An automated office task processing method and system based on multi-agent cooperation

Abstract

The application relates to the technical field of intelligent agents, in particular to an automatic office task processing method and system based on multi-agent cooperation. The automatic office task processing method based on multi-agent cooperation performs deep analysis on user instructions, identifies core intentions and requirements, and generates an execution plan; a tool execution agent reads original data and transmits the data to a content generation agent, the content generation agent inserts a preset PPT template in combination with data analysis results, and sends a PPT report to the user. The automatic office task processing method and system based on multi-agent cooperation deploy a localized model by using an edge computing node, realize autonomous decomposition and cross-application cooperation of tasks, have multi-modal interaction and logical reasoning capabilities, can seamlessly integrate enterprise software ecology, provide an end-to-end automatic solution, significantly improve the office automation level of enterprises, reduce manual operation errors, and protect enterprise data privacy through private deployment.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent technology, and in particular to an automated office task processing method and system based on multi-agent collaboration. Background Technology

[0002] Currently, with the rapid development and widespread application of Large Language Models (LLM), AI-assisted office work is gradually becoming more common and an important tool for improving work efficiency, widely used in document processing, data analysis, and communication collaboration. However, despite significant technological advancements, the following pain points still exist in practical applications: (1) Limited task complexity: Existing AI assistants (such as the ChatGPT plugin) are mostly "lone wolf" and can usually only perform a single task or respond to simple instructions. They are difficult to handle complex tasks that require multi-step logical judgment and cross-software collaboration. For example, when creating a business report, AI cannot automatically extract data from Excel spreadsheets, perform statistical analysis, and visualize the results in PPT to generate charts, which means that users still need to manually operate multiple steps.

[0003] (2) Illusion and accuracy issues: General large models are prone to illusion when processing enterprise-specific business data (such as financial statement formats and industry terminology). Due to a lack of domain expertise or contextual understanding, the output results may not conform to the norms or even introduce erroneous information. This not only affects the quality of decision-making but may also lead to compliance risks.

[0004] (3) Data security risks: There is a risk of leakage when the cloud-based general model processes internal enterprise data. Enterprises have extremely high requirements for the protection of sensitive information (such as customer information and financial records). Therefore, they urgently need private and localized intelligent solutions to ensure that data is processed in a controllable environment and to avoid external access or abuse.

[0005] (4) Low level of automation: Existing technologies mostly remain at the level of text generation and lack the ability to directly operate specific software such as Office / WPS. Users still need to manually perform operations such as copying and pasting, formatting adjustment or software interaction. The automation process is incomplete, which limits the deep integration of AI in office scenarios.

[0006] Therefore, there is an urgent need for an intelligent agent system that can understand complex instructions, invoke specific tools, and collaboratively complete tasks, just like human employees.

[0007] To address the problems of logical confusion, inability to operate specific software, and high data security risks associated with single AI models in handling complex office tasks in existing technologies, this invention proposes an automated office task processing method and system based on multi-agent collaboration. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a simple and efficient automated office task processing method and system based on multi-agent collaboration.

[0009] This invention is achieved through the following technical solution: An automated office task processing method based on multi-agent collaboration includes the following steps: Step S1: Instruction Parsing and Planning After the user inputs a command, the task planning agent performs in-depth analysis of the command to identify the user's core intent and needs; Based on the analysis results, the task planning agent generates an execution plan, which includes four key stages: data extraction, data analysis, document generation and formatting. The goals, tools and outputs of each stage are clearly defined to provide structured guidance for subsequent processes. Step S2, Tool Invocation and Data Acquisition The task planning agent assigns the data extraction task to the tool execution agent, which then launches the integrated Excel plugin. The Excel add-in automatically connects to the file at the specified path, reads the raw data, and uses built-in algorithms to filter and calculate, extracting custom key indicators. In step S2, the tool execution agent also performs preliminary data cleaning to ensure data quality and lay the foundation for subsequent analysis.

[0010] Step S3: Content Generation and Arrangement The tool execution agent passes the extracted data to the content generation agent, which combines the data analysis results to write a structured analysis report text, including an overview, findings, and recommendations. Meanwhile, the task planning agent coordinates resources and inserts the generated text and automatically generated charts into preset PPT templates, supporting custom adjustments to the layout and style to maintain a professional and aesthetically pleasing appearance.

[0011] Step S4, Result Output and Feedback Once the document is generated, the final PPT report will be automatically saved to the path specified by the user, and a completion status notification will be sent to the user, including the file location and the generated summary. In step S4, users can view the PPT report and rate it based on the accuracy of the content and the satisfaction with the format; the rating data is collected and used for subsequent model fine-tuning to optimize the performance and adaptability of the agent.

[0012] An automated office task processing system based on multi-agent collaboration adopts a four-layer architecture: user layer, agent layer, tool layer, and edge computing layer. This layered design ensures modularity and scalability of task execution. The system is used to implement the aforementioned methods and includes: The user layer includes a user command interaction module, which is responsible for receiving office task commands input by users through natural language and implementing multimodal interaction to improve user experience; The user command interaction module supports text and voice input as well as multilingual processing (including languages ​​from various countries and ethnic groups both domestically and internationally). Voice input integrates noise reduction and voice recognition optimization technologies. It supports intent recognition, based on a pre-trained natural language processing model (NLP) and combined with context analysis to accurately parse user needs. At the same time, it filters invalid or ambiguous instructions through a rule engine and machine learning classifier to ensure the clarity and executability of instructions.

[0013] In addition, the user command interaction module is also responsible for providing real-time feedback, including confirming commands or requesting supplementary information, to enhance the smoothness of the interaction.

[0014] The agent layer includes the task planning agent, the tool execution agent, the content generation agent, and the multi-agent collaboration hub; The task planning agent is responsible for deep intent understanding of user commands based on the large language model LLM, using knowledge graphs to break down macro tasks into executable atomic steps, and generating a directed acyclic graph (DAG) to optimize the execution order. The task planning agent adopts the ReAct (Reasoning + Acting) framework, combined with CoT (Cooperation of Thought) technology for multi-step reasoning, and dynamically adjusts the plan to deal with anomalies; at the same time, it integrates reinforcement learning to learn optimization strategies from historical tasks and improve planning accuracy.

[0015] The tool's execution agent is responsible for calling specific APIs or RPA scripts to operate local or cloud-based office software, supporting cross-platform compatibility. The tool's Executor Agent operates Excel / WPS via the COM interface to read and write data and perform formula calculations. It automatically generates slides and applies enterprise templates through the PPTX library, or calls the enterprise's internal ERP / CRM interface to obtain real-time data. It extends support for manipulating Word documents, sending Outlook emails, and automating browser operations, ensuring coverage of common office scenarios.

[0016] The content generation agent is responsible for automatically generating report text, chart annotations, and presentation scripts based on the data extracted by the tool's execution agent, using a template engine and generative AI model. It supports multi-format output, while also performing syntax checking and content optimization to improve readability. The multi-agent collaboration hub is responsible for enabling asynchronous communication between the task planning agent, the tool execution agent, and the content generation agent through the message bus, thereby achieving collaborative work among the agents. The message bus adopts an event-driven architecture to ensure highly reliable and low-latency data exchange.

[0017] The tool layer is responsible for encapsulating the operation interfaces for Microsoft Office, WPS Office, browsers, and enterprise internal systems, providing a unified and standardized access layer to reduce the complexity of agent invocation.

[0018] The tool layer develops a standardized ToolKit, including the read_excel function, write_ppt function, and send_email function, and expands upon it with more tools, including data visualization, API gateway calling, and workflow orchestration. The tool layer adopts a plug-in design, supports third-party extensions and custom integrations, and achieves efficient communication through RESTful API or gRPC protocol, ensuring flexibility and maintainability.

[0019] The edge computing layer is responsible for deploying lightweight large models on local terminals or enterprise intranet edge servers to reduce latency and enhance privacy protection; Sensitive data is processed locally, and the edge computing layer only uploads the de-identified instructions or results to the cloud to ensure data security.

[0020] Lightweight large models are optimized through pruning and quantization techniques to maintain high-performance inference in resource-constrained environments, and are deployed using containerization to achieve rapid scaling and isolation.

[0021] Data anonymization employs differential privacy and data masking technologies to prevent information leakage, while also supporting offline mode to ensure service continuity when the network is unstable, thereby improving the overall reliability and compliance of the system.

[0022] An automated office task processing device based on multi-agent collaboration includes a memory and a processor; the memory is used to store a computer program, and the processor is used to implement the above-described method when executing the computer program.

[0023] A readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0024] The beneficial effects of this invention are as follows: This automated office task processing method and system based on multi-agent collaboration has multimodal interaction and logical reasoning capabilities, can seamlessly integrate with the enterprise software ecosystem, provide end-to-end automation solutions, and utilizes edge computing nodes to deploy localized models, thereby realizing autonomous task decomposition and cross-application collaboration, which significantly improves the level of enterprise office automation, reduces human error, and protects enterprise data privacy through private deployment. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Appendix Figure 1 This is a schematic diagram of the automated office task processing method based on multi-agent collaboration of the present invention.

[0027] Appendix Figure 2 This is a schematic diagram of the architecture of the automated office task processing system based on multi-agent collaboration of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions in the embodiments of this invention will be clearly and completely described below in conjunction with the embodiments of this invention. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0029] This automated office task processing method based on multi-agent collaboration includes the following steps: Step S1: Instruction Parsing and Planning After the user inputs a command, the task planning agent performs in-depth analysis of the command to identify the user's core intent and needs; Based on the analysis results, the task planning agent generates a detailed execution plan, which includes four key stages: data extraction, data analysis, document generation and formatting. It also clearly defines the goals, tools and outputs of each stage, providing structured guidance for subsequent processes. Step S2, Tool Invocation and Data Acquisition The task planning agent assigns the data extraction task to the tool execution agent, which then launches the integrated Excel plugin. The Excel add-in automatically connects to a file at a specified path, reads the raw data, and uses built-in algorithms to filter and calculate, extracting custom key indicators such as sales revenue and growth rate. In step S2, the tool execution agent also performs preliminary data cleaning to ensure data quality and lay the foundation for subsequent analysis.

[0030] Step S3: Content Generation and Arrangement The tool execution agent passes the extracted data to the content generation agent, which combines the data analysis results to write a structured analysis report text, including an overview, findings, and recommendations. Meanwhile, the task planning agent coordinates resources and precisely inserts the generated text and automatically generated charts (such as chart agents creating visualizations based on data) into preset PPT templates, supporting custom adjustments to layout and style to maintain a professional and aesthetically pleasing appearance.

[0031] Step S4, Result Output and Feedback Once the document is generated, the final PPT report will be automatically saved to the path specified by the user, and a completion status notification will be sent to the user, including the file location and the generated summary. In step S4, users can view the PPT report and rate it based on the accuracy of the content and the satisfaction with the format; the rating data is collected and used for subsequent model fine-tuning to optimize the performance and adaptability of the agent.

[0032] This automated office task processing system based on multi-agent collaboration adopts a four-layer architecture: user layer, agent layer, tool layer, and edge computing layer. This layered design ensures the modularity and scalability of task execution. The system is used to implement the aforementioned methods, including: The user layer includes a user command interaction module, which is responsible for receiving office task commands input by users through natural language (such as "generate a PPT report based on last month's sales data") and implementing multimodal interaction to improve user experience. The user command interaction module supports text and voice input as well as multilingual processing (including languages ​​from various countries and ethnic groups both domestically and internationally). Voice input integrates noise reduction and voice recognition optimization technologies. It supports intent recognition, based on a pre-trained natural language processing model (NLP) and combined with context analysis to accurately parse user needs. At the same time, it filters invalid or ambiguous instructions through a rule engine and machine learning classifier to ensure the clarity and executability of instructions.

[0033] In addition, the user command interaction module is also responsible for providing real-time feedback, including confirming commands or requesting supplementary information, to enhance the smoothness of the interaction.

[0034] The agent layer includes the task planning agent, the tool execution agent, the content generation agent, and the multi-agent collaboration hub; The task planning agent is responsible for performing deep intent understanding of user commands based on large language models (LLMs) such as Qwen and Llama3. It uses knowledge graphs to break down macro tasks into executable atomic steps and generates directed acyclic graphs (DAGs) to optimize the execution order. The following is a code example of how to use the ReAct framework and CoT (CoT) to decompose user instructions into a directed acyclic graph (DAG) and schedule the tool to execute the agent: class PlannerAgent: def __init__(self, llm_model="Qwen-72B"): self.llm = LLMWrapper(model=llm_model) self.memory = VectorMemory() # Long-term memory vector library def parse_and_plan(self, user_instruction): # 1. Intent Recognition and Context Retrieval context = self.memory.retrieve_similar_tasks(user_instruction) # 2. Reasoning based on the ReAct framework prompt = f""" [System] You are a Planner Agent. [Context] {context} [User Instruction] {user_instruction} [Task] Break down the instruction into atomic steps (DAGformat). [Output Format] JSON list of steps with dependencies. """ plan_json = self.llm.generate(prompt) # 3. Generate the DAG execution graph dag = self.build_dag(plan_json) return dag class ExecutorAgent: def execute_step(self, step): tool_name = step['tool'] params = step['params'] # Dynamically Loading Tool Plugins tool_instance = ToolRegistry.get(tool_name) try: # Perform specific operations, such as manipulating Excel or calling APIs. result = tool_instance.run(**params) return result except Exception as e: # Trigger exception handling mechanism (see code snippet 3) return self.handle_exception(e, step) # Main process scheduling def main_workflow(user_input): planner = PlannerAgent() executor = ExecutorAgent() # Step 1: Instruction Parsing and Planning dag = planner.parse_and_plan(user_input) # Step 2: Execute according to DAG order for step in dag.topological_sort(): if step['type'] == 'data_extraction': data = executor.execute_step(step) elif step['type'] == 'content_generation': # Passing data to the generating agent pass The task planning agent adopts the ReAct (Reasoning + Acting) framework, combined with CoT (Cooperation of Thought) technology for multi-step reasoning, and dynamically adjusts the plan to deal with anomalies; at the same time, it integrates reinforcement learning to learn optimization strategies from historical tasks and improve planning accuracy.

[0035] The tool's execution agent is responsible for calling specific APIs or RPA scripts to operate local or cloud-based office software, supporting cross-platform compatibility. The tool's Executor Agent operates Excel / WPS via the COM interface to read and write data and perform formula calculations. It automatically generates slides and applies enterprise templates through the PPTX library, or calls the enterprise's internal ERP / CRM interface to obtain real-time data. It extends support for manipulating Word documents, sending Outlook emails, and automating browser operations, ensuring coverage of common office scenarios.

[0036] The content generation agent is responsible for automatically generating report text, chart annotations, and presentation scripts based on the data extracted by the tool's execution agent, using a template engine and generative AI model. It supports multi-format output (such as PDF and HTML), while also performing syntax checking and content optimization to improve readability. The content generation agent is responsible for writing copy, summarizing data, and creating charts and graphs, combining corporate brand guidelines and style libraries to generate professional and consistent content.

[0037] The multi-agent collaboration hub is responsible for enabling asynchronous communication between the task planning agent, the tool execution agent, and the content generation agent through the message bus, thereby achieving collaborative work among the agents. The message bus adopts an event-driven architecture (such as RabbitMQ or Kafka) to ensure highly reliable and low-latency data exchange.

[0038] The intelligent agent possesses both short-term and long-term memory. Short-term memory is specifically used to store the context information of the current task, ensuring data continuity and real-time performance during the execution of a single task. Long-term memory is built upon an enterprise knowledge base and past task experience, and is efficiently stored, indexed, and retrieved through a vector database. This hierarchical memory system enables the intelligent agent to dynamically invoke historical experience and knowledge when handling complex tasks, significantly improving the continuity, accuracy, and adaptability of task processing. The vector database utilizes semantic embedding technology to achieve similarity matching, helping the intelligent agent intelligently associate relevant memory fragments, thereby optimizing decision-making processes and execution efficiency.

[0039] Anomaly Handling Mechanism: When a tool call fails, such as due to file being occupied, insufficient resources, or external interface anomalies, the agent automatically triggers a retry mechanism with preset retries and delay intervals to attempt recovery. If the retry still fails, the agent intelligently analyzes the cause of the failure and automatically switches to a backup plan, such as enabling alternative tools, adjusting parameters, or executing a degradation process, ensuring the task continues uninterrupted. The entire anomaly handling process is fully automated, requiring no manual intervention, thus enhancing the system's robustness, fault tolerance, and operational efficiency.

[0040] Human-Machine Collaboration Mode: At critical decision points, such as the generation of reports involving large-scale fund approvals, high-risk operations, or compliance checks, the intelligent agent automatically identifies the risk level and pauses execution. Simultaneously, it sends a detailed confirmation request to designated personnel, including a contextual summary and potential impact analysis, awaiting human review and authorization. This "human-in-the-loop" safety control mode fully leverages the intelligent agent's high-speed processing and data integration capabilities while ensuring the prudence and reliability of critical decisions through human supervision, achieving a perfect balance between automation and security. After human confirmation, the intelligent agent continues execution, adjusts tasks, or records learning experiences based on feedback, forming a closed-loop management system of intelligent and human collaboration, enhancing the transparency and credibility of the overall workflow.

[0041] The tool layer is responsible for encapsulating the operation interfaces for Microsoft Office, WPS Office, browsers, and enterprise internal systems, providing a unified and standardized access layer to reduce the complexity of agent invocation.

[0042] The tool layer develops a standardized ToolKit, including the read_excel function, write_ppt function, and send_email function, and expands upon it with more tools, including data visualization, API gateway calling, and workflow orchestration. The tool layer adopts a plug-in design, supports third-party extensions and custom integrations, and achieves efficient communication through RESTful API or gRPC protocol, ensuring flexibility and maintainability.

[0043] The edge computing layer is responsible for deploying lightweight large models (such as LoRA-based fine-tuned models) on local terminals or enterprise intranet edge servers to reduce latency and enhance privacy protection. Sensitive data (such as financial data) is processed locally, and the edge computing layer only uploads the de-identified instructions or results to the cloud to ensure data security.

[0044] Data anonymization employs differential privacy and data masking technologies to prevent information leakage, while also supporting offline mode to ensure service continuity when the network is unstable, thereby improving the overall reliability and compliance of the system.

[0045] Lightweight large models are optimized through pruning and quantization techniques to maintain high-performance inference in resource-constrained environments, and are deployed using containerization to achieve rapid scaling and isolation.

[0046] Load the lightweight, quantized, and LoRA-tuned large model at the edge (local terminal or intranet server) to ensure that the data does not leave the domain. Code example is as follows: import torch from transformers import AutoModelForCausalLM, AutoTokenizer,BitsAndBytesConfig class EdgeComputingNode: def __init__(self, model_path="local-model-lora"): self.model_path = model_path self.device = "cuda" if torch.cuda.is_available() else "cpu" self.model = None def load_lightweight_model(self): # 1. Configure quantization parameters (4-bit quantization to reduce video memory usage) quantization_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16 ) # 2. Load the base model and LoRA adapter self.tokenizer = AutoTokenizer.from_pretrained(self.model_path) self.model = AutoModelForCausalLM.from_pretrained( self.model_path, quantization_config=quantization_config, device_map="auto" ) # 3. Load LoRA fine-tuning weights for specific enterprises self.model.load_adapter(f"{self.model_path} / lora-adapter-enterprise") print("Edge Model Loaded Successfully with LoRA") def process_sensitive_data(self, raw_data): # Data anonymization (differential privacy / masking) masked_data = self.apply_data_masking(raw_data) # Local inference, no cloud upload inputs = self.tokenizer(masked_data, return_tensors="pt").to(self.device) outputs = self.model.generate(**inputs, max_new_tokens=512) return self.tokenizer.decode(outputs[0], skip_special_tokens=True) def apply_data_masking(self, data): # Simulated data anonymization logic return data.replace("Confidential", "[REDACTED]") This automated office task processing device based on multi-agent collaboration includes a memory and a processor; the memory is used to store computer programs, and the processor is used to execute the methods implemented when the computer programs are executed.

[0047] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0048] Memory can be used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.

[0049] The readable storage medium stores a computer program, the method of which is implemented when the computer program is executed by a processor.

[0050] Compared with existing technologies, this automated office task processing method and system based on multi-agent collaboration has the following characteristics: (1) Significantly improved office efficiency: By introducing advanced automation technology, the manual cross-software operations that originally required several hours (such as data import and export, format conversion or report generation) were optimized into automated processes that took minutes. This not only significantly reduced task completion time, but also freed up employee resources, allowing them to focus on more creative and strategic work, thereby improving overall organizational productivity.

[0051] (2) Reduced error rate: By automating tasks, common data errors in manual copy-pasting processes, such as missing information, inconsistent formatting, or input errors, are completely eliminated. The automated process ensures the accuracy and consistency of data processing, reduces the cost of rework and error correction due to human negligence, and thus improves the overall work quality and reliability.

[0052] (3) Improved data security: Core data processing is completed at the edge, meaning that sensitive information does not need to be uploaded to the cloud, effectively reducing the risk of data leakage and unauthorized access. This design strictly complies with enterprise data compliance requirements, ensuring the security and privacy of customer and business data, and enhancing the enterprise's trustworthiness and compliance.

[0053] (4) Flexible and scalable: By adding new ToolKit modules, the intelligent agent can quickly adapt to new office software or business systems without large-scale reconstruction or interruption of existing processes. This modular design enables the solution to be easily expanded as technology develops and business needs change, ensuring long-term applicability and return on investment, and supporting continuous innovation and growth of enterprises.

[0054] The foregoing has provided a detailed description of an automated office task processing method and system based on multi-agent collaboration, as exemplified by this invention. This section uses specific examples to illustrate the principles and implementation methods of the invention. These examples are merely for the purpose of helping to understand the core ideas of this invention. All other embodiments obtained by those skilled in the art without creative effort, without departing from the principles of this invention, should fall within the scope of protection of this invention.

Claims

1. A method for automating office task processing based on multi-agent collaboration, characterized in that: Includes the following steps: Step S1: Instruction Parsing and Planning After the user inputs a command, the task planning agent performs in-depth analysis of the command to identify the user's core intent and needs; Based on the analysis results, the task planning agent generates an execution plan, which includes four key stages: data extraction, data analysis, document generation and formatting. The goals, tools and outputs of each stage are clearly defined to provide structured guidance for subsequent processes. Step S2, Tool Invocation and Data Acquisition The task planning agent assigns the data extraction task to the tool execution agent, which then launches the integrated Excel plugin. The Excel add-in automatically connects to the file at the specified path, reads the raw data, and uses built-in algorithms to filter and calculate, extracting custom key indicators. Step S3: Content Generation and Arrangement The tool execution agent passes the extracted data to the content generation agent, which combines the data analysis results to write a structured analysis report text, including an overview, findings, and recommendations. Meanwhile, the task planning agent coordinates resources and inserts the generated text and automatically generated charts into the preset PPT template, supporting custom adjustments to the layout and style; Step S4, Result Output and Feedback Once the document is generated, the PPT report will be automatically saved to the path specified by the user, and a completion status notification will be sent to the user, including the file location and the generated summary.

2. The automated office task processing method based on multi-agent collaboration according to claim 1, characterized in that: In step S2, the tool execution agent also performs preliminary data cleaning to ensure data quality and lay the foundation for subsequent analysis.

3. The automated office task processing method based on multi-agent collaboration according to claim 1, characterized in that: In step S4, users can view the PPT report and rate it based on the accuracy of the content and the satisfaction with the format. The scoring data was collected and used for subsequent model fine-tuning to optimize the agent's performance and adaptability.

4. An automated office task processing system based on multi-agent collaboration, characterized in that: A four-layer architecture consisting of a user layer, an intelligent agent layer, a tool layer, and an edge computing layer is adopted. The layered design ensures the modularity and scalability of task execution, and is used to implement the method described in any one of claims 1 to 3. The user layer includes a user command interaction module, which is responsible for receiving office task commands input by users through natural language and implementing multimodal interaction to improve user experience; The agent layer includes task planning agents, tool execution agents, content generation agents, and a multi-agent collaboration hub; The task planning agent is responsible for deep intent understanding of user commands based on the large language model LLM, using knowledge graphs to decompose macro tasks into executable atomic steps, and generating a directed acyclic graph (DAG) to optimize the execution order. The tool executes an intelligent agent, which is responsible for calling specific APIs or RPA scripts to operate local or cloud-based office software, and supports cross-platform compatibility; The tool executes an intelligent agent that operates Excel / WPS through a COM interface to read and write data and perform formula calculations. It automatically generates slides and applies enterprise templates through the PPTX library, or calls the enterprise's internal ERP / CRM interface to obtain real-time data. Extended support for Word document manipulation, Outlook email sending, and browser automation ensures coverage of common office scenarios; The content generation agent is responsible for automatically generating report text, chart annotations, and presentation scripts based on the data extracted by the tool's execution agent, using a template engine and generative AI model. It supports multi-format output, while also performing syntax checks and content optimization to improve readability. The multi-agent collaboration hub is responsible for enabling asynchronous communication between task planning agents, tool execution agents, and content generation agents through a message bus, thereby achieving collaborative work among the agents. The message bus adopts an event-driven architecture to ensure highly reliable and low-latency data exchange; The tool layer is responsible for encapsulating the operation interfaces for Microsoft Office, WPS Office, browsers, and enterprise internal systems, providing a unified and standardized access layer; The tool layer adopts a plug-in design, supports third-party extensions and custom integrations, and communicates through RESTful API or gRPC protocol to ensure flexibility and maintainability; The edge computing layer is responsible for deploying lightweight large models on local terminals or enterprise intranet edge servers to reduce latency and enhance privacy protection; Sensitive data is processed locally, and the edge computing layer only uploads the de-identified instructions or results to the cloud to ensure data security. The lightweight large model is optimized through pruning and quantization techniques, and its deployment adopts containerization to achieve rapid expansion and isolation.

5. The automated office task processing system based on multi-agent collaboration according to claim 4, characterized in that: The user command interaction module supports text and voice input as well as multilingual processing. The voice input integrates noise reduction and voice recognition optimization technologies. It supports intent recognition, based on a pre-trained natural language processing model (NLP) and combined with context analysis to accurately parse user needs. At the same time, it filters invalid or ambiguous instructions through a rule engine and machine learning classifier to ensure the clarity and executability of instructions. In addition, the user command interaction module is also responsible for providing real-time feedback, including confirming commands or requesting supplementary information, to enhance the smoothness of the interaction.

6. The automated office task processing system based on multi-agent collaboration according to claim 4, characterized in that: The task planning agent adopts the ReAct framework and combines the CoT (Cooperation of Thought) technology for multi-step reasoning to dynamically adjust the plan to deal with anomalies; at the same time, it integrates reinforcement learning to learn optimization strategies from historical tasks and improve planning accuracy.

7. The automated office task processing system based on multi-agent collaboration according to claim 4, characterized in that: The tool layer develops a standardized ToolKit, including the read_excel, write_ppt, and send_email functions, and expands upon it with more tools, including data visualization, API gateway calls, and workflow orchestration.

8. The automated office task processing system based on multi-agent collaboration according to claim 4, characterized in that: Data anonymization employs differential privacy and data masking technologies to prevent information leakage, while also supporting offline mode to ensure service continuity when the network is unstable, thereby improving the overall reliability and compliance of the system.

9. An automated office task processing device based on multi-agent collaboration, characterized in that: It includes a memory and a processor; the memory is used to store a computer program, and the processor is used to execute the computer program to implement the method as described in any one of claims 1 to 3.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 3.

Images