method, which aims to solve the above-mentioned technical problems.

[0010] This invention is implemented as follows: a multi-agent cooperative communication system, comprising:

[0011] Multiple personal AI agent modules are deployed on the corresponding user's terminal device or cloud. These modules are used to interact with the user in natural language and maintain the user's task context and private data.

[0012] A communication proxy gateway, coupled to multiple personal AI agent modules, is used to establish direct communication links between personal AI agent modules of different users, and to perform protocol adaptation and cross-model semantic conversion on messages transmitted in the communication links.

[0013] A shared context storage module, coupled to the communication proxy gateway, is used to persistently store shared context information generated by multiple personal AI agent modules during collaboration.

[0014] The distributed permission management module is coupled to the personal AI agent module, the communication agent gateway, and the shared context storage module, respectively, and is used to authorize and verify communication requests and data access requests between personal AI agent modules based on a preset permission policy.

[0015] The communication proxy gateway is configured to route and send messages generated by the personal AI agent module of the first user directly to the personal AI agent module of the second user in the communication format of "first user → first user's personal AI agent module → second user's personal AI agent module → second user".

[0016] Furthermore, the communication proxy gateway specifically includes:

[0017] The unified identity resolution unit is used to resolve a user's identity identifier into a network-addressable agent address;

[0018] The session management unit is used to establish, maintain, and release logical communication sessions between multiple personal AI agent modules;

[0019] The protocol adaptation unit is used to convert the private interface protocols used by different personal AI agent modules into a unified standardized interaction protocol.

[0020] A cross-model translation engine is used for bidirectional semantic and data format conversion between the first major model used by the first user's personal AI agent module and the second major model used by the second user's personal AI agent module.

[0021] Furthermore, the shared context storage module adopts a concurrent synchronization mechanism based on conflict-free copy data type to maintain data consistency when multiple personal AI agent modules simultaneously perform read and write operations on the shared context.

[0022] Furthermore, the shared context storage module also supports snapshot and rollback functions, allowing any individual AI agent module to restore the collaborative context to a specified historical state.

[0023] Furthermore, the cross-model translation engine employs meta-knowledge distillation techniques between large models to pre-train a lightweight semantic alignment network. This semantic alignment network takes the message output of the source model as input and outputs a message representation that the target model can understand.

[0024] Furthermore, the distributed permission management module specifically includes:

[0025] The identity authentication submodule is used to verify the legitimate identity of the individual AI agent module that initiated the communication request;

[0026] The policy decision submodule is used to store and execute the preset permission policies for each user; the permission policies include at least one of the following: allowed peer users for communication, allowed data range for access, and allowed operation types;

[0027] The operation audit submodule is used to record all communication and data access behaviors across personal AI agent modules;

[0028] The message sending and receiving and acknowledgment unit is used to support three communication modes: point-to-point, multicast, and broadcast, using asynchronous message queues or reliable UDP protocols, and provides message delivery acknowledgment, retransmission, and timeout handling mechanisms.

[0029] Furthermore, the personal AI agent module also includes an intent classifier, which is used to automatically identify whether the user's input command needs to trigger cross-AI communication, and automatically trigger the communication agent gateway when needed.

[0030] Another objective of this invention is to provide a multi-agent cooperative communication method, implemented through the aforementioned multi-agent cooperative communication system, comprising the following steps:

[0031] S1. The first user's personal AI agent module receives and parses the first user's natural language instructions, and parses out the work content and task objectives that need to be collaborated with the second user's personal AI agent module.

[0032] S2. The first user's personal AI agent module calls the communication agent gateway to initiate a request to establish a communication connection with the second user's personal AI; the request carries the identity identifier of the first user, the identity identifier of the second user, and the type of communication intent requested.

[0033] S3. Verify whether the request complies with the first user's sending permission and the second user's preset receiving policy through the distributed permission management module. If the verification fails, return insufficient permission information to the first user's personal AI agent module and terminate the process; if the verification is successful, continue to step S4.

[0034] S4. The identity of the second user is resolved to the network address of the second user's communication proxy gateway through the communication proxy gateway, and an encrypted transport layer secure connection is initiated.

[0035] S5. After the connection is established, the first user's communication proxy gateway will convert the internal message format of the first user's personal AI agent module into a standardized interactive message.

[0036] S6. Send the converted standardized interactive message to the second user's communication proxy gateway;

[0037] S7. After receiving the standardized interaction message, the second user's communication proxy gateway performs the reverse transformation of step S5 and delivers the final message to the second user's personal AI agent module.

[0038] S8. After receiving the final message, the second user's personal AI agent module automatically converts it into a form that the second user can understand and presents it to the second user according to the notification rules preset by the second user.

[0039] Furthermore, step S3 specifically includes:

[0040] Sub-step P1: The first user's distributed permission management module retrieves the first user's private key from local secure storage and signs the request content;

[0041] Sub-step P2: Send the signed request to the second user's distributed permission management module;

[0042] Sub-step P3: The second user's distributed permission management module verifies the signature validity, and then queries the second user's policy database to determine whether the triple "first user → second user → intent type" is allowed;

[0043] Sub-step P4: If allowed, generate a temporary session token and return the token to the first user's communication proxy gateway; if not allowed, return a rejection message and the reason.

[0044] In step S5, if the first major model used by the first user's personal AI agent module is different from the second major model used by the second user's personal AI agent module, then a bidirectional mapping conversion of semantics and format is performed through a cross-model translation engine.

[0045] The cross-model translation engine performs the following sub-steps:

[0046] Sub-step T1: Obtain a series of message vectors or text fragments output by the source large model used by the source personal AI agent module and the model identifier of the target large model used by the target personal AI agent module;

[0047] Sub-step T2: Load the semantic alignment mapping table corresponding to the target large model from the pre-stored knowledge base;

[0048] Sub-step T3: Based on the semantic alignment mapping table, perform word segmentation, intent extraction, and structured representation on the source message to generate an intermediate language representation that does not depend on a specific large model;

[0049] Sub-step T4: Based on the intermediate language representation, and combined with the prompt template style and parameter format of the target large model, generate instruction strings or structured data that the target large model can directly understand and execute;

[0050] Sub-step T5: If the target large model does not support certain special functions of the source large model, it will automatically be downgraded to text description or simulation approximation function;

[0051] In step S6, an acknowledgment timer is started. If an acknowledgment is received, the process proceeds to step S7. If no acknowledgment is received within the timeout period, the transmission is retransmitted using the exponential backoff algorithm. If the transmission fails after a preset number of retransmissions, a communication failure error is returned.

[0052] In step S1 or S8, when a personal AI agent module generates information that needs to be shared across AIs, it simultaneously calls the write interface of the shared context storage module to store the information along with a timestamp, version number, and the identifier of the collaboration group to which it belongs. Before executing subsequent tasks, any user's personal AI agent module first reads the latest global context of the corresponding collaboration group from the shared context storage module and merges it into its own local context. When the shared contexts of multiple personal AI agent modules conflict, the shared context storage module resolves the conflict based on the automatic merging rules of conflict-free copy data types and notifies all relevant personal AI agent modules of the resolved status.

[0053] Furthermore, the conflict resolution method based on shared context specifically includes the following steps:

[0054] Each context entry is accompanied by a version vector, which records the modification count of the entry by each participant;

[0055] When multiple personal AI agent modules update the same context entry simultaneously, the shared context storage module receives multiple write requests and determines whether there is a conflict based on the partial order relationship of the version vector.

[0056] If conflicts exist, they will be automatically merged according to the following rules:

[0057] For numeric fields, take the maximum or average value;

[0058] For text fields, if they originate from different individual AI agent modules, use structured merging;

[0059] For the task status field, the conflict handling strategy defined by the state machine is adopted;

[0060] A new version vector is generated based on the merged new context entries and broadcast to all online personal AI agent modules within the collaboration group.

[0061] The multi-agent collaborative communication system provided by this invention enables the personal AI agent module of the first user to directly establish a communication link with the personal AI agent module of the second user, realizing the direct transmission of work content, the sharing and continuation of context, and solving the interoperability problem across platforms and large models, while ensuring the security of permissions and data during the communication process. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the structure of a multi-agent cooperative communication system provided in an embodiment of the present invention.

[0063] Figure 2 This is a flowchart illustrating the multi-agent cooperative communication method provided in an embodiment of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0065] like Figure 1 As shown, in one embodiment of the present invention, a multi-agent cooperative communication system is provided, comprising:

[0066] Multiple personal AI agent modules are deployed on the terminal device or cloud of the corresponding user. They are used to receive and understand the natural language instructions of the user, interact with the user in natural language, maintain the user's task context and private data, and generate message content that needs to be passed to other users' personal AI agent modules.

[0067] A communication proxy gateway, deployed on the network side or user terminal side, is coupled with multiple personal AI agent modules to establish direct communication links between personal AI agent modules of different users, and to perform protocol adaptation and cross-model semantic conversion on messages transmitted in the communication links, thereby realizing addressing, connection establishment, message routing and protocol conversion between different personal AI agent modules.

[0068] A shared context storage module, coupled to the communication proxy gateway, is used to persistently store shared context information (such as task context, intermediate results, decision logs, and version information) generated by multiple personal AI agent modules during the collaboration process.

[0069] The distributed permission management module is coupled to the personal AI agent module, the communication agent gateway, and the shared context storage module, respectively. It is used to authorize and verify the communication requests and data access requests between personal AI agent modules based on preset permission policies, so as to manage the access control, data sharing scope, and operation permissions between user personal AI agent modules.

[0070] The communication proxy gateway is configured to route and send messages generated by the personal AI agent module of the first user (human) directly to the personal AI agent module of the second user (human) in the communication format of "first user → first user's personal AI agent module → second user's personal AI agent module → second user (Human-AI-AI-Human)," and vice versa.

[0071] In practical applications, the connection relationships and working principles of each module are as follows:

[0072] The first user's personal AI agent module connects to the first user's communication proxy gateway via an internal API, and the second user's personal AI agent module similarly connects to the second user's communication proxy gateway. An encrypted connection is established between the two communication proxy gateways via the internet or a private enterprise network. The system operates in a Human-AI-AI-Human format.

[0073] Human→AI (First User Side): The first user issues collaborative instructions to their personal AI agent module using natural language.

[0074] AI→AI (Cross-user side): The first user's personal AI agent module parses the instruction into a structured message, which is then passed through the communication agent gateway, the shared context storage module, and the distributed permission management module in sequence, and directly transmitted to the second user's personal AI agent module; this process does not rely on manual relay by the first or second user.

[0075] AI→Human (Second User Side): After receiving the message, the second user's personal AI agent module presents the information to the second user (human) and can wait for feedback. The reverse path is similar.

[0076] In a preferred embodiment of the present invention, the personal AI agent module further includes an intent classifier for automatically identifying whether the user-inputted instruction needs to trigger cross-AI communication, and automatically triggering the communication agent gateway when necessary.

[0077] In a preferred embodiment of the present invention, the communication proxy gateway specifically includes:

[0078] The unified identity resolution unit is used to resolve a user's identity identifier into a network-addressable agent address;

[0079] The session management unit is used to establish, maintain, and release logical communication sessions between multiple personal AI agent modules;

[0080] The protocol adaptation unit is used to convert the private interface protocols used by different personal AI agent modules into a unified standardized interaction protocol (such as an A2A protocol based on HTTP / JSON or an equivalent custom protocol).

[0081] A cross-model translation engine is used to perform bidirectional semantic and data format conversion between the first major model (such as model A) used by the first user's personal AI agent module and the second major model (such as model B) used by the second user's personal AI agent module, so as to enable direct communication between major models from different companies.

[0082] Preferably, the cross-model translation engine employs meta-knowledge distillation techniques between large models to pre-train a lightweight semantic alignment network. This semantic alignment network takes the message output of the source model as input and outputs a message representation that the target model can understand.

[0083] In a preferred embodiment of the present invention, the shared context storage module supports content-addressable read and write operations and provides an independent secure storage space for each collaborative group; the shared context storage module internally uses a vector database or key-value store and adopts a concurrent synchronization mechanism based on conflict-free replicated data type (CRDT) to maintain data consistency when multiple personal AI agent modules perform read and write operations on the shared context simultaneously.

[0084] Preferably, the shared context storage module also supports snapshot and rollback functions, allowing any individual AI agent module to restore the collaborative context to a historically specified state (at any time) to correct inconsistencies caused by information errors.

[0085] In a preferred embodiment of the present invention, the distributed permission management module specifically includes:

[0086] The identity authentication submodule is used to verify the legitimate identity of the individual AI agent module that initiates the communication request based on digital certificates or decentralized identity identifiers.

[0087] The policy decision submodule is used to store and execute the preset permission policies for each user. The permission policies include at least one of the following: allowed peer users to communicate, allowed data access range, and allowed operation types (e.g., which other users' personal AI agent modules are allowed to read which task contexts of mine, which types of instructions are allowed to be executed, etc.).

[0088] The operation audit submodule is used to record all communication and data access behaviors across personal AI agent modules for post-event auditing and anomaly detection.

[0089] The message sending and receiving and acknowledgment unit is integrated into the communication proxy gateway. It is used to support three communication modes: point-to-point, multicast, and broadcast, using asynchronous message queues or reliable UDP protocols, and provides message delivery acknowledgment, retransmission, and timeout handling mechanisms.

[0090] In terms of technical implementation, when the personal AI agent module of the first user needs to send work content to the personal AI agent module of the second user, the distributed permission management module first verifies the permission. After successful verification, the session management unit establishes a temporary session, and the protocol adaptation unit converts the private format message of the first user's personal AI agent module into a standard format. If necessary, the cross-model translation engine performs semantic conversion between large models. The converted message is sent to the communication proxy gateway of the second user via the message sending and receiving and confirmation unit. The communication proxy gateway performs a reverse conversion and then delivers it to the personal AI agent module of the second user.

[0091] During communication, the shared context generated by the individual AI agent modules of both parties is automatically or on-demand written to the shared context storage module. Subsequently, either party can retrieve the latest collaboration status from this shared context storage module, thus achieving continuous context synchronization. The entire communication link can use TLS 1.3 or a higher encryption protocol. All shared data is encrypted at the field level during storage, and the encryption keys are distributed according to policies by the distributed permission management module.

[0092] like Figure 2As shown, in another embodiment of the present invention, a multi-agent cooperative communication method is also provided, which is implemented through the above-described multi-agent cooperative communication system, and specifically includes the following steps:

[0093] S1 (Human→AI): The first user's personal AI agent module receives and parses the first user's natural language instructions, and parses out the work content and task objectives that need to be collaborated with the second user's personal AI agent module.

[0094] In practical applications, the first user issues collaborative instructions to their personal AI agent module using natural language (e.g., "Please ask Li Si why the parameter of that function is negative?"). The personal AI agent module parses the instructions and extracts the task content, intent type, and target user identifier.

[0095] S2 (AI Initiates Connection): The first user's personal AI agent module calls the communication agent gateway to initiate a request to establish a communication connection with the second user's personal AI; the request carries the identity identifier of the first user, the identity identifier of the second user, and the type of communication intent requested.

[0096] S3 (Permission Verification): The distributed permission management module verifies whether the request complies with the sending permission of the first user and the receiving policy preset by the second user. If the verification fails, the system returns insufficient permission information to the first user's personal AI agent module and terminates the process. If the verification is successful, the system continues to execute step S4.

[0097] For example, the distributed permission management module verifies the request: it verifies whether the first user has the right to send such a message to the second user; and it verifies whether the second user is allowed to receive such a message. If the verification fails, it returns "insufficient permissions" and terminates the process; if it succeeds, it generates a temporary session token.

[0098] S4 (Identity Resolution and Connection Establishment): The unified identity resolution unit in the communication proxy gateway resolves the identity identifier of the second user into the network address of the communication proxy gateway of the second user, and initiates an encrypted transport layer security connection (TLS connection) through the session management unit.

[0099] After the connection is established (S5, protocol conversion and cross-model translation), the protocol adaptation unit in the first user's communication proxy gateway converts the internal message format of the first user's personal AI agent module into a standardized interactive message (such as JSON-RPC). If the first major model used by the first user's personal AI agent module is different from the second major model used by the second user's personal AI agent module, then the cross-model translation engine performs a bidirectional mapping conversion of semantics and format: the output of the source major model is converted into an intermediate semantic graph, and then instructions / data that the target major model can understand are generated.

[0100] S6 (AI→AI Sending and Confirmation): The message sending and receiving and confirmation unit sends the converted standardized interactive message to the second user's communication proxy gateway and starts the confirmation timer. If confirmation is received, proceed to step S7. If no confirmation is received after timeout, retransmit according to the exponential backoff algorithm (up to 3 times). If retransmission still fails after a preset number of times (e.g., 2 times), return a communication failure error.

[0101] S7 (Receiving and Inverse Conversion): After receiving the standardized interaction message, the second user's communication proxy gateway performs the inverse conversion (including model translation inverse mapping and protocol restoration) which is the opposite of step S5, and delivers the final message to the second user's personal AI agent module.

[0102] After receiving the final message, the S8 (AI→Human) personal AI agent module for the second user automatically converts it into a form that the second user can understand (such as natural language summaries, structured prompts, or voice broadcasts) and presents it to the second user (human) according to the notification rules preset by the second user. If the message contains content that requires the second user to make a decision or provide feedback, the personal AI agent module for the second user waits for the second user to input a response.

[0103] Step S9 (optional reverse process): If the second user responds, repeat the Human-AI-AI-Human process described above (second user → second user's personal AI agent module → first user's personal AI agent module → first user) until collaboration is complete. Either party may initiate a session termination request, and the communication proxy gateway will release session resources.

[0104] In a preferred embodiment of the present invention, step S3 specifically includes:

[0105] Sub-step P1: The first user's distributed permission management module retrieves the first user's private key from local secure storage and signs the request content;

[0106] Sub-step P2: Send the signed request to the second user's distributed permission management module;

[0107] Sub-step P3: The second user's distributed permission management module verifies the signature validity, and then queries the second user's policy database to determine whether the triple "first user → second user → intent type" is allowed;

[0108] Sub-step P4: If allowed, generate a temporary session token (validity period is configurable, default is 30 minutes) and return the token to the first user's communication proxy gateway; if not allowed, return a rejection message and reason.

[0109] In a preferred embodiment of the present invention, in order to solve the technical problem that large cross-company models cannot communicate directly, the cross-model translation engine performs the following sub-steps:

[0110] Sub-step T1: Obtain a series of message vectors or text fragments output by the source large model used by the source personal AI agent module and the model identifier of the target large model used by the target personal AI agent module;

[0111] Sub-step T2: Load the semantic alignment mapping table corresponding to the target large model from the pre-stored knowledge base. This mapping table can be obtained through offline training. The training data is the parallel corpus generated by the source large model and the target large model under the same input.

[0112] Sub-step T3: Based on the semantic alignment mapping table, the source message is segmented, intent extracted, and structured to generate an intermediate language representation (an abstract semantic graph) that does not depend on a specific large model.

[0113] Sub-step T4: Based on the intermediate language representation, and combined with the prompt template style and parameter format of the target large model, generate instruction strings or structured data that the target large model can directly understand and execute;

[0114] Sub-step T5: If the target large model does not support certain special functions of the source large model (such as function calls), it will automatically be downgraded to text description or simulation approximation functions, and a warning mark will be added.

[0115] In a preferred embodiment of the present invention, the above method further includes (throughout the entire process) the following in order to achieve context sharing and continuity:

[0116] Step C1: In step S1 or S8, when the personal AI agent module generates information that needs to be shared across AIs, it simultaneously calls the write interface of the shared context storage module to store the information along with a timestamp, version number, and the identifier of the collaboration group to which it belongs, and writes it to the shared context storage module.

[0117] Step C2: Before executing subsequent tasks, each user's personal AI agent module first reads the latest global context of the corresponding collaboration group from the shared context storage module and merges it into its own local context.

[0118] Step C3: When a conflict occurs in the shared context of multiple (e.g., two) personal AI agent modules, the shared context storage module resolves the conflict based on the automatic merging rules of the conflict-free copy data type (CRDT) and notifies all relevant personal AI agent modules of the resolved status; if automatic merging is not possible, human arbitration is requested.

[0119] Specifically, to address the information asymmetry problem arising from the independent operation of multiple personal AI agent modules, the conflict resolution method based on shared context includes the following steps:

[0120] Each context entry is accompanied by a version vector, which records the modification count of the entry by each participant;

[0121] When multiple personal AI agent modules update the same context entry simultaneously, the shared context storage module receives multiple write requests and determines whether there is a conflict based on the partial order relationship of the version vector.

[0122] If conflicts exist, they will be automatically merged according to the following rules:

[0123] For numeric fields, take the maximum or average value (preset by the collaboration group).

[0124] For text-based fields, if they originate from different personal AI agent modules, a structured merging process is used (the two texts are appended to the field as opinions from different sources and marked as "awaiting human arbitration").

[0125] For the task status field, a conflict handling strategy defined by the state machine is adopted (for example, if one AI is marked "completed" and another is marked "in progress", then "completed" is used and the overwrite log is recorded).

[0126] A new version vector is generated based on the merged new context entries and broadcast to all online personal AI agent modules within the collaboration group.

[0127] Preferably, when automatic merging cannot be determined, the shared context storage module generates an arbitration request and sends it to the personal AI agent module of the user with the highest authority in the collaboration group (such as the project leader). The user makes the final decision through natural language, and the system learns this decision for subsequent similar conflicts.

[0128] Compared with the prior art, the technical solution provided by the embodiments of the present invention has the following significant advantages:

[0129] 1. Enables direct communication between personal AIs, eliminating manual intermediaries: This invention, through the establishment of a communication proxy gateway and unified identity resolution, session management, and protocol adaptation units, allows a first user's personal AI to directly send messages to a second user's personal AI, fully supporting the Human-AI-AI-Human communication format. Users no longer need to manually copy AI-generated content to third-party communication software, nor does the recipient need to re-enter their personal AI, thus completely eliminating human delays and operational burdens in the information relay process and significantly improving team collaboration efficiency.

[0130] 2. Avoiding errors and distortions in information transmission: In existing technologies, information undergoes multiple transformations—"AI → sender user → communication software → receiver user → receiver AI"—making it highly susceptible to semantic misinterpretation, data omissions, or formatting errors. This invention employs end-to-end AI-to-AI direct communication. Messages maintain their structure and integrity under the processing of a cross-model translation engine and protocol adaptation unit, and reliable delivery is ensured through message confirmation and retransmission mechanisms. This effectively avoids information errors caused by manual forwarding and ensures the accuracy of collaborative results.

[0131] 3. Achieving cross-AI shared context and eliminating information asymmetry: This embodiment of the invention sets up a shared context storage module, which can persistently store context information such as task history, intermediate results, and decision logs among multiple individual AIs, and supports concurrent synchronization and conflict resolution based on CRDT. Each user's individual AI can read and write the global context at any time, thereby maintaining a consistent understanding of the collaborative state. Compared with the situation in the prior art where AIs are isolated and act independently, this embodiment of the invention fundamentally solves the problem of repeated correction and communication coordination caused by missing context.

[0132] 4. Breaking down communication barriers between large models from different companies: The cross-model translation engine in this embodiment of the invention employs semantic alignment mapping technology, which can convert the output of the source large model into instructions or data that the target large model can directly understand, without requiring a unified underlying model or all users to use the same AI platform. This enables individual AIs from different vendors (such as large models developed by different companies) to work collaboratively without obstacles, breaking through the technical bottleneck of existing closed ecosystems and the inability to communicate across models, and has broad compatibility and practical value.

[0133] 5. Provides fine-grained distributed access control to ensure information security: This invention incorporates identity authentication, policy decision-making, and operation auditing modules. While allowing personal AIs to communicate with each other, users can precisely control which other users' AIs can access which contexts and perform which operations. All cross-AI communication and data access behaviors are recorded, and communication links are encrypted using TLS, while shared data is encrypted at the field level. Compared to the risks of permission abuse and data leakage caused by simply "sharing personal AI" in existing technologies, this invention significantly enhances enterprise-level security and compliance.

[0134] 6. Support for automatic conflict resolution and recovery of context: The conflict resolution method based on version vectors in this embodiment of the invention can automatically merge or arbitrate when multiple AIs modify the same context simultaneously, and restore to any historical state through snapshot and rollback functions. This not only solves the problem of information asymmetry, but also avoids data chaos caused by concurrent operations, enabling stable and reliable operation of multi-agent collaboration.

[0135] 7. Reduce team communication costs and truly achieve AI-driven team collaboration: Combining the above advantages, this invention allows each user to interact naturally with their own personal AI, indirectly and efficiently collaborating with the personal AIs of all other team members. All work content, context, and decision records flow automatically between AIs. Teams no longer need to hold frequent synchronous meetings, manually organize results from multiple parties, or rely on third-party communication software for information alignment, thereby significantly reducing communication costs and truly upgrading AI from a personal tool to an equal participant in team collaboration.

[0136] The following specific embodiment illustrates in detail how the present invention is implemented in the daily work of a company and its beneficial effects. This embodiment is used to achieve direct communication and collaboration between different user-specific AI agent modules.

[0137] Example 1: This example describes a scenario where two employees in the marketing department of a company use their respective personal AI agent modules to collaborate on cross-model tasks.

[0138] I. Scenario Setting: Zhang San, a marketing specialist at a company, uses a personal AI agent module based on a large model A (e.g., a commercial large model), while Li Si, a finance specialist, uses a personal AI agent module based on a large model B (another vendor's large model). The two need to collaborate on a marketing campaign budget approval: Zhang San needs to send key data (total campaign budget, itemized expenses) from a budget application form for a "618 online promotion" to Li Si and request Li Si to confirm its compliance with financial regulations. Under existing technology, Zhang San needs to manually copy the AI-generated budget form to WeChat and send it to Li Si, who then manually inputs it into their AI for calculation. This process is prone to data errors, omissions, or formatting issues, and the repeated communication between the two is time-consuming. This invention, based on the aforementioned multi-agent collaborative communication system, enables direct communication between Zhang San's personal AI and Li Si's personal AI, eliminating the need for manual intermediaries.

[0139] II. System Configuration: According to the technical solution provided in the embodiments of the present invention, the core components required for the multi-agent cooperative communication system are pre-deployed:

[0140] Personal AI Agent Module: Zhang San and Li Si each installed client software on their office computers, bound their own large model accounts, and generated identity identifiers based on digital certificates.

[0141] Communication Proxy Gateway: A gateway server is deployed on the company's intranet, configured with a unified identity resolution unit (interfacing with the company's AD domain), a session management unit (using Redis to store sessions), a protocol adaptation unit (supporting bidirectional conversion between private protocols and standardized protocols of large model A and large model B), and a cross-model translation engine (a semantic alignment model pre-trained using the company's historical communication corpus).

[0142] Shared context storage module: Deploy a database cluster to store the shared context of the collaboration group (such as budget approval status, modification records, etc.).

[0143] Distributed access control module: Zhang San and Li Si each set access policies. Zhang San's policy allows Li Si's AI to read the "budget application" field in its shared context; Li Si's policy allows Zhang San's AI to send a message with the intent of "budget approval" to him, requiring the message to include a complete expense breakdown.

[0144] The above modules together constitute a complete multi-agent collaborative communication system, supporting a closed-loop process from human command input, inter-agent communication to execution result feedback.

[0145] III. Execution Process:

[0146] Step S1 (Human→AI): Zhang San opens his personal AI client on his computer and speaks in natural language: "Xiao A (his AI nickname), please send the key data from the '618 Promotion Budget Sheet' I just gave you—total budget of 50,000 yuan, including 30,000 yuan for advertising, 15,000 yuan for gifts, and 5,000 yuan for logistics—to the AI ​​of finance manager Li Si, and ask him if it complies with the company's budget regulations." The voice input is converted into text, and the intent classifier in the personal AI agent module recognizes the instruction as "cross-user collaboration," automatically extracting the target user "Li Si," the intent "budget approval," and the structured content (total budget of 50,000 yuan, advertising fee of 30,000 yuan, gift fee of 15,000 yuan, and logistics fee of 5,000 yuan).

[0147] Step S2 (AI initiates connection request): Zhang San's personal AI calls the local communication proxy gateway interface to initiate a connection request. The parameters include the sender identifier, receiver identifier, intent type, and payload data. After receiving the request, the gateway forwards it to the distributed permission management module.

[0148] Step S3 (Permission Verification): The distributed permission management module verifies Zhang San's sending permission and Li Si's receiving policy (Li Si's policy explicitly allows budget approval intent from Zhang San). After successful verification, a session token with a validity period of 30 minutes is generated and returned to Zhang San's gateway.

[0149] Step S4 (Identity Resolution and Connection Establishment): The unified identity resolution unit resolves Li Si's identity identifier into the IP address and public key fingerprint of his gateway. Zhang San's gateway establishes a two-way authenticated encrypted connection with Li Si's gateway using an encryption protocol.

[0150] Step S5 (Protocol Adaptation and Cross-Model Translation): The protocol adaptation unit in Zhang San's gateway converts the private message format within Zhang San's personal AI into a standardized interaction message, including the protocol version, sender and receiver identifiers, intent type, and payload data. Since the target Li Si uses the large model B, the cross-model translation engine is invoked: converting the standardized message into an instruction format that the large model B can directly understand—translating the payload data into a natural language prompt: "Please approve the following budget: total budget 50,000 yuan, advertising fee 30,000 yuan, gift fee 15,000 yuan, logistics fee 5,000 yuan. Does this comply with company regulations?" A structured backup is also retained.

[0151] Step S6 (AI→AI Sending and Acknowledgment): The message sending and receiving unit sends the translated message to Li Si's gateway through the message queue and starts the acknowledgment timer. Upon receiving the message, Li Si's gateway immediately returns an acknowledgment frame, and Zhang San's gateway receives the acknowledgment.

[0152] Step S7 (Receiving and Reverse Conversion): After receiving the message, Li Si's gateway performs reverse conversion: the prompt in the large model B format is restored to a standard message and delivered to Li Si's personal AI agent module.

[0153] Step S8 (AI → Human): After receiving the message, Li Si's personal AI (nickname "Xiao B"), according to Li Si's preset notification rules (pop-up reminders during work hours), pops up a notification in the lower right corner of the computer screen: "Zhang San's AI has sent a budget approval request: total budget of 50,000 yuan. Please confirm whether it complies with regulations." The expense details are also displayed in a table format. After reading it, Li Si verbally replies to his AI: "Tell Zhang San that the advertising expense of 30,000 yuan exceeds the department's quarterly limit (limit of 25,000 yuan). It is recommended to reduce it to 25,000 yuan. Other items are acceptable."

[0154] Step S9 (Reverse Process): Li Si's AI transmits the above response back through the same multi-agent collaborative communication system path (Li Si → Li Si AI → Gateway → Zhang San Gateway → Zhang San AI → Zhang San). Upon receiving this, Zhang San's AI reads aloud: "Li Si's feedback: The advertising expense of 30,000 yuan exceeds the department's quarterly limit of 25,000 yuan. It is recommended to reduce it to 25,000 yuan." Zhang San modifies the budget based on this feedback. The entire round-trip communication takes approximately 1.8 seconds (excluding human thinking and input time).

[0155] Step S10 (Context Writing and Sharing): During message transmission, the AIs of both parties automatically write key decisions into the shared context storage module. Specifically, when Zhang San's AI sends a message, it writes the relevant information of the budget application (budget identifier, total budget, status pending approval) into the collaboration group's context; after Li Si's AI replies, it updates the entry with the adjusted budget data, the status "adjusted," the recommender, and the reason for the adjustment. Subsequently, when the marketing manager needs to check the approval status of the budget, their AI can directly read the latest result from the shared context without needing to ask Zhang San or Li Si again.

[0156] IV. Beneficial Effects of This Embodiment: As can be clearly seen from the above-mentioned budget approval case in the company's daily work, the multi-agent collaborative communication system constructed in this embodiment of the invention achieves the following technical effects:

[0157] Direct communication: Zhang San and Li Si's personal AI can directly transmit budget data without human intermediaries (such as WeChat for Business or email). The operation steps have been simplified from the traditional "copy the table → paste → send → receive → input into AI" to a single voice command, significantly improving efficiency.

[0158] Cross-model compatibility: Semantic and format differences between large model A and large model B are transparently handled by the cross-model translation engine, and both AIs can accurately understand each other's messages, avoiding data misalignment or semantic misunderstanding caused by different models.

[0159] Context sharing: Key parameters and modification suggestions for budget approval are automatically written to shared storage, which can be directly accessed by subsequent collaborators (such as marketing managers), avoiding duplicate communication and information asymmetry.

[0160] Access control security: The distributed access control module ensures that only authorized intents and authorized users can initiate communication, and the communication content is encrypted throughout the process, which meets the enterprise's financial data security requirements.

[0161] Execution loop: The entire process forms a complete multi-agent collaborative communication system loop of "human command → agent communication → agent execution → result feedback to humans", which makes AI a truly active participant in team collaboration, rather than a passive personal tool.

[0162] It should be understood that although the steps in the flowcharts of the embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0163] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods.

[0164] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A multi-agent cooperative communication system, characterized in that, include: Multiple personal AI agent modules are deployed on the corresponding user's terminal device or cloud. These modules are used to interact with the user in natural language and maintain the user's task context and private data. A communication proxy gateway, coupled to multiple personal AI agent modules, is used to establish direct communication links between personal AI agent modules of different users, and to perform protocol adaptation and cross-model semantic conversion on messages transmitted in the communication links. A shared context storage module, coupled to the communication proxy gateway, is used to persistently store shared context information generated by multiple personal AI agent modules during collaboration. The distributed permission management module is coupled to the personal AI agent module, the communication agent gateway, and the shared context storage module, respectively, and is used to authorize and verify communication requests and data access requests between personal AI agent modules based on a preset permission policy. The communication proxy gateway is configured to route and send messages generated by the personal AI agent module of the first user directly to the personal AI agent module of the second user in the communication format of "first user → first user's personal AI agent module → second user's personal AI agent module → second user".

2. The multi-agent cooperative communication system according to claim 1, characterized in that, The communication proxy gateway specifically includes: The unified identity resolution unit is used to resolve a user's identity identifier into a network-addressable agent address; The session management unit is used to establish, maintain, and release logical communication sessions between multiple personal AI agent modules; The protocol adaptation unit is used to convert the private interface protocols used by different personal AI agent modules into a unified standardized interaction protocol. A cross-model translation engine is used for bidirectional semantic and data format conversion between the first major model used by the first user's personal AI agent module and the second major model used by the second user's personal AI agent module.

3. The multi-agent cooperative communication system according to claim 1, characterized in that, The shared context storage module adopts a concurrent synchronization mechanism based on conflict-free copy data type to maintain data consistency when multiple personal AI agent modules simultaneously perform read and write operations on the shared context.

4. The multi-agent cooperative communication system according to claim 3, characterized in that, The shared context storage module also supports snapshot and rollback functions, allowing any individual AI agent module to restore the collaborative context to a specified historical state.

5. The multi-agent cooperative communication system according to claim 3, characterized in that, The cross-model translation engine employs meta-knowledge distillation techniques between large models, pre-training a lightweight semantic alignment network. This semantic alignment network takes the message output of the source model as input and outputs a message representation that the target model can understand.

6. The multi-agent cooperative communication system according to claim 1, characterized in that, The distributed permission management module specifically includes: The identity authentication submodule is used to verify the legitimate identity of the individual AI agent module that initiated the communication request; The policy decision submodule is used to store and execute the preset permission policies for each user; the permission policies include at least one of the following: allowed peer users for communication, allowed data range for access, and allowed operation types; The operation audit submodule is used to record all communication and data access behaviors across personal AI agent modules; The message sending and receiving and acknowledgment unit is used to support three communication modes: point-to-point, multicast, and broadcast, using asynchronous message queues or reliable UDP protocols, and provides message delivery acknowledgment, retransmission, and timeout handling mechanisms.

7. The multi-agent cooperative communication system according to claim 1, characterized in that, The personal AI agent module also includes an intent classifier, which automatically identifies whether the user's input command needs to trigger cross-AI communication and automatically triggers the communication agent gateway when needed.

8. A multi-agent cooperative communication method, implemented using the multi-agent cooperative communication system according to any one of claims 1-7, characterized in that, Includes the following steps: S1. The first user's personal AI agent module receives and parses the first user's natural language instructions, and parses out the work content and task objectives that need to be collaborated with the second user's personal AI agent module. S2. The first user's personal AI agent module calls the communication agent gateway to initiate a request to establish a communication connection with the second user's personal AI; the request carries the identity identifier of the first user, the identity identifier of the second user, and the type of communication intent requested. S3. Verify whether the request complies with the sending permissions of the first user and the receiving strategy preset by the second user through the distributed permission management module. If the verification fails, return the permission insufficient information to the personal AI agent module of the first user and terminate the process. If the verification is successful, proceed to step S4; S4. The identity of the second user is resolved to the network address of the second user's communication proxy gateway through the communication proxy gateway, and an encrypted transport layer secure connection is initiated. S5. After the connection is established, the first user's communication proxy gateway will convert the internal message format of the first user's personal AI agent module into a standardized interactive message. S6. Send the converted standardized interactive message to the second user's communication proxy gateway; S7. After receiving the standardized interaction message, the second user's communication proxy gateway performs the reverse transformation of step S5 and delivers the final message to the second user's personal AI agent module. S8. After receiving the final message, the second user's personal AI agent module automatically converts it into a form that the second user can understand and presents it to the second user according to the notification rules preset by the second user.

9. The multi-agent cooperative communication method according to claim 8, characterized in that, Step S3 specifically includes: Sub-step P1: The first user's distributed permission management module retrieves the first user's private key from local secure storage and signs the request content; Sub-step P2: Send the signed request to the second user's distributed permission management module; Sub-step P3: The distributed permission management module of the second user verifies the validity of the signature, and then queries the policy database of the second user to determine whether the triple "first user → second user → intent type" is allowed; Sub-step P4: If allowed, generate a temporary session token and return the token to the first user's communication proxy gateway; if not allowed, return a rejection message and the reason. In step S5, if the first major model used by the first user's personal AI agent module is different from the second major model used by the second user's personal AI agent module, then a bidirectional mapping conversion of semantics and format is performed through a cross-model translation engine. The cross-model translation engine performs the following sub-steps: Sub-step T1: Obtain a series of message vectors or text fragments output by the source large model used by the source personal AI agent module and the model identifier of the target large model used by the target personal AI agent module; Sub-step T2: Load the semantic alignment mapping table corresponding to the target large model from the pre-stored knowledge base; Sub-step T3: Based on the semantic alignment mapping table, perform word segmentation, intent extraction, and structured representation on the source message to generate an intermediate language representation that does not depend on a specific large model; Sub-step T4: Based on the intermediate language representation, and combined with the prompt template style and parameter format of the target large model, generate instruction strings or structured data that the target large model can directly understand and execute; Sub-step T5: If the target large model does not support certain special functions of the source large model, it will automatically be downgraded to text description or simulation approximation function; In step S6, an acknowledgment timer is started. If an acknowledgment is received, the process proceeds to step S7. If no acknowledgment is received within the timeout period, the transmission is retransmitted using the exponential backoff algorithm. If the transmission fails after a preset number of retransmissions, a communication failure error is returned. In step S1 or S8, when a personal AI agent module generates information that needs to be shared across AIs, it simultaneously calls the write interface of the shared context storage module to store the information along with a timestamp, version number, and the identifier of the collaboration group to which it belongs. Before executing subsequent tasks, any user's personal AI agent module first reads the latest global context of the corresponding collaboration group from the shared context storage module and merges it into its own local context. When the shared contexts of multiple personal AI agent modules conflict, the shared context storage module resolves the conflict based on the automatic merging rules of conflict-free copy data types and notifies all relevant personal AI agent modules of the resolved status.

10. The multi-agent cooperative communication method according to claim 9, characterized in that, The conflict resolution method using shared context specifically includes the following steps: Each context entry is accompanied by a version vector, which records the modification count of the entry by each participant; When multiple personal AI agent modules update the same context entry simultaneously, the shared context storage module receives multiple write requests and determines whether there is a conflict based on the partial order relationship of the version vector. If conflicts exist, they will be automatically merged according to the following rules: For numeric fields, take the maximum or average value; For text fields, if they originate from different individual AI agent modules, use structured merging; For the task status field, the conflict handling strategy defined by the state machine is adopted; A new version vector is generated based on the merged new context entries and broadcast to all online personal AI agent modules within the collaboration group.

Images