Multi-agent cooperation system

Abstract

The invention provides a multi-agent cooperation system, which comprises a multi-agent cooperation engine module, a task decomposition and distribution module and a tool calling normal form module, and is characterized in that the task decomposition and distribution module is used for decomposing tasks and distributing decomposed sub-tasks to agents; the multi-agent cooperation engine module is used for managing cooperation and communication between agents; and the tool calling module enables the intelligent agent to dynamically call different tools. According to the multi-agent collaboration system, the collaboration efficiency among the agents, the flexibility of task decomposition, the adaptivity of tool calling and the intelligent level of user interaction are remarkably improved, and remarkable technical advantages and effects are shown in complex task processing and dynamic environment processing.

Description

Technical Field

[0001] The present invention relates to an intelligent agent system, and in particular to a multi-agent collaboration system. Background Art

[0002] In the current field of multi-agent systems and AI technology, there are several aspects:

[0003] 1. Multi-agent system based on predefined rules

[0004] Traditional multi-agent systems usually rely on pre-set communication protocols and task allocation rules, and the collaboration between agents is carried out through symbolic languages ​​or pre-defined message formats. Common multi-agent collaboration methods include distributed blackboard systems or subscription / publishing messaging mechanisms.

[0005] In these systems, functions such as task allocation and tool calling are often based on fixed algorithms or rules, lacking flexible adaptive mechanisms. The interaction between agents mainly relies on symbolic languages ​​or predefined instruction sets, and cannot carry out complex natural language communication and dynamic decision-making.

[0006] 2 Multi-agent collaboration system based on reinforcement learning

[0007] In recent years, reinforcement learning technology has been gradually applied to multi-agent systems, where agents learn strategies by interacting with the environment to achieve collaborative or competitive goals. Each agent adjusts its behavior based on a reward mechanism, but this collaborative approach relies on a large amount of training data and lacks direct natural language communication capabilities between agents.

[0008] Reinforcement learning agents are able to optimize decisions through environmental feedback, but because there are no standardized processes for operations such as task decomposition and tool calling, the system's adaptability and task execution flexibility are still limited and cannot be dynamically adjusted according to complex task requirements.

[0009] 3. Robotic Process Automation (RPA) System

[0010] RPA systems are usually used to automate specific tasks. Through predefined rules and workflows, the system can complete efficient task allocation and tool calling. In RPA systems, task allocation and tool calling are based on clear workflow design. The intelligent agent lacks adaptive capabilities and cannot be flexibly adjusted according to task requirements.

[0011] RPA can complete tasks in fixed processes, but when faced with complex and changing tasks, it cannot achieve large-scale multi-agent collaboration, and lacks the ability to process natural language and decompose large-scale tasks.

[0012] Current multi-agent systems are limited in the following ways:

[0013] (1) Insufficient collaboration between agents: Existing multi-agent systems often lack efficient collaborative capabilities in task execution. Especially when dealing with complex tasks, there are bottlenecks in communication and information sharing between agents, resulting in a decrease in the efficiency of the overall system.

[0014] (2) Complexity of task decomposition and allocation: Multi-agent systems need to deal with complex task decomposition and allocation problems, but traditional methods mostly rely on pre-set rules or simple task segmentation and are unable to adjust task decomposition and optimize collaboration in real time according to dynamic environments.

[0015] (3) Insufficient flexibility in tool calling: Existing systems often use a static calling mechanism when calling tools, which makes it difficult to dynamically allocate and call tools based on task requirements or the state of the agent, limiting the system’s adaptability and flexibility.

[0016] (4) Engineering challenges of cross-platform integration: Multi-agent collaborative systems face engineering challenges in cross-platform integration and deployment, especially the incompatibility of front-end and back-end technology stacks and limited system scalability, which lead to high development, deployment and maintenance costs.

[0017] (5) Lack of natural language interaction capabilities: Existing intelligent agent systems have limited performance in natural language interaction, making it difficult to support complex collaborative communication and task negotiation among multiple agents through natural language, which reduces the collaborative capabilities of intelligent agents in complex task environments.

[0018] (6) Insufficient large-scale data processing and real-time response capabilities: Faced with large-scale data streams in complex scenarios, traditional multi-agent systems often perform poorly in real-time response and data processing, especially when high-dimensional information is involved, which is prone to delays and bottlenecks. Summary of the invention

[0019] The present invention provides a multi-agent collaboration system, which aims to solve a series of problems in existing multi-agent systems in collaboration, task allocation and tool calling. For example, traditional multi-agent systems rely on predefined communication protocols and rules for collaboration, and the collaboration efficiency between agents is often low, especially when dealing with tasks in complex or dynamic environments. There are bottlenecks in information sharing and decision-making collaboration.

[0020] The technical solution is as follows:

[0021] A multi-agent collaboration system includes a multi-agent collaboration engine module, a task decomposition and assignment module and a tool calling paradigm module. The task decomposition and assignment module is used to decompose tasks and assign the decomposed subtasks to agents; the multi-agent collaboration engine module is used to manage collaboration and communication between agents; and the tool calling module enables agents to dynamically call different tools.

[0022] The multi-agent collaboration engine module is used to promote natural language communication and information sharing between agents through a large language model, so that each agent can communicate with other agents through the large language model, so that each agent can dynamically share information according to task requirements, and can resolve task conflicts between multiple agents through a negotiation mechanism.

[0023] The multi-agent collaboration engine module can use symbolic language and machine learning driven non-language models to replace large language models.

[0024] The multi-agent collaboration engine module can implement collaboration between multiple agents by using distributed signal-based collaboration, artificial bee colonies or group intelligence collaboration.

[0025] The task decomposition and allocation module understands the task context through natural language, is responsible for decomposing complex tasks into manageable subtasks, and allocates tasks according to the capabilities and current status of the intelligent agent, so as to realize task analysis, intelligent decomposition and dynamic allocation; the task decomposition and allocation module can be replaced by a rule engine and expert system, or a machine learning algorithm.

[0026] The tool calling paradigm module can analyze task requirements through a large language model and determine the most suitable tool to complete the task, so as to achieve tool adaptation and calling, dynamic tool selection, and tool result analysis. It can be replaced by API-based static calling or decision tree or conditional logic.

[0027] The present invention also includes an interaction paradigm module, a task execution and monitoring module, an engineering facility and integration module, and a data storage and processing module; the interaction paradigm module is responsible for managing the interaction between the agent and the user and the internal communication between the agents through a unified language interaction interface; the task execution and monitoring module is used to manage the tracking and feedback of the agent's execution of tasks; the engineering facility and integration module is used to realize the front-end and back-end integration and deployment of the multi-agent collaboration system; the data storage and processing module is used to manage large-scale data processing tasks in the multi-agent collaboration system.

[0028] The interaction paradigm module includes the following interaction modes:

[0029] (1) User-agent interaction: Users input tasks through natural language, and the system automatically recognizes the task content and understands and processes it through the language model;

[0030] (2) Agent-agent interaction: Agents exchange information through a unified language interface to ensure the collaborative work of multi-agent systems in complex tasks;

[0031] (3) Interaction optimization: Based on user feedback and task execution results, the system continuously optimizes the interaction method.

[0032] The task execution and monitoring module includes the following processing contents:

[0033] (1) Task progress monitoring: The system continuously monitors the agent's task execution progress and displays the status, progress, and expected completion time of each subtask through a graphical interface;

[0034] (2) Exception handling: When an exception or delay occurs during task execution, the system can dynamically adjust the task execution order or reallocate tasks based on the importance and current status of the tasks;

[0035] (3) Feedback mechanism: Users and agents provide real-time feedback on task execution, and the system adjusts its execution strategy based on the feedback.

[0036] The data storage and processing module includes the following processing contents:

[0037] (1) Distributed storage: Use distributed databases or cloud storage services to support efficient storage and access of large-scale data;

[0038] (2) Real-time data processing: supports real-time processing and analysis of data generated during task execution;

[0039] (3) Historical data backtracking: Users and systems can backtrack historical task data to analyze and optimize the execution of future tasks.

[0040] The multi-agent collaboration system, through the organic combination of a multi-agent collaboration engine, a task decomposition and allocation module, a tool calling paradigm module, an interactive paradigm module and supporting engineering facilities, has significant technical advantages over the prior art, and significantly improves the collaboration efficiency, adaptability and cross-platform engineering application capabilities of the multi-agent system. The present invention implements a multi-agent collaboration engine based on a large language model, so that the agents can communicate and collaborate efficiently through natural language; through the task decomposition paradigm, the decomposition and allocation of tasks are dynamically adjusted to improve the adaptability and efficiency of the agent system in a complex environment; the tool calling paradigm is adopted to enable the system to flexibly call the required tools according to the task requirements and the state of the agent, and enhance the adaptive ability; supporting engineering facilities are provided, combined with the Javascript front end and the PythonDjango back end, to solve the cross-platform integration problem of the multi-agent collaboration system in practical applications.

[0041] The following analyzes its advantages and effects from multiple perspectives:

[0042] 1. Improve the collaborative efficiency between agents

[0043] Existing technical problems: Traditional multi-agent systems rely on predefined communication protocols and rules for collaboration. The collaboration efficiency between agents is often low, especially when dealing with tasks in complex or dynamic environments. There are bottlenecks in information sharing and decision-making collaboration.

[0044] This invention introduces a large language model to allow agents to exchange information and coordinate tasks through natural language, greatly simplifying the complexity of communication between agents. The language model can understand and infer the implicit information in the task and reach a consensus with other agents through natural language, reducing the limitations of traditional symbolic communication.

[0045] Through natural language interaction, communication between intelligent agents is more flexible and adaptive, which improves the collaborative efficiency of the system in complex tasks. Especially when intelligent agents need to make joint decisions or dynamically adjust tasks, it can significantly reduce response time and improve overall task completion efficiency.

[0046] 2. Dynamic task decomposition and allocation optimization

[0047] Existing technical problems: Traditional task decomposition and allocation methods usually rely on static rules or algorithms, which make it difficult to optimize tasks according to real-time conditions. This makes multi-agent systems often inefficient when handling complex tasks and unable to fully utilize the capabilities and resources of each agent.

[0048] The present invention adopts a task decomposition and allocation mechanism based on a large language model, which can analyze task complexity and agent status in real time, dynamically generate subtasks through the task decomposition paradigm, and reasonably allocate them to the most suitable agent. Compared with static rules, the reasoning ability of the language model enables the system to better adapt to complex task scenarios and environmental changes.

[0049] The decomposition and allocation of tasks are more flexible, and can be adjusted in real time according to the state of the intelligent agent and task requirements in a dynamic environment, which significantly improves the task execution efficiency and resource utilization, and reduces resource waste and delays caused by improper task allocation.

[0050] 3. Flexibility and adaptability of tool calls

[0051] Existing technical problems: The tool calling mechanism of traditional multi-agent systems is usually based on a preset calling process or fixed API binding, which cannot be flexibly called according to task requirements, resulting in poor system adaptability and difficulty in meeting the diverse tool requirements in complex tasks.

[0052] This invention introduces a tool calling paradigm, whereby the agent can flexibly call appropriate tools based on the understanding of the task requirements by the large language model. This tool calling paradigm has the ability to dynamically adjust, and can change the called tools or services at any time according to the changes in the task, without relying on a fixed calling process.

[0053] The system can flexibly adapt to a variety of tools during task execution, and can switch or expand the calling range of tools when needed, which improves the flexibility and adaptability of tool calling and ensures the continuity and stability of task execution, especially when processing heterogeneous tasks.

[0054] 4. Natural language interaction and user experience optimization

[0055] Existing technical problems: User interaction in existing multi-agent systems is mainly achieved through graphical interfaces or predefined instructions. The interaction between users and agents is often limited to fixed input and output formats, which reduces the flexibility and intuitiveness of the user experience.

[0056] The present invention uses the natural language interaction paradigm, allowing users to interact with the intelligent agent in a natural language manner without having to master complex technical instructions or system logic. This interaction method not only improves the usability of the system, but also automatically recognizes user intentions through language models and responds quickly, improving the level of intelligent interaction.

[0057] The user experience is more natural and intuitive. The intelligent agent can dynamically adjust the task execution plan according to the user's language input, reducing the user's learning cost and improving the user-friendliness of the system, which is especially suitable for operation scenarios of non-technical users.

[0058] 5. Engineering facilities improve the scalability and maintainability of the system

[0059] Existing technical problems: In actual engineering applications, multi-agent systems often face problems such as incompatibility between front-end and back-end technologies, difficulty in deployment, and poor scalability, resulting in high maintenance costs and insufficient flexibility of the system.

[0060] The present invention adopts the supporting engineering facilities of Javascript front-end and Python Django back-end to solve the engineering problems of multi-agent system in front-end and back-end interaction and cross-platform deployment. The Javascript framework of the front-end supports complex interactive interface and real-time data update, and the Django framework of the back-end provides stable task scheduling and data processing capabilities.

[0061] The scalability and maintainability of the system have been significantly improved, allowing developers to more easily integrate the front-end and back-end and expand functions. At the same time, the system deployment is more flexible, supporting rapid implementation and maintenance in a variety of application scenarios, reducing development and operating costs.

[0062] 6. Real-time response and large-scale data processing capabilities

[0063] Existing technical problems: Traditional multi-agent systems often have slow response speeds due to computing bottlenecks or communication delays when processing large-scale data and real-time tasks, affecting the overall performance of the system.

[0064] The present invention significantly improves the real-time response capability of the system through distributed data storage and real-time data processing modules, combined with the reasoning ability of the language model. The system can maintain efficient data processing and task execution in a large-scale data environment, and ensure fast access and sharing of data through distributed storage technology.

[0065] The system has stronger real-time and robustness when processing large-scale data, and can provide stable task execution and data processing capabilities in complex task scenarios, meeting the requirements for high efficiency and low latency.

[0066] In summary, the present invention significantly improves the collaborative efficiency between agents, the flexibility of task decomposition, the adaptability of tool calls, and the intelligent level of user interaction through a multi-agent collaborative system driven by a large language model. Compared with the prior art, the present invention has shown significant technical advantages and effects in processing complex tasks and dynamic environments, and solves many problems in the practical application of multi-agent systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is an application diagram of the multi-agent collaboration system. DETAILED DESCRIPTION

[0068] The multi-agent collaboration system is composed of multiple modules that work together and use a large language model to drive the collaboration, task decomposition and tool calling of multiple agents. It specifically includes a multi-agent collaboration engine module, a task decomposition and allocation module and a tool calling paradigm module. The task decomposition and allocation module is used to decompose tasks and assign the decomposed subtasks to agents; the multi-agent collaboration engine module is used to manage collaboration and communication between agents; and the tool calling module enables agents to dynamically call different tools.

[0069] In addition, it also includes an interaction paradigm module, a task execution and monitoring module, an engineering facility and integration module, and a data storage and processing module; the interaction paradigm module is responsible for managing the interaction between the agent and the user and the internal communication between the agents through a unified language interaction interface; the task execution and monitoring module is used to manage the tracking and feedback of the agent's execution of tasks; the engineering facility and integration module is used to achieve front-end and back-end integration and deployment of the multi-agent collaborative system; the data storage and processing module is used to manage large-scale data processing tasks in the multi-agent collaborative system.

[0070] The following is a detailed description of each module:

[0071] 1. Multi-agent collaboration engine module

[0072] The multi-agent collaboration engine module is the core module of the system, responsible for managing the collaboration and communication between agents. The multi-agent collaboration engine module promotes natural language communication and information sharing between agents through large language models (such as the GPT series model), thereby achieving efficient collaboration of complex tasks. It includes the following functions:

[0073] (1) Natural language driven collaboration: Each agent can communicate with other agents through a natural language model to explain its task progress, requirements, and status. This greatly improves the flexibility of agent collaboration.

[0074] (2) Information sharing and reasoning: Agents can dynamically share information based on task requirements and use large language models to reason and determine the optimal collaboration strategy.

[0075] (3) Conflict resolution: When task conflicts arise between multiple agents, the collaborative engine can use the negotiation mechanism and the reasoning ability of the language model to reach a consensus and resolve the conflict.

[0076] The multi-agent collaboration engine module can also use symbolic language. Although the large language model has advantages in natural language processing, symbolic language or domain-specific communication protocols can be used to replace natural language. For example, the agent can use a predefined symbolic protocol (Symbolic Communication Protocol) to achieve similar interaction effects. This method has higher efficiency and simplicity in certain specific industries or application scenarios, especially in scenarios where complex natural language processing is not required.

[0077] The multi-agent collaboration engine module can also use non-language models driven by machine learning. In addition to language-based collaboration, reinforcement learning or multi-agent game models can also be used to replace natural language interactions. These models allow agents to coordinate with each other through behavior and reward mechanisms without relying on language input. For example, through policy learning, agents can make autonomous decisions based on the environment and the behavior of other agents.

[0078] Furthermore, the multi-agent collaboration engine module is used for collaboration between multiple agents, and the collaboration mode can also be implemented by distributed signal-based collaboration, artificial bee colony or swarm intelligence, as described below:

[0079] Collaboration based on distributed signals: Agent collaboration can also replace language model-driven collaboration through distributed signaling mechanisms. For example, agents can communicate with each other through shared data structures or signaling mechanisms such as blackboard systems or publish / subscribe-based messaging systems. This approach has advantages in large distributed systems or low-latency environments, especially in scenarios where language reasoning is not required.

[0080] Artificial bee colony or swarm intelligence: Agent collaboration can also be replaced by swarm intelligence collaboration methods. Based on artificial bee colony algorithms, particle swarm algorithms, etc., agents can achieve collaboration through local information transmission and group behavior patterns.

[0081] 2. Task decomposition and allocation module

[0082] The task decomposition and assignment module is an important part of the system. It uses a large language model to decompose tasks, understands the task context through natural language, is responsible for decomposing complex tasks into manageable subtasks, and assigns tasks according to the capabilities and current status of the agent. This module not only relies on preset rules, but also can achieve dynamic decomposition and optimization through a large language model. It includes the following functions:

[0083] (1) Task analysis: After the system receives a complex task, the task decomposition module analyzes the task requirements and uses a large language model to understand the task objectives and context.

[0084] (2) Intelligent decomposition: Through the reasoning capability of large language models, the module can intelligently decompose a task into multiple subtasks based on factors such as task complexity and time constraints.

[0085] (3) Dynamic allocation: The module dynamically adjusts the allocation of tasks based on the current status, resource availability, and capabilities of each agent to ensure efficient completion of tasks.

[0086] The task decomposition and allocation module can be replaced by a rule engine and an expert system, or a machine learning algorithm.

[0087] Rule engines and expert systems: They can be replaced by traditional rule-based task decomposition systems or expert systems. These systems use predefined rules and knowledge bases to determine how to decompose tasks and are suitable for task execution in specific fields. For example, in manufacturing or logistics systems, rule engines may automate task allocation based on specific business logic or expert experience.

[0088] Traditional machine learning algorithms: It is also possible to use machine learning algorithms that do not use language models for task decomposition. For example, supervised learning-based models can be trained to identify the subtask structure of complex tasks and automatically assign them to appropriate agents based on historical data. Although these models do not rely on natural language understanding, they can optimize task allocation by analyzing large amounts of task data.

[0089] 3. Tool call paradigm module

[0090] The tool calling paradigm module establishes a flexible tool calling paradigm, which enables the agent to dynamically call different tools or services according to task requirements. This module can analyze task requirements through a large language model and determine the most suitable tool to complete the task. It includes the following functions:

[0091] (1) Tool adaptation and calling: The module has the function of tool library management and can automatically adapt and call appropriate tools (such as APIs, external services, computing resources, etc.) according to the task requirements of the intelligent agent.

[0092] (2) Dynamic selection: When it is necessary to adjust or add tools during a task, the module will make real-time judgments based on the status and requirements of the current task through a large language model and call the corresponding tools.

[0093] (3) Tool result analysis: The results after tool invocation can be analyzed and processed through a large language model, thereby further promoting the execution of the task.

[0094] The tool calling paradigm module relies on a large language model to dynamically select and call tools according to task requirements, ensuring the flexibility and adaptability of tool calling. It can also adopt API-based static calling or decision trees or conditional logic.

[0095] Static API-based calling: You can use the traditional API static calling mode to define the toolset required for the task in advance and match the preset tools according to the task type. For example, for highly repetitive task scenarios, using a predefined tool calling process instead of a language model-driven tool calling can simplify system design.

[0096] Decision trees or conditional logic: Alternatively, the tool calling paradigm can also be implemented through decision trees or conditional logic. According to the different attributes of the task, the decision tree can statically determine the calling order and selection of tools. Although it does not have the flexibility of a large language model, it is suitable for scenarios where the tool calling process is relatively fixed.

[0097] 4. Interaction Paradigm Module

[0098] The interactive paradigm module is responsible for managing the interaction between agents and users and the internal communication between agents. Through the natural language interaction paradigm, users can interact with the system through language instructions, and agents can also interact through language. This allows the module to ensure smooth natural language communication between users and agents through a unified language interaction interface. It includes the following functions:

[0099] (1) User-agent interaction: Users can input tasks through natural language, and the system will automatically recognize the task content and understand and process it through the language model.

[0100] (2) Agent-agent interaction: Agents exchange information through a unified language interface to ensure the collaborative work of multi-agent systems in complex tasks.

[0101] (3) Interaction optimization: Based on user feedback and task execution results, the system will continuously optimize the interaction method so that the agent's response is more in line with user expectations.

[0102] The interaction paradigm module may be replaced by a graphical user interface interaction, or by voice recognition and command control.

[0103] Graphical User Interface (GUI) Interaction: Traditional graphical user interface (GUI) can be used instead of natural language interaction, and users interact with the system by selecting menus or buttons. Although GUI lacks the flexibility of natural language, it is more intuitive in scenarios with frequent operations and clear options, and is suitable for industrial control, customer service and other fields.

[0104] Speech recognition and command control: In terms of natural language interaction, a command control system based on speech recognition can also be used instead. The user issues a fixed-format command through speech, and the system understands the command through speech recognition technology and performs the corresponding operation. Although this method simplifies the processing of natural language, it still maintains the convenience of voice interaction.

[0105] 5.Task execution and monitoring module

[0106] The task execution and monitoring module is responsible for tracking and providing feedback on the agent's task execution. It provides real-time task progress monitoring, ensures the transparency of the task execution process, and intervenes and adjusts when necessary. It includes the following functions:

[0107] (1) Task progress monitoring: The system continuously monitors the agent’s task execution progress and displays the status, progress, and expected completion time of each subtask through a graphical interface.

[0108] (2) Exception handling: When an exception or delay occurs during task execution, the system can dynamically adjust the task execution order or reallocate tasks based on the importance and current status of the tasks.

[0109] (3) Feedback mechanism: Users and agents can provide real-time feedback on task execution. The system adjusts the execution strategy based on the feedback to ensure that the task is completed on time and accurately.

[0110] 6. Engineering facilities and integrated modules

[0111] The engineering facility module is mainly responsible for the front-end and back-end integration and deployment of the system, ensuring that the multi-agent system can run efficiently in actual application scenarios. This module uses the front-end Javascript and back-end Python Django framework technology stack. It includes the following functions:

[0112] (1) Front-end interactive interface (Javascript): The front-end uses Javascript technology to provide users with an intuitive and interactive interface through which users can submit tasks, view task progress, and interact with the intelligent agent.

[0113] (2) Backend management and task scheduling (Django): The backend uses the Python Django framework to provide core services such as task management, agent scheduling, and data processing. It seamlessly connects with the frontend to achieve real-time data synchronization and task execution feedback.

[0114] (3) Cross-platform deployment and expansion: The system is designed to be deployable across platforms, supporting a variety of application scenarios based on cloud computing and local deployment, and ensuring good scalability and maintainability of the system through a microservice architecture.

[0115] The front-end and back-end of the engineering facilities and integrated modules can be replaced by other methods, as shown in the following:

[0116] Front-end interactive interface uses other front-end frameworks: Other front-end frameworks such as React, Vue.js or Angular can be used to replace the Javascript front-end framework. These frameworks provide similar functions, but may have different development efficiency or performance in specific application scenarios. For example, the React framework performs well in building single-page applications (SPA), while Angular is suitable for building complex enterprise-level applications.

[0117] Backend using other backend frameworks: The backend can use other technology stacks such as Node.js, Flask or Ruby on Rails to replace Python Django. Node.js, as an event-driven non-blocking model, can perform better in high-concurrency scenarios, while Flask is suitable for small and fast development application scenarios.

[0118] The backend management framework can be replaced by containerization and microservice architecture: In addition, the backend system of the present invention can also be replaced by a microservice architecture, and efficient management and expansion of the system can be achieved through container technologies such as Docker and Kubernetes, thereby improving the flexibility and scalability of the system.

[0119] 7.Data storage and processing module

[0120] The data storage and processing module is responsible for large-scale data processing tasks in the system, supporting data sharing and task result storage between agents. This module provides reliable storage and fast retrieval services for information during task execution. It includes the following functions:

[0121] (1) Distributed storage: Use distributed databases or cloud storage services to support efficient storage and access of large-scale data and ensure the persistence of task results and intermediate data.

[0122] (2) Real-time data processing: The module supports real-time processing and analysis of data generated during task execution, providing data support for the dynamic decision-making of the intelligent agent.

[0123] (3) Historical data backtracking: Users and systems can backtrack historical task data to analyze and optimize the execution of future tasks.

[0124] The data storage and processing module uses distributed storage and real-time data processing modules to manage large-scale data streams and provide real-time feedback and historical data storage during task execution. Its distributed storage and real-time data processing can be replaced by other solutions, specifically:

[0125] Traditional database solutions: You can replace traditional centralized databases (such as MySQL and PostgreSQL) to manage data storage. These databases are stable and easy to maintain in small and medium-scale application scenarios, and are suitable for most scenarios that do not require distributed processing.

[0126] Edge computing and IoT integration: When processing real-time data, edge computing technology can be used to complete some data processing tasks at edge nodes close to the data source, thereby reducing data transmission delays and improving real-time response capabilities. This approach is particularly suitable for use in IoT and smart device networks.

[0127] Through the above-mentioned structure and modules, the multi-agent collaboration system of the present invention can effectively improve the coordination, adaptability and scalability of complex task processing, and solve many problems of existing systems in terms of agent collaboration, task decomposition, tool calling, etc.

[0128] Through alternative solutions, the technical scope of the present invention can be further expanded to cover a variety of solutions for achieving similar goals, thereby strengthening patent protection and preventing others from circumventing patent restrictions by changing certain structures or technical means to achieve similar invention purposes.

[0129] like Figure 1 As shown, the multi-agent collaboration system, when applied, includes the following steps:

[0130] S1: Communicate with the user input information through the interactive paradigm module to obtain the task;

[0131] S2: Decompose the task into multiple subtasks through the task decomposition and assignment module and assign them to different agents;

[0132] S3: The agents communicate with each other through a multi-agent collaboration engine module, and the agents can call tools through a tool calling paradigm module;

[0133] S4: The system implements real-time task progress monitoring through the task execution and monitoring module;

[0134] S5: The system provides reliable storage and fast retrieval services for information during task execution through data storage and processing modules;

[0135] S6: The final solution is sent to users through engineering facilities and integrated modules.

[0136] The definitions of abbreviations and key terms used in the present invention are as follows:

[0137] API Application Programming Interface, application programming interface

[0138] NLP Natural Language Processing

[0139] LM Language Model

[0140] LLM Large Language Model

[0141] CPU Central Processing Unit

[0142] GPU Graphic Processing Unit, graphics processor

[0143] Agent

[0144] Workflow

[0145] The present invention has the following characteristics:

[0146] 1. Multi-agent collaboration engine based on large language model. The present invention uses a multi-agent collaboration engine driven by a large language model (such as the GPT series) to enable agents to coordinate and collaborate on complex tasks through natural language. The engine can understand the communication content between agents, infer task requirements, and dynamically optimize collaboration strategies.

[0147] It significantly improves the communication efficiency between agents and the flexibility of task collaboration, and is particularly suitable for multi-agent collaborative task scenarios in complex and dynamic environments.

[0148] 2. Intelligent optimization mechanism for task decomposition and allocation, using large language models to intelligently decompose complex tasks, and dynamically assign tasks to the most suitable agents by understanding the context and requirements of the tasks. This task decomposition and allocation mechanism based on semantic understanding is superior to traditional static rules or allocation methods based on preset conditions.

[0149] It improves the flexibility and efficiency of task decomposition and allocation, can adapt to task requirements of different scenarios and complexity, and reduce resource waste and time delays during task execution.

[0150] 3. Adaptive paradigm of tool calling: the system uses the reasoning ability of the large language model to dynamically call appropriate tools or services according to task requirements, and adjusts the use of tools in real time according to the intermediate results of task execution. Different from the traditional static calling mechanism, the present invention can adjust the calling of tools in real time according to the changes of tasks.

[0151] The flexibility of tool calling and the adaptability of task execution have been enhanced, especially when facing complex or changing task requirements, which can ensure the smooth completion of tasks and dynamically optimize the execution process.

[0152] 4. Natural language interaction paradigm: interact with users through natural language. Users can input tasks in natural language, and the system understands and decomposes tasks through language models, without the need for users to master complex technical instructions. This natural language interaction paradigm is also used for communication and information sharing between intelligent agents.

[0153] It improves the user experience of the system, allowing non-technical users to easily operate the system, lowers the threshold for using the system, and improves the efficiency of collaboration between intelligent agents.

[0154] 5. The front-end and back-end supporting engineering facilities adopt the supporting architecture of Javascript front-end and Python Django back-end to ensure the cross-platform integration, deployment and expansion of multi-agent systems in practical applications, especially in large-scale data processing and high-concurrency task scenarios.

[0155] The system's scalability, maintainability and cross-platform adaptability are improved, and it is particularly suitable for multi-agent collaboration and complex task execution in enterprise-level applications.

Claims

1. A multi-agent collaborative system, characterized in that: It includes a multi-agent collaboration engine module, a task decomposition and allocation module and a tool calling paradigm module. The task decomposition and allocation module is used to decompose tasks and allocate the decomposed subtasks to agents; the multi-agent collaboration engine module is used to manage collaboration and communication between agents; The tool calling module enables the agent to dynamically call different tools.

2. The multi-agent collaboration system according to claim 1, characterized in that: The multi-agent collaboration engine module is used to promote natural language communication and information sharing between agents through a large language model, so that each agent can communicate with other agents through the large language model, so that each agent can dynamically share information according to task requirements, and can resolve task conflicts between multiple agents through a negotiation mechanism.

3. The multi-agent collaboration system according to claim 2, characterized in that: The multi-agent collaboration engine module can use symbolic language and machine learning driven non-language models to replace large language models.

4. The multi-agent collaboration system according to claim 2, characterized in that: The multi-agent collaboration engine module can implement collaboration between multiple agents by using distributed signal-based collaboration, artificial bee colonies or group intelligence collaboration.

5. The multi-agent collaboration system according to claim 1, characterized in that: The task decomposition and allocation module understands the task context through natural language, is responsible for decomposing complex tasks into manageable subtasks, and allocates tasks according to the capabilities and current status of the intelligent agent, so as to realize task analysis, intelligent decomposition and dynamic allocation; the task decomposition and allocation module can be replaced by a rule engine and expert system, or a machine learning algorithm.

6. The multi-agent collaboration system according to claim 1, characterized in that: The tool calling paradigm module can analyze task requirements through a large language model and determine the most suitable tool to complete the task, so as to achieve tool adaptation and calling, dynamic tool selection, and tool result analysis. It can be replaced by API-based static calling or decision tree or conditional logic.

7. The multi-agent collaboration system according to claim 1, characterized in that: It also includes an interaction paradigm module, a task execution and monitoring module, an engineering facility and integration module, and a data storage and processing module; the interaction paradigm module is responsible for managing the interaction between the agent and the user and the internal communication between the agents through a unified language interaction interface; the task execution and monitoring module is used to manage the tracking and feedback of the agent's execution of tasks; the engineering facility and integration module is used to realize the front-end and back-end integration and deployment of the multi-agent collaboration system; the data storage and processing module is used to manage large-scale data processing tasks in the multi-agent collaboration system.

8. The multi-agent collaboration system according to claim 7, characterized in that: The interaction paradigm module includes the following interaction modes: (1) User-agent interaction: Users input tasks through natural language, and the system automatically recognizes the task content and understands and processes it through the language model; (2) Agent-agent interaction: Agents exchange information through a unified language interface to ensure the collaborative work of multi-agent systems in complex tasks; (3) Interaction optimization: Based on user feedback and task execution results, the system continuously optimizes the interaction method.

9. The multi-agent collaboration system according to claim 1, characterized in that: The task execution and monitoring module includes the following processing contents: (1) Task progress monitoring: The system continuously monitors the agent's task execution progress and displays the status, progress, and expected completion time of each subtask through a graphical interface; (2) Exception handling: When an exception or delay occurs during task execution, the system can dynamically adjust the task execution order or reallocate tasks based on the importance and current status of the tasks; (3) Feedback mechanism: Users and agents provide real-time feedback on task execution, and the system adjusts its execution strategy based on the feedback.

10. The multi-agent collaboration system according to claim 1, characterized in that: The data storage and processing module includes the following processing contents: (1) Distributed storage: Use distributed databases or cloud storage services to support efficient storage and access of large-scale data; (2) Real-time data processing: supports real-time processing and analysis of data generated during task execution; (3) Historical data backtracking: Users and systems can backtrack historical task data to analyze and optimize the execution of future tasks.

Images