Methods, systems, devices, processors, and readable storage media for large-scale multi-agent collaborative processing in securities business.

Abstract

This invention relates to a method for large-scale multi-agent collaborative processing of securities business, comprising the following steps: generating a directed acyclic graph (DAG) depicting the dependencies between subtasks; grouping roles according to subtask types and assigning different atomic subtasks to agent groups with corresponding functions; conducting multiple rounds of review of intermediate outputs through an annotation-based review mechanism, triggering a debate-based generation mechanism for comprehensive decision-making tasks, and outputting intermediate results after cross-validation; and generating the final business output. This invention employs the method, system, device, processor, and computer-readable storage medium for large-scale multi-agent collaborative processing of securities business, based on core mechanisms such as task graph generation and decomposition, role grouping, and annotation-based review, and combined with a three-layer architecture design of a basic model layer, an agent orchestration layer, and a business adaptation layer. This achieves high-precision, high-efficiency, and high-reliability intelligent services in automating complex securities business processes.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and large language models, and particularly to the field of multi-agent system technology. Specifically, it refers to a method, system, device, processor, and computer-readable storage medium for realizing large-scale multi-agent collaborative processing of securities business. Background Technology

[0002] 1.1 Current Status of Intelligent Securities Business and Limitations of Single-Agent Technology With the rise of Large Language Models (LLMs), intelligent applications in the financial securities field have ushered in new opportunities. Currently, securities companies have begun exploring the application of large-model intelligent agents, but this mainly remains at the level of "single agent function stacking". This early application approach attempts to cope with complex business processes through a single model or simply stacking multiple independent functional modules. For example, in scenarios such as intelligent investment advisory, information processing, and compliance review, they often rely on a single large model to directly generate suggestions or process data.

[0003] However, given the massive scale of securities business, the complexity of its operational processes, and the extremely high requirements for real-time performance, the single-agent model faces significant technical bottlenecks: 1. Complex processes and delayed responses: A single large model struggles to effectively break down and execute complex financial tasks involving multiple steps and long processes. Due to the lack of effective task planning and decomposition mechanisms, a single model often results in complex operations and delayed responses when dealing with processes requiring cross-stage collaboration (such as the entire process from market monitoring to investment decision-making to compliance review), failing to meet the business's demand for immediacy.

[0004] 2. Illusions and Compliance Risks: The financial sector is highly specialized and subject to stringent regulations. Relying solely on an LLM model to generate investment advice or research report summaries often leads to "illusions" due to insufficient real-time data, a lack of professional knowledge base, or insufficient depth of reasoning, resulting in erroneous output. Furthermore, a single model lacks built-in self-checking and error-correction mechanisms, making it difficult to ensure that the output fully complies with stringent industry compliance requirements, thus posing significant compliance risks.

[0005] 3. Insufficient domain depth: General-purpose models often lack sufficient domain depth when dealing with securities-specific terminology, complex logic, and dynamic market data. For example, when generating in-depth investment research reports, a single model struggles to simultaneously ensure data accuracy, logical rigor, and innovative perspectives, resulting in inconsistent output quality.

[0006] 1.2 Current Status of Existing Multi-Agent Collaboration Techniques for Large Models To overcome the limitations of single models, academia and industry have begun to explore large model multi-agent systems (LLM-MAS). Current mainstream research focuses on using multiple LLM-driven agents to collaboratively solve complex problems, with collaboration paradigms mainly including centralized control, decentralized interaction, and hybrid architecture.

[0007] Centralized control paradigm: Represented by MetaGPT, this paradigm uses a central controller (such as a project manager agent) to uniformly schedule tasks and assign them to agents playing specific roles (such as product managers or engineers). This model performs well in structured tasks such as software development, but when faced with dynamic, non-standardized, and complex scenarios in securities business, the central controller faces enormous scheduling pressure and lacks flexibility.

[0008] Decentralized interaction paradigm: Represented by AutoGen, this allows agents to automatically negotiate and cooperate through natural language dialogue. While improving the autonomy of the system, in an open dialogue environment lacking clear constraints, agents are prone to getting stuck in ineffective loops or deviating from the task objective, resulting in low task completion efficiency and difficulty in ensuring the consistency and traceability of output results.

[0009] Hybrid architecture paradigms: Frameworks such as CAMEL attempt to combine the advantages of both, but in practical applications, existing hybrid architectures often lack refined modeling tailored to the characteristics of securities business. For example, the lack of explicit expression of task dependencies (such as DAG task graphs) leads to low collaboration efficiency and huge resource consumption among agents when processing massive amounts of information in parallel or conducting multiple rounds of investment research and deduction.

[0010] Furthermore, most existing multi-agent systems focus on collaboration mechanisms for general tasks, lacking specialized collaboration primitives for securities business scenarios. For example, they lack "annotation-based review" mechanisms to simulate the review process of human experts, "debate-based generation" mechanisms to improve decision-making quality through the collision of multiple viewpoints, and efficient "semantic context sharing" mechanisms to ensure the consistency of information among multiple agents.

[0011] 1.3 Core Challenges of Existing Technologies in Securities Business Applications Despite the theoretical advantages of multi-agent collaboration technology, its practical application in securities business still faces several core technological challenges. These issues directly restrict the efficiency and reliability of intelligent applications: Challenge 1: Lack of ability to decompose complex tasks and dynamically schedule them Securities business tasks typically consist of multiple tightly coupled subtasks (e.g., "acquiring market data" → "generating research report summaries" → "compliance checks"). Existing multi-agent systems lack a unified task graph generation and dynamic scheduling mechanism. Most systems rely on pre-set fixed processes or simple linear execution, failing to intelligently construct directed acyclic graphs (DAGs) to characterize task dependencies based on the dynamic characteristics of user requests. This results in the system's inability to flexibly adjust the execution order and parallel strategies of agents when facing sudden market events or personalized user needs, leading to low task execution efficiency and even logical deadlocks.

[0012] Challenge 2: Lack of domain knowledge-based automated verification and quality closed-loop In financial scenarios, the accuracy and compliance of output are paramount. Existing agent collaboration processes are mostly unidirectional, lacking effective mechanisms for "annotated review" and "debate-based generation." Specifically, when an agent generates intermediate results (such as investment advice or data summaries), there is a lack of dedicated review agents to provide rule-based compliance annotations and semantic-based logical verification. Furthermore, there is a lack of debate mechanisms among multiple agents to expose potential errors. This lack of a quality loop makes it difficult for the system to eliminate illusions and correct errors through multiple rounds of review and discussion, as human expert teams can, resulting in insufficient credibility of the final output and making it unsuitable for direct use in high-risk financial decisions.

[0013] Challenge 3: Context fragmentation and information silo effect Multi-agent collaboration requires a high degree of information sharing and contextual consistency. However, existing systems often rely on simple message passing for inter-agent communication, lacking a unified "semantic context management system." This results in each agent only accessing partial information when executing tasks, unable to access a globally shared knowledge base (such as historical Q&A, legal provisions, and real-time market data). This fragmentation of information not only leads to redundant queries and wasted computing power, but more seriously, it causes discrepancies in the understanding of the same task among different agents, creating information silos and undermining the integrity and coherence of collaboration. This makes it difficult to support complex securities business scenarios that require long-term contextual memory and multiple rounds of interaction. Summary of the Invention

[0014] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, system, device, processor and its computer-readable storage medium for large-scale multi-agent collaborative processing of securities business, which meets the requirements of high task processing efficiency, high accuracy and low development cost.

[0015] To achieve the above objectives, the present invention provides a method, system, apparatus, processor, and computer-readable storage medium for large-scale multi-agent collaborative processing of securities business models, as well as the following: This method for achieving multi-agent collaborative processing of large models in securities business is characterized by the following steps: (1) Receive the user's natural language business request, generate a directed acyclic graph (DAG) that describes the subtask dependencies, and decompose the business request into atomic subtasks that can be executed in parallel; (2) Role grouping is performed according to the subtask type, and different atomic subtasks are assigned to intelligent agent groups with corresponding functions. Dynamic scheduling and parallel processing of subtasks are realized based on a centralized scheduling strategy. (3) The intermediate outputs are reviewed in multiple rounds through the annotation-based review mechanism. For comprehensive decision-making tasks, a debate-based generation mechanism is triggered, and the intermediate results are output after cross-validation. (4) Through a globally shared semantic context management system, information synchronization and consistency maintenance among all intelligent agents are achieved, the execution results of all subtasks are aggregated, and the final business output that meets the requirements of the securities industry is generated.

[0016] Preferably, step (1) specifically includes the following steps: Generate a directed acyclic graph Where V is the set of atomic subtasks and E is the set of directed edges; define the task state. ,in Let be the set of subtasks that have been completed at time t. The set of executable subtasks is the set of all nodes whose prerequisites have been completed.

[0017] Preferably, step (2) specifically includes the following steps: Establish a mapping relationship between task nodes and intelligent agent groups. ,in The system is a collection of intelligent agents; a centralized scheduling strategy is introduced based on the multi-agent Markov decision process. The state space Includes task graph progress, completed nodes, and quality signals. Resource load ; where the action space This indicates the selection of subtasks and the execution agent.

[0018] Preferably, step (2) further includes the following steps: The intelligent agents are divided into information collection group, text parsing group, investment research and analysis group, content generation group, compliance review group, and output processing group. The intelligent agents within each group work closely together, and the intelligent agents between groups interact loosely through a message bus. The same subtask can be executed in parallel by multiple intelligent agents and cross-validated.

[0019] Preferably, the annotation-based review mechanism in step (3) is specifically as follows: The review agent reviews the intermediate results generated by the execution agent segment by segment, adds structured annotations, and feeds them back to the execution agent for correction. The review cycle can be repeated multiple times until the review is passed or the preset maximum number of retries is reached.

[0020] Preferably, the debate-based generation mechanism in step (3) is specifically as follows: Multiple agent instances with different policy labels are launched. Each agent generates an independent solution based on the same shared context. The agents exchange arguments and rebuttals through a message bus. After a preset number of rounds of debate, the judging agent outputs the optimal result by summarizing the advantages and disadvantages of each solution.

[0021] Preferably, the semantic context management system in step (4) uses shared memory to store key contexts, including the user's original request, task graph information, intermediate results of each stage, and external retrieval knowledge; and attaches a unique semantic label to each data item, and the agent obtains relevant information by subscribing to the corresponding label.

[0022] Preferably, the semantic context management system supports incremental context updates and version control, with each update generating a unique version number; sensitive information is tagged with access control labels, allowing only authorized agents to access it.

[0023] Preferably, the method further includes log auditing and end-to-end tracing steps, specifically including the following processes: Record all agent message interactions, tool calls, generated content, and execution time, and store the logs in a structured format; track the agent call chain for each task through end-to-end tracing to locate performance bottlenecks and abnormal behaviors.

[0024] Preferably, the method further includes a system self-optimization feedback loop step, specifically including the following processing steps: Regularly analyze the execution quality, response latency, and token consumption of each agent; optimize scheduling strategies, agent prompts, and task allocation rules using reinforcement learning based on the overall business utility function; and dynamically adjust based on historical data. Hyperparameters.

[0025] This system, designed for large-scale multi-agent collaborative processing in securities business, is characterized by the following features: The basic model and tools layer is used to provide underlying intelligent computing power, financial data retrieval, market information query and knowledge enhancement capabilities; The agent container and orchestration layer, connected to the basic model and tool layer, is used to realize dynamic task graph generation, multi-agent role grouping, annotation-based review, debate-based generation, and collaborative execution scheduling. The business adaptation layer, connected to the aforementioned agent container and orchestration layer, is used to encapsulate multi-agent collaboration capabilities into standardized securities business interfaces, enabling integration with existing business systems. The intelligent agent container and orchestration layer include: An agent container is used to encapsulate and run multiple agents with different role labels in a containerized manner. The message bus, employing a high-throughput message queue system, is used to transmit messages and events between intelligent agents; The agent scheduler, connected to the message bus and agent container, is used to execute centralized scheduling strategies according to the task graph and role grouping mapping, and to manage the lifecycle of agents and task orchestration. The semantic context management module, built on the model context protocol, is connected to the message bus and is used to maintain a globally shared session context.

[0026] Preferably, the basic model and tool layer, the intelligent agent container and orchestration layer, and the business adaptation layer constitute a three-layer composite mapping system, which is represented by the following formula: ; in, Based on the basic model and tools layer, For intelligent agent containers and orchestration layers, This is a business adaptation layer.

[0027] Preferably, the basic model and tool layer includes a large model base, a set of tool plugins, and a vector knowledge base. The large model base is based on a general financial model and is equipped with N scenario-based fine-tuning sub-models. The set of tool plugins encapsulates financial data retrieval, computing engine, and communication interface. The vector knowledge base stores research reports, legal provisions, historical Q&A, and market data.

[0028] Preferably, the agent scheduler uniformly manages the lifecycle and task orchestration of agents, achieving load balancing and automatic fault recovery; when an agent fails to execute, it automatically restarts the instance or assigns another agent to take over the task; and it has a built-in tracking subsystem to monitor the execution chain and resource usage of each agent in real time.

[0029] The main feature of this device for enabling large-scale multi-agent collaborative processing in securities business is that the device includes: A processor is configured to execute computer-executable instructions; The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the various steps of the above-described method for multi-agent collaborative processing of large-scale models for securities business.

[0030] The processor for implementing large-scale multi-agent collaborative processing in securities business is characterized by being configured to execute computer-executable instructions, which, when executed by the processor, implement the various steps of the aforementioned method for implementing large-scale multi-agent collaborative processing in securities business.

[0031] The computer-readable storage medium is characterized in that it stores a computer program that can be executed by a processor to implement the various steps of the above-described method for large-scale multi-agent collaborative processing of securities business.

[0032] The present invention employs a method, system, apparatus, processor, and computer-readable storage medium for large-scale multi-agent collaborative processing of securities business. By constructing a large-scale multi-agent collaborative architecture for securities business, based on core mechanisms such as task graph generation and decomposition, role grouping, annotation-based review, context sharing, and debate-based generation, and combined with a three-layer architecture design of basic model layer, agent orchestration layer, and business adaptation layer, it achieves high-precision, high-efficiency, and high-reliability intelligent services in solving the automation of complex securities business processes. Attached Figure Description

[0033] Figure 1 This is an architecture diagram of the system for implementing large-scale multi-agent collaborative processing of securities business according to the present invention.

[0034] Figure 2 This is an architecture diagram of an embodiment of the system for implementing large-scale multi-agent collaborative processing of securities business according to the present invention.

[0035] Figure 3 This is a schematic diagram of the business process of the method for implementing large-scale multi-agent collaborative processing of securities business according to the present invention.

[0036] Figure 4 This is a schematic diagram of a multi-agent Markov decision process model for the method of multi-agent collaborative processing of large models for securities business in this invention.

[0037] Figure 5 This is a schematic diagram of the multi-agent cooperative sequence of the method for large-scale multi-agent cooperative processing of securities business according to the present invention.

[0038] Figure 6This is a schematic diagram of the intelligent decomposition of a directed acyclic graph (DAG) for the method of multi-agent collaborative processing of large models for securities business in this invention.

[0039] Figure 7 This diagram illustrates the agent containerization encapsulation technology and lifecycle management mechanism of the method for realizing large-scale multi-agent collaborative processing in securities business according to the present invention. Detailed Implementation

[0040] To more clearly describe the technical content of the present invention, the following description is provided in conjunction with specific embodiments.

[0041] The present invention provides a method for multi-agent collaborative processing of large models in securities business, comprising the following steps: (1) Receive the user's natural language business request, generate a directed acyclic graph (DAG) that describes the subtask dependencies, and decompose the business request into atomic subtasks that can be executed in parallel; (2) Role grouping is performed according to the subtask type, and different atomic subtasks are assigned to intelligent agent groups with corresponding functions. Dynamic scheduling and parallel processing of subtasks are realized based on a centralized scheduling strategy. (3) The intermediate outputs are reviewed in multiple rounds through the annotation-based review mechanism. For comprehensive decision-making tasks, a debate-based generation mechanism is triggered, and the intermediate results are output after cross-validation. (4) Through a globally shared semantic context management system, information synchronization and consistency maintenance among all intelligent agents are achieved, the execution results of all subtasks are aggregated, and the final business output that meets the requirements of the securities industry is generated.

[0042] In a preferred embodiment of the present invention, step (1) specifically includes the following steps: Generate a directed acyclic graph Where V is the set of atomic subtasks and E is the set of directed edges; define the task state. ,in Let be the set of subtasks that have been completed at time t. The set of executable subtasks is the set of all nodes whose prerequisites have been completed.

[0043] In a preferred embodiment of the present invention, step (2) specifically includes the following steps: Establish a mapping relationship between task nodes and intelligent agent groups. ,in The system is a collection of intelligent agents; a centralized scheduling strategy is introduced based on the multi-agent Markov decision process. The state space Includes task graph progress, completed nodes, and quality signals. Resource load ; where the action space This indicates the selection of subtasks and the execution agent.

[0044] In a preferred embodiment of the present invention, step (2) further includes the following steps: The intelligent agents are divided into information collection group, text parsing group, investment research and analysis group, content generation group, compliance review group, and output processing group. The intelligent agents within each group work closely together, and the intelligent agents between groups interact loosely through a message bus. The same subtask can be executed in parallel by multiple intelligent agents and cross-validated.

[0045] As a preferred embodiment of the present invention, the annotation-based review mechanism in step (3) is specifically as follows: The review agent reviews the intermediate results generated by the execution agent segment by segment, adds structured annotations, and feeds them back to the execution agent for correction. The review cycle can be repeated multiple times until the review is passed or the preset maximum number of retries is reached.

[0046] As a preferred embodiment of the present invention, the debate-style generation mechanism in step (3) is specifically as follows: Multiple agent instances with different policy labels are launched. Each agent generates an independent solution based on the same shared context. The agents exchange arguments and rebuttals through a message bus. After a preset number of rounds of debate, the judging agent outputs the optimal result by summarizing the advantages and disadvantages of each solution.

[0047] As a preferred embodiment of the present invention, the semantic context management system in step (4) uses shared memory to store key contexts, including user original requests, task graph information, intermediate results of each stage, and external retrieval knowledge; and attaches a unique semantic label to each data item, and the agent obtains relevant information by subscribing to the corresponding label.

[0048] As a preferred embodiment of the present invention, the semantic context management system supports incremental context updates and version control, generating a unique version number with each update; sensitive information is tagged with permission control labels, allowing only authorized agents to access it.

[0049] In a preferred embodiment of the present invention, the method further includes the steps of log auditing and end-to-end tracing, specifically including the following processing steps: Record all agent message interactions, tool calls, generated content, and execution time, and store the logs in a structured format; track the agent call chain for each task through end-to-end tracing to locate performance bottlenecks and abnormal behaviors.

[0050] In a preferred embodiment of the present invention, the method further includes a system self-optimization feedback closed loop step, specifically including the following processing steps: Regularly analyze the execution quality, response latency, and token consumption of each agent; optimize scheduling strategies, agent prompts, and task allocation rules using reinforcement learning based on the overall business utility function; and dynamically adjust based on historical data. Hyperparameters.

[0051] The present invention relates to a system for large-scale multi-agent collaborative processing of securities business models, wherein the system comprises: The basic model and tools layer is used to provide underlying intelligent computing power, financial data retrieval, market information query and knowledge enhancement capabilities; The agent container and orchestration layer, connected to the basic model and tool layer, is used to realize dynamic task graph generation, multi-agent role grouping, annotation-based review, debate-based generation, and collaborative execution scheduling. The business adaptation layer, connected to the aforementioned agent container and orchestration layer, is used to encapsulate multi-agent collaboration capabilities into standardized securities business interfaces, enabling integration with existing business systems. The intelligent agent container and orchestration layer include: An agent container is used to encapsulate and run multiple agents with different role labels in a containerized manner. The message bus, employing a high-throughput message queue system, is used to transmit messages and events between intelligent agents; The agent scheduler, connected to the message bus and agent container, is used to execute centralized scheduling strategies according to the task graph and role grouping mapping, and to manage the lifecycle of agents and task orchestration. The semantic context management module, built on the model context protocol, is connected to the message bus and is used to maintain a globally shared session context.

[0052] In a preferred embodiment of the present invention, the basic model and tool layer, the intelligent agent container and orchestration layer, and the business adaptation layer constitute a three-layer composite mapping system, which is represented by the following formula: ; in, Based on the basic model and tools layer, For intelligent agent containers and orchestration layers, This is a business adaptation layer.

[0053] As a preferred embodiment of the present invention, the basic model and tool layer includes a large model base, a tool plugin set, and a vector knowledge base. The large model base is based on a general financial model and is equipped with N scenario-based fine-tuning sub-models. The tool plugin set encapsulates financial data retrieval, a computing engine, and communication interfaces. The vector knowledge base stores research reports, legal provisions, historical Q&A, and market data.

[0054] In a preferred embodiment of the present invention, the agent scheduler uniformly manages the lifecycle and task orchestration of agents, and realizes load balancing and automatic fault recovery; when an agent fails to execute, it automatically restarts the instance or assigns another agent to take over the task; it has a built-in tracking subsystem to monitor the execution link and resource usage of each agent in real time.

[0055] The present invention provides an apparatus for implementing large-scale multi-agent collaborative processing in securities business, wherein the apparatus comprises: A processor is configured to execute computer-executable instructions; The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the various steps of the above-described method for multi-agent collaborative processing of large-scale models for securities business.

[0056] The present invention discloses a processor for implementing large-scale multi-agent collaborative processing for securities business, wherein the processor is configured to execute computer-executable instructions, and when the computer-executable instructions are executed by the processor, the various steps of the above-described method for implementing large-scale multi-agent collaborative processing for securities business are implemented.

[0057] The computer-readable storage medium of the present invention stores a computer program thereon, which can be executed by a processor to implement the various steps of the above-described method for large-scale multi-agent collaborative processing of securities business.

[0058] This invention proposes a multi-agent collaborative architecture and method for large-scale securities business, aiming to solve the problems of process fragmentation, response lag, insufficient accuracy, and high compliance risks in existing single-agent models when handling complex securities business. The core lies in constructing a multi-agent collaborative processing system based on a unified collaborative mechanism model and a hierarchical system architecture. Through key mechanisms such as task graph-driven processing, role grouping, annotation-based review, context sharing, and debate-based generation, it achieves automated decomposition, specialized division of labor, and highly reliable execution of complex tasks in securities business.

[0059] Overall architecture as follows Figure 1 As shown, the multi-agent collaborative architecture consists of a three-layer composite mapping system: a basic model and tool layer, an agent container and orchestration layer, and a business adaptation layer. This can be abstractly represented as: in, It establishes a basic model and tool layer to realize the mapping from "text / structured input + tool calls" to "semantic output"; It serves as a container and orchestration layer for intelligent agents, enabling the mapping from "task state" to "scheduling decisions and multi-agent collaborative trajectories"; For the business adaptation layer, it realizes the mapping from "collaboration trajectory" to "business-consumable results and indicators".

[0060] The system is designed for a given business distribution. Maximize the overall business utility function : in, : Scheduling and cooperation strategy parameters; Basic model and tool selection parameters; Business-side strategies and template parameters.

[0061] The components and their implementation methods are described in detail below with reference to the accompanying drawings.

[0062] Module 1: Collaboration Mechanism Modeling Unit This module establishes the basic mathematical model and operating mechanism for multi-agent collaborative processing, and includes five core sub-mechanisms.

[0063] 1. Task graph generation and decomposition mechanism To address the complex, multi-step nature of securities business tasks, a directed acyclic graph (DAG) is used for structured representation. A business request is modeled as a task graph. ,in: For a set of subtasks, each node This corresponds to an atomic subtask, such as "acquiring market data", "generating research report summaries", or "compliance checks". Let be a set of directed edges. express Depends on The output of .

[0064] Introducing a task graph generator , will the user's natural language request Mapped to a task graph: in These are the parameters for structured programming based on a large model. In actual implementation, By using "mind chain + tool call template" to guide the large language model to gradually produce node sets and dependencies, and then organize them into a DAG structure.

[0065] Define task status ,in For a moment The set of completed nodes. The set of executable subtasks is: This means that all prerequisite nodes have been completed. This mechanism makes task decomposition and execution naturally correspond to a multi-stage decision-making process, providing a mathematical basis for subsequent agent scheduling.

[0066] 2. Role grouping mechanism Based on the task graph, tasks are divided, and agents are organized into several functional groups, each responsible for a specific sub-task or phase. Assume there exists a set of agents in the system. , No. Each agent has a strategy Tool Capability Set And role labels (such as "information analysis", "investment advisory decision-making", "compliance review", etc.). Among them, Agent in this invention refers to intelligent agent.

[0067] Define the mapping from task nodes to groups of agents. ,in To undertake sub-tasks A group of intelligent agents. For example, information nodes. Mapped to "Information Processing Group": .

[0068] When a task is executed, the scheduler starts from... Select the node to be executed and from Select a specific Agent Execution, forming a joint action This mechanism enables close collaboration within groups and loosely coupled interaction between groups, improving system scalability.

[0069] 3. Annotation-based review mechanism To improve the accuracy and reliability of the output of multi-agent systems, an annotation-based review mechanism is introduced: the review agent reviews the intermediate outputs of other agents and adds annotations, providing suggestions for modification or highlighting problems. For example, in the scenario of automatic research report generation, the review agent adds annotations such as "Data source is unclear here" or "Conclusion lacks risk warnings" to the draft of the writing agent, and the writing agent makes corrections after feedback.

[0070] This mechanism is applicable to key financial processes such as compliance review and risk assessment, ensuring that outputs undergo multiple rounds of checks, thus enhancing credibility. The process simulates human editing workflows, utilizing the language understanding capabilities of intelligent agents to identify and correct logical errors, inconsistencies, or violations.

[0071] 4. Context sharing mechanism By introducing a semantic context management system, efficient information sharing among intelligent agents can be achieved, avoiding information silos. Specifically, shared memory or a knowledge base is used to store key contexts, including user information, intermediate results, and external knowledge.

[0072] Before executing a task, each agent retrieves relevant data from the shared context. For example, the information parsing agent writes the summary into the shared context for the investment strategy agent to access. To address information filtering and access control, a semantic tagging mechanism is designed to push only relevant content to the required agents and restrict access to sensitive information. This mechanism reduces redundant queries and ensures collaborative consistency.

[0073] 5. Debate-based generation mechanism Drawing inspiration from the truth-discovery process in human debates, a debate-based generation mechanism is incorporated: two or more intelligent agents engage in dialogue and debate on a certain issue, presenting arguments and rebuttals, with the system or a third party ultimately evaluating the agents and selecting the best output.

[0074] In the securities field, this can be used for investment decision support: multiple "virtual investment advisors" propose and debate strategies, revealing their advantages and disadvantages to assist human decision-making. The implementation can be configured with debate rounds. Evaluation criteria. Through multi-agent debate, the system iteratively optimizes its answers in a human-team-like manner, improving robustness and credibility.

[0075] 6. Mathematical Modeling of Collaboration Mechanisms The entire multi-agent collaborative process is modeled as a multi-agent Markov decision process (MMDP) with centralized training and distributed execution. Among them, the state space : It includes task graph progress, completed nodes, and quality signals. (such as BLEU / ROUGE or predictive scoring), resource load Action space : This indicates the selection of subtasks and the execution agent; state transition kernel The agent generates a new state; reward function. Overall quality, delay, and cost; Discount factor.

[0076] Introducing a centralized scheduling strategy The goal is to maximize overall business efficiency. Single-step return is defined as: in, Scoring of key output quality (such as accuracy, summary score, compliance score); For response delay; The cost of token consumption or computing power; To balance hyperparameters.

[0077] 7. Unified collaboration process Generate a task graph based on the user's task and break it down into sub-tasks; Assign a dedicated intelligent agent group to each type of subtask, and exchange information by sharing context; The process is executed sequentially within the group, with key results reviewed and annotated by a review agent. For problems requiring comprehensive decision-making, multi-agent debates are triggered, and cross-validation yields high-quality answers.

[0078] Module Two: System Architecture Design Unit A three-layer architecture is proposed to implement the above collaboration mechanism: basic model and tool layer, intelligent agent container and orchestration layer, and business adaptation layer.

[0079] 1. Basic Model and Tool Layer Provides core intelligent capabilities, including: Large Model Base: It adopts a "1+N" model system, that is, a general financial model as the base, combined with N scenario-based fine-tuning sub-models, integrating the advantages of open source and self-developed technology to support diversified businesses; Tool plugins: Encapsulate interfaces for financial data retrieval, market data query, computing engine, email / SMS sending, etc., and provide them to intelligent agents for invocation through the LangChain class mechanism; Vector database / knowledge base: Stores research reports, regulations, historical Q&A, and other knowledge, and supports retrieval-enhanced generation (RAG).

[0080] This layer is accessible to the upper layer via a standard API.

[0081] 2. Intelligent Agent Containers and Orchestration Layers The core of the architecture is responsible for running and managing multiple agents and coordinating their interactions.

[0082] Containerized encapsulation of intelligent agents: Each intelligent agent is encapsulated in a containerized manner, possessing an independent operating environment and resource control, and is registered to the scheduling and communication framework; Message bus: A high-throughput message queue is used as the bus to transmit messages and events between intelligent agents. A publish / subscribe mechanism is adopted to achieve decoupling, broadcasting and multicasting, support context sharing, and record message flow for audit tracing; Agent Scheduling and Tracking: The scheduler uniformly manages the lifecycle and task orchestration of agents. It allocates or starts corresponding agent containers based on the task graph and assigns tasks via the message bus. It is responsible for creation, destruction, load balancing, and maintaining the agent state machine (idle, running, error), restarting or replacing instances when necessary. The tracking subsystem monitors the execution chain, records the call chain and latency, and identifies bottlenecks. Session and Context Management: Maintaining multi-agent sessions and isolating different user contexts. Each session is treated as a logical workflow, containing all relevant agents and message trajectories. Communication is aggregated using a unique session ID to prevent information leakage. The context management module interfaces with a knowledge base to inject and update context before and after processing. Log Auditing and Security: Records all critical interaction logs (messages, tool calls, generated content), stores them in a structured manner, and supports post-event traceability and compliance checks. Integrates with the security module to achieve sensitive information shielding and interception of sensitive operations.

[0083] This layer acts as the "operating system" of a multi-agent system, enabling "orchestration-driven collaboration" and allowing the upper layers to focus on business logic.

[0084] 3. Business Adaptation Layer By combining capabilities with specific securities business, we provide terminal service interfaces.

[0085] Business module API encapsulation: General multi-agent collaboration processes are encapsulated as APIs for use by mobile apps, web applications, etc., enabling integration with existing systems. Different business functions define their own agent orchestration processes while sharing a unified architecture. Strategy Configuration and Template Customization: Configure agent behavior according to business needs. For example, customize system prompts in intelligent question answering, limiting response style and compliant language; configure review rules in compliance review, triggering manual review when violations are found; Interface and Frontend Interaction: Provide REST APIs or message interfaces to handle frontend input and output. For interactive applications, implement dialogue state management and multi-turn context connection, and package the results after calling the lower-level agent and return them to the user.

[0086] The three-tier architecture features a clear division of labor, distinct layers, and easy scalability: the bottom layer provides intelligence and tools, the middle layer implements general collaboration mechanisms, and the top layer enables business customization. Each layer connects through standard interfaces, reducing complexity.

[0087] Module 3: Implementation of Typical Application Scenarios This has been proven in practice in large-scale application architectures, and the following two cases illustrate this.

[0088] 1. Information Processing Intelligent Agent Production Line Faced with a massive amount of financial information and research reports, we build an AIGC content factory pipeline to achieve automated processing and personalized distribution: Information gathering intelligent agent: scrapes news, announcements, research reports, etc.; Content parsing agent: extracts key information, performs sentiment analysis, and tags; Abstract Generating Intelligent Agents: Key Points and Brief Commentary; Quality review agent: Annotates and reviews to ensure accuracy and objectivity; Push scheduling intelligent agent: Generates personalized information streams based on users' holdings and preferences.

[0089] The process relies on a "1+N" architecture, with each intelligent agent working collaboratively to achieve real-time, contextualized, and precise information distribution.

[0090] 2. Intelligent Question Answering and Investment Advisory Agent To address the massive influx of inquiries and personalized investment advisory needs from securities firms, a conversational intelligent service front-end has been deployed: Intelligent question-answering agent: Fine-tuned with business knowledge, it understands customers' natural language questions; The search agent searches for relevant content in the regulatory database and product manuals to assist in answering questions; Multi-turn dialogue management agent: maintaining context and enabling contextual responses; Investment advisory decision-making AI: Generates investment advice based on user risk preferences, market conditions, etc., and then submits it to a question-and-answer AI for response.

[0091] Module 4: Alternative Technical Solutions 1. Heterogeneous model support scheme It supports the coexistence of open-source, closed-source APIs, self-developed models, and multimodal models. The intelligent agent framework calls any backend model through a unified interface and intelligently selects the appropriate model based on the task. A cross-model translation agent is introduced to ensure that the outputs of different models can be fused, achieving model independence.

[0092] 2. Intelligent Agent Evolution and Scheduling Scheme As the scale of the intelligent agents increases, a reinforcement learning mechanism is introduced to autonomously optimize the agent combination and processes based on historical performance data, enabling the collaborative strategy to evolve on its own. Elastic scheduling is implemented at runtime: the number of agent instances and GPU resource allocation are dynamically adjusted according to load and task type to avoid resource waste or overload.

[0093] 3. Self-monitoring and evaluation plan It incorporates built-in self-evaluation and safety constraint mechanisms. A supervised agent is configured to continuously evaluate output quality, forming a self-supervised feedback loop. For example, the meta-agent periodically checks the consistency between question-and-answer results and the fact base, or monitors whether the dialogue deviates from its intent and corrects it immediately, significantly reducing the cost of manual annotation.

[0094] In a specific embodiment of the present invention, the system processing flow mainly includes: user request reception and task graph generation, multi-agent task scheduling based on a cooperation mechanism, collaborative execution under a three-layer architecture, and final result aggregation and output. The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0095] Step 1: User Request Reception and Task Graph Generation After the system starts, the business adaptation layer receives natural language requests (x) from users on the mobile app or web frontend via REST API or messaging interface. For example, a user might enter, "Please generate an investment research report on the new energy industry, focusing on the impact of recent policy changes on leading companies."

[0096] Upon receiving the request, the system invokes the task planning module. This generator, based on the Large Language Model (LLM), guides the model to gradually produce a set of subtask nodes and dependencies through a "thinking chain + tool call template" approach, ultimately organizing them into a Directed Acyclic Graph (DAG) structure. For the above investment research report generation request, the generated task graph may contain the following nodes: To obtain the latest policy information for the new energy industry, To obtain market performance data of leading companies, To analyze policy texts and extract key impact points, To analyze corporate financial data and stock price trends, To generate a research report summary, For compliance audit, To format it into the final research report.

[0097] edge set Define the dependencies between subtasks, for example Depends on The output, Depends on The output, Depends on and The output, Depends on The output, Depends on The output of .

[0098] System initialization task status ,in Initially, the set is empty, and the set of executable subtasks is calculated. This refers to nodes where all prerequisite dependencies have been completed.

[0099] Step 2: Multi-agent scheduling and context initialization The agent scheduler, based on the task graph And character grouping mapping Each subtask is assigned a corresponding agent group. The system predefines several agent types and their role labels, for example: Information gathering agent (role tag: data collection), policy analysis agent (role tag: text analysis), investment research analysis agent (role tag: strategy generation), research report writing agent (role tag: content generation), compliance review agent (role tag: compliance review), and formatting agent (role tag: output processing).

[0100] Scheduler from The system selects the currently executable subtask node and chooses a specific agent instance from the corresponding agent group for execution. Simultaneously, the system initializes a shared context through the semantic context management system. This context, built on the Model Context Protocol (MCP), contains the user's original request, task graph information, and intermediate results generated by each agent. Each session uses a unique session ID to aggregate communications in the message bus, ensuring information isolation within the same session.

[0101] Step 3: Multi-agent cooperative execution based on message bus The system adopts a message-driven architecture, using a high-throughput message queue system as a bus to transmit messages and events between agents. The following example, an investment research report generation task, illustrates the collaborative execution process of the agents: 1. Information Collection and Analysis Stage Scheduler from Select node (To obtain the latest policy information for the new energy industry) and (Acquire market performance data of leading enterprises) and assign them to the data collection agent instances in the information collection agent group.

[0102] The data collection agent calls tool plugins through the basic model and tool layer: For It calls the search engine API to retrieve recent policy documents; for It calls the market data query API to retrieve stock price data and financial reports of leading companies. The collected results are published via the message bus and written to the shared context.

[0103] node and Update after completion ,at this time Updated to include and The scheduler will Assigned to the policy analysis agent Assigned to investment research and analysis intelligent agents.

[0104] The policy parsing agent reads from the shared context The collected policy texts are used for key information extraction and impact analysis using LLM, and the analysis results are written back to the shared context. The investment research and analysis agent reads from the shared context. Collect market data, conduct financial analysis and trend assessment, and write the analysis results back to the shared context.

[0105] 2. Research report generation and annotation-based review stage and After completion, Updated to include (Generate research report summary). The scheduler will... Assigned to the intelligent agent responsible for writing research reports.

[0106] The research report writing agent reads policy analysis and investment research results from the shared context and generates a draft report based on a pre-set research report template. During the generation process, the agent follows the system prompts configured in the business adaptation layer for the securities field to ensure that the response style and compliant terminology meet the requirements. The generated draft is then written back to the shared context.

[0107] At this point, the scheduler triggers the annotation-based review mechanism, which will... (Compliance review) is assigned to the compliance review agent. As the review agent, the compliance review agent reviews and annotates the draft research report, for example: "The source of the data cited in the policy impact analysis section is unclear", "The investment recommendation section lacks risk warnings", and "The conclusions are inconsistent with the data analysis logic".

[0108] The compliance review agent feeds back the annotations to the research report writing agent via the message bus and updates the shared context. The research report writing agent then revises the report based on the annotations, and the revised draft is resubmitted for compliance review. This review cycle can run multiple times until the compliance review is passed or the maximum number of retries is reached. This mechanism is similar to the editing process in human writing, utilizing the agent's own language understanding capabilities to identify and correct logical errors, inconsistencies, or violations.

[0109] 3. Debate-based generation (if comprehensive decision-making is required) If a user request involves a comprehensive decision-making issue, such as "comparing the investment value of Company A and Company B and providing recommendations," the system can trigger a debate-style generation mechanism. The scheduler launches multiple investment advisory decision-making agent instances (such as value investing agents, growth investing agents, and risk-averse agents). Each agent generates investment recommendations based on the same data in a shared context, according to its own strategy.

[0110] Each agent engages in multiple rounds of dialogue and debate regarding its own proposals, presenting arguments and rebuttals to each other via a message bus. For example: Agent A: "Based on P / E valuation, Company A's current valuation is lower than the industry average; a buy recommendation is given." Agent B: "However, Company A's revenue growth has slowed recently, its cash flow is negative, and the risk is high." Agent C: "Although Company B has a high valuation, its R&D investment continues to grow, and its long-term growth potential is better."

[0111] After the debate reaches a preset number of rounds N, the system's pre-defined evaluation rules or a third-party evaluation agent comprehensively considers the advantages and disadvantages of each solution and outputs the best answer. Through multi-agent debate, the system can iteratively optimize the answer in a human-like team discussion manner, improving credibility and robustness.

[0112] Step 4: Ensuring Context Sharing and Information Consistency Throughout the execution process, all agents access and update shared information through the semantic context management system. This system is implemented based on the MCP protocol and includes the following key technologies: Shared memory storage: A memory database such as Redis is used to store the key context accessible to all agents, including user-provided information, intermediate results at each stage, and externally retrieved knowledge. Each agent, when executing a task, first retrieves the task-related data from the shared context, processes it, and then writes the results back.

[0113] Semantic Tagging Mechanism: To address information filtering and access control issues, the system attaches semantic tags (such as "policy information," "financial data," "draft," and "review comments") to each data item. Agents subscribe to tag types of interest, and the message bus only pushes content with relevant tags to the necessary agents. Sensitive information (such as user identity information) is tagged with access control labels, allowing only authorized roles such as compliance agents to access it.

[0114] Incremental Updates and Version Control: The context management module supports incremental updates, avoiding the overhead of full data transfer. It also maintains a version number for each update to ensure data consistency for the agent and prevent conflicts caused by concurrent modifications.

[0115] For example, after the information analysis agent completes the market news summary, it writes the result to the shared context in the format "{tag:'news summary', version:v1, content:...}". Subsequently, the investment strategy agent subscribes to the "news summary" tag to obtain the latest version of the data at the start of the task, eliminating the need to repeatedly query the original information.

[0116] Step 5: Business Adaptation and Result Output Once the task is completed and all child nodes are finished, the business adaptation layer aggregates the collaboration trajectory into the final business-consumable result. Specific implementations include: Result aggregation: The formatting agent reads the full text of the final research report, compliance review records, and debate process summaries from the shared context and encapsulates them in the format required by the business (such as PDF, HTML, or JSON).

[0117] Business rule injection: Based on business scenario requirements, the adaptation layer executes customized strategies before outputting results. For example, in intelligent question-answering scenarios, the dialogue question-answering agent transforms the investment advice generated by the investment advisory decision-making agent into conversational, conversational responses; in institutional client scenarios, the adaptation layer adds a detailed audit evidence chain, including the source page number, processing agent ID, timestamp, etc., for each piece of data.

[0118] Interface Response: The final result is returned to the caller via a REST API or messaging interface. For user interaction applications, the adaptation layer is also responsible for connecting multi-turn dialogue contexts, storing the current response in the session context for use in subsequent turns.

[0119] Step 6: Log Auditing and End-to-End Tracing The system records all key interactions to the log system, including: The message content and timestamp received by the intelligent agent; The tools invoked and the results returned; The generated intermediate and final content; The execution time and state transitions of each intelligent agent.

[0120] Logs are stored in a structured format and analyzed periodically by the audit module to check for abnormal behavior or unauthorized output. Meanwhile, the tracking subsystem enables end-to-end tracing, recording the agent call chain and execution time for each task to help pinpoint performance bottlenecks or abnormal agent behavior. For example, if the delay in generating a research report exceeds a threshold, the tracking system can quickly determine whether the delay is due to slow processing by the policy analysis agent or multiple rounds of compliance review.

[0121] Step 7: System Optimization and Feedback Loop This invention supports a system optimization mechanism based on historical performance data. The system periodically analyzes the execution quality of each agent. (e.g., answer accuracy rate, compliance score), response latency and Token consumption And calculate the overall business utility: Single-step return hyperparameters Adjust dynamically according to business needs.

[0122] Scheduling strategy based on utility function Optimization can be achieved through reinforcement learning. For example, if excessive latency is found due to frequent multi-round loops in the compliance review process, the pass threshold for annotation-based reviews can be adjusted, or more explicit constraints can be added to the prompts. If excessive token consumption is found for a particular agent, a lightweight model can be used to replace or the prompts can be optimized to reduce useless dialogue. Through this iterative optimization, the system continuously improves itself, achieving the optimal balance between quality, latency, and cost.

[0123] This invention provides an intelligent processing method and system for securities business based on multi-agent collaboration and dynamic directed acyclic graphs (DAGs). It aims to solve core technical problems currently faced in the securities field when applying large models, such as rigid processes of single agents, difficulties in decomposing complex and long-link tasks, lack of automated verification loops for domain knowledge, and fragmented context among multiple agents. This system achieves automated and highly reliable processing of complex scenarios in securities business, from market monitoring and research report generation to compliance review, through deep integration of automatic task graph generation technology, annotation-based review and debate-based generation mechanisms, and a unified semantic context management system.

[0124] Specifically, this invention first employs a dynamic task graph generation mechanism based on user intent. This mechanism intelligently constructs a directed acyclic graph (DAG) depicting task dependencies according to the real-time needs of securities business, replacing the traditional scheduling mode of preset fixed processes or simple linear execution. This mechanism effectively solves the problems of logical deadlock and response lag caused by the system's inability to flexibly adjust the execution order and parallel strategies of agents when facing sudden market events or personalized needs, achieving dynamic scheduling and efficient execution of complex tasks. Simultaneously, this invention constructs a dual quality verification closed loop including "annotation-based review" and "debate-based generation." Addressing the extremely high accuracy and compliance requirements of financial scenarios, the system introduces a dedicated review agent to perform rule-based compliance annotations and semantic-based logical verification on intermediate results, and exposes potential errors through the collision of viewpoints among multiple agents. This mechanism simulates the review process of a human expert team, significantly eliminating the "illusion" phenomenon of large models and solving the problems of insufficient self-checking and error correction mechanisms and insufficient credibility of output results in existing unidirectional execution processes, ensuring that the final output meets strict industry regulatory standards.

[0125] Furthermore, this invention introduces a unified semantic context management system, breaking down information silos in multi-agent collaboration. By establishing a globally shared knowledge base (covering historical Q&A, legal provisions, real-time market data, etc.), efficient information sharing and context consistency maintenance among agents are achieved. This avoids misunderstanding biases and redundant queries caused by partial information acquisition, ensuring the integrity and coherence of collaboration when handling complex securities transactions requiring long contextual memory and multi-round interactions, significantly improving the system's reasoning depth and decision-making quality. Through the aforementioned dynamic orchestration, closed-loop verification, and context fusion throughout the entire process, this invention provides securities companies with a highly flexible, reliable, and domain-deep intelligent solution, effectively promoting the digital transformation of securities business from "single-point intelligence" to "swarm intelligence."

[0126] The specific technical effects of this invention are as follows: 1. This invention solves the problem of automatic decomposition and collaborative execution of complex securities business processes, significantly improving task processing efficiency: It employs a task graph generation and decomposition mechanism based on a Directed Acyclic Graph (DAG), which intelligently maps user natural language requests into a structured set of sub-task nodes and dependencies. This decomposes complex financial tasks into atomic sub-tasks that can be processed in parallel by different intelligent agents. Furthermore, a role-based grouping mechanism organizes intelligent agents into several functional groups, such as an information processing group responsible for information collection and processing, an investment research group responsible for strategy generation, and a compliance group responsible for review and approval. This achieves close collaboration among agents within a group and loosely coupled interaction between groups via a message bus. This "high cohesion, low coupling" design effectively improves system scalability; adding new business processes only requires adding the corresponding intelligent agent group without affecting other parts. In practical applications, in the scenario of information processing intelligent agent pipeline, the pipeline collaboration of information collection intelligent agent, content parsing intelligent agent, summary generation intelligent agent, quality review intelligent agent and push scheduling intelligent agent enables real-time, scenario-based and precise support for information. Users have changed from passively browsing massive amounts of information to the system actively pushing customized information streams, which greatly reduces information overload and significantly improves information processing efficiency.

[0127] 2. Reduces the "illusion" risk of a single large model and significantly improves the accuracy and compliance of output results: This invention introduces an annotation-based review mechanism and a debate-based generation mechanism, constructing a closed-loop process of multi-round quality control and cross-validation. The annotation-based review mechanism involves a specific review agent reviewing and annotating the intermediate outputs of other agents, providing suggestions for modification or highlighting problems. For example, in the scenario of automatically generating research reports, the review agent adds annotations such as "data source unclear" or "conclusions lack risk warnings" to the draft report generated by the writing agent and provides feedback for correction, forming an editing process similar to human writing. The debate-based generation mechanism allows multiple agents to engage in dialogue and debate on a certain issue or decision, presenting arguments and rebuttals to each other, with the system or a third party judging the agents and selecting the best answer. In investment decision support scenarios, this mechanism allows multiple "virtual investment advisors" to provide strategies and debate them, revealing the advantages and disadvantages of different strategies for human reference in decision-making. Through multi-agent debate, the system can iteratively optimize answers in a human-like team discussion manner, effectively exposing fallacies and improving the reliability of conclusions. In compliance review scenarios, the annotation-based review mechanism can be used to annotate potential compliance issues, mark uncertain conclusions, and ensure that the final output has undergone multiple rounds of review to meet the stringent requirements of securities business for the credibility and compliance of the output results.

[0128] 3. Optimizes communication overhead and computing cost in multi-agent collaboration, achieving a dynamic balance between quality, latency, and cost: This invention introduces a centralized scheduling strategy to model the multi-agent collaboration process as a multi-agent Markov decision process (MMDP), and defines a reward function that takes into account factors such as service quality, latency, and computing cost. in To rate the quality, In response to latency, The cost of token consumption or computing power is determined by hyperparameters. A dynamic balance of factors is struck. Based on this, a semantic context management system is introduced through a context-sharing mechanism. This system uses shared memory or a knowledge base to store key contexts accessible to all agents. Each agent retrieves task-related data from the shared context before executing a task, avoiding redundant queries. Simultaneously, a semantic tagging mechanism is designed to push relevant content only to the agents that need it, ensuring that sensitive information is only accessible to compliant agents. This mechanism reduces redundant communication and repetitive computation while ensuring collaborative consistency. It keeps the additional token consumption and latency associated with collaboration within an acceptable range while maintaining improved quality, achieving optimal cost-effectiveness.

[0129] 4. Supports large-scale concurrency and dynamic resource scheduling, achieving horizontal scalability: This invention adopts a three-layer architecture design. Each agent is encapsulated using containerization technology at the agent container and orchestration layers, giving it an independent operating environment and resource control. All agent containers are registered under the scheduling and communication framework. The message bus uses a high-throughput message queue system to transmit messages and events between agents, ensuring decoupling between components. Agents do not need to directly call each other; they only need to subscribe to / publish relevant topic messages. The agent scheduler uniformly manages the lifecycle and task orchestration of agents, responsible for agent creation, destruction, load balancing, etc., to fully utilize computing resources. The tracking subsystem monitors the execution chain of each agent in real time, recording the agent call chain and time consumption for each task, helping to locate performance bottlenecks or abnormal agent behavior. The scheduler maintains the agent state machine, tracks agent states, and restarts or replaces agent instances when necessary to ensure successful task completion. Through simulated concurrency testing, by gradually increasing the number of concurrent sessions or the number of agent instances started, the system performance exhibits near-linear scalability. With linear increases in resources, throughput also increases linearly, effectively supporting the large-scale concurrency requirements of securities business.

[0130] 5. This invention achieves modular reuse of agent capabilities, significantly reducing development costs for new business scenarios: Through a business adaptation layer, the general multi-agent collaboration process is encapsulated into business module APIs for various business systems to call. Different businesses define their own agent orchestration processes, but all run on a unified architecture. The business adaptation layer configures strategies and customizes templates for agent behavior based on actual business needs. For example, in the intelligent question-answering scenario, system prompts in the securities field are customized for the question-answering agent, limiting its response style and compliant language; in the compliance review scenario, the review agent is configured to follow a list of regulatory rules, and special processes such as escalation to manual review are adopted when prohibited content is involved. By statistically analyzing the development workload required for new business scenario access and the reuse rate of existing agent components, this invention achieves a significant improvement in development reusability. New feature development extensively calls existing agent services and reuses existing dialogue processes, fully demonstrating the advantages of modular reuse and significantly reducing the development and maintenance costs of new scenario access.

[0131] 6. Supporting the intelligent transformation of the entire securities business process and achieving significant business results: This invention has deployed multi-agent collaboration modules in key business areas such as information, investment advisory, risk control, and investment research, forming an "AI-driven" business support system. The 1+N architecture provides a unified and powerful large model foundation and dedicated sub-model support for each agent. Multi-agent orchestration allows AI capabilities to permeate every business process, solidifying expert experience into reusable agents. Through message bus linkage, the agent cluster collaboratively constitutes a new infrastructure for the digital transformation of securities firms. At the business indicator level, customer satisfaction scores and usage frequency in the intelligent question-and-answer scenario have significantly improved; customer conversion rates and asset growth rates have seen positive changes after the investment advisory module was launched; and the reduction in manpower due to the introduction of AI is considerable. The improvements in user experience and business indicators directly prove the effectiveness of the architecture of this invention.

[0132] For the specific implementation scheme of this embodiment, please refer to the relevant descriptions in the above embodiments, which will not be repeated here.

[0133] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0134] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0135] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0136] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0137] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The corresponding program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiments.

[0138] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0139] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0140] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0141] The present invention employs a method, system, apparatus, processor, and computer-readable storage medium for large-scale multi-agent collaborative processing of securities business. By constructing a large-scale multi-agent collaborative architecture for securities business, based on core mechanisms such as task graph generation and decomposition, role grouping, annotation-based review, context sharing, and debate-based generation, and combined with a three-layer architecture design of basic model layer, agent orchestration layer, and business adaptation layer, it achieves high-precision, high-efficiency, and high-reliability intelligent services in solving the automation of complex securities business processes.

[0142] In this specification, the invention has been described with reference to specific embodiments thereof. However, it will be apparent that various modifications and variations can be made without departing from the spirit and scope of the invention. Therefore, the specification and drawings should be considered illustrative rather than restrictive.

Claims

1. A method for implementing multi-agent collaborative processing of large models in securities business, characterized in that, The method includes the following steps: (1) Receive the user's natural language business request, generate a directed acyclic graph (DAG) that describes the subtask dependencies, and decompose the business request into atomic subtasks that can be executed in parallel; (2) Role grouping is performed according to the subtask type, and different atomic subtasks are assigned to intelligent agent groups with corresponding functions. Dynamic scheduling and parallel processing of subtasks are realized based on a centralized scheduling strategy. (3) The intermediate outputs are reviewed in multiple rounds through the annotation-based review mechanism. For comprehensive decision-making tasks, a debate-based generation mechanism is triggered, and the intermediate results are output after cross-validation. (4) Through a globally shared semantic context management system, information synchronization and consistency maintenance among all intelligent agents are achieved, the execution results of all subtasks are aggregated, and the final business output that meets the requirements of the securities industry is generated.

2. The method for multi-agent collaborative processing of large models for securities business as described in claim 1, characterized in that, Step (1) specifically includes the following steps: Generate a directed acyclic graph Where V is the set of atomic subtasks and E is the set of directed edges; define the task state. ,in Let be the set of subtasks that have been completed at time t. The set of executable subtasks is the set of all nodes whose prerequisites have been completed.

3. The method for multi-agent collaborative processing of large models for securities business as described in claim 1, characterized in that, Step (2) specifically includes the following steps: Establish a mapping relationship between task nodes and intelligent agent groups. ,in The system is a collection of intelligent agents; a centralized scheduling strategy is introduced based on the multi-agent Markov decision process. The state space Includes task graph progress, completed nodes, and quality signals. Resource load ; where the action space This indicates the selection of subtasks and the execution agent.

4. The method for multi-agent collaborative processing of large models for securities business as described in claim 1, characterized in that, Step (2) further includes the following steps: The intelligent agents are divided into information collection group, text parsing group, investment research and analysis group, content generation group, compliance review group, and output processing group. The intelligent agents within each group work closely together, and the intelligent agents between groups interact loosely through a message bus. The same subtask can be executed in parallel by multiple intelligent agents and cross-validated.

5. The method for multi-agent collaborative processing of large models for securities business as described in claim 1, characterized in that, The annotation-based review mechanism in step (3) is specifically as follows: The review agent reviews the intermediate results generated by the execution agent segment by segment, adds structured annotations, and feeds them back to the execution agent for correction. The review cycle can be repeated multiple times until the review is passed or the preset maximum number of retries is reached.

6. The method for multi-agent collaborative processing of large models for securities business as described in claim 1, characterized in that, The debate-based generation mechanism in step (3) is specifically as follows: Multiple agent instances with different policy labels are launched. Each agent generates an independent solution based on the same shared context. The agents exchange arguments and rebuttals through a message bus. After a preset number of rounds of debate, the judging agent outputs the optimal result by summarizing the advantages and disadvantages of each solution.

7. The method for multi-agent collaborative processing of large models for securities business as described in claim 1, characterized in that, The semantic context management system in step (4) uses shared memory to store key contexts, including user original requests, task graph information, intermediate results of each stage, and external retrieval knowledge; it attaches a unique semantic label to each data item, and the agent obtains relevant information by subscribing to the corresponding label.

8. The method for multi-agent collaborative processing of large models for securities business as described in claim 7, characterized in that, The semantic context management system supports incremental context updates and version control, generating a unique version number with each update; sensitive information is tagged with access control labels, allowing only authorized agents to access it.

9. The method for multi-agent collaborative processing of large models for securities business as described in claim 1, characterized in that, The method also includes log auditing and end-to-end tracing steps, specifically including the following processes: Record all agent message interactions, tool calls, generated content, and execution time, and store the logs in a structured format; track the agent call chain of each task through full-link tracing to locate performance bottlenecks and abnormal behaviors.

10. The method for multi-agent collaborative processing of large models for securities business as described in claim 1, characterized in that, The method also includes a system self-optimization feedback closed-loop step, specifically including the following processing steps: Regularly analyze the execution quality, response latency, and token consumption of each agent; optimize scheduling strategies, agent prompts, and task allocation rules using reinforcement learning based on the overall business utility function; and dynamically adjust based on historical data. Hyperparameters.

11. A system for implementing large-scale multi-agent collaborative processing of securities business models, characterized in that, The system includes: The basic model and tools layer is used to provide underlying intelligent computing power, financial data retrieval, market information query and knowledge enhancement capabilities; The agent container and orchestration layer, connected to the basic model and tool layer, is used to realize dynamic task graph generation, multi-agent role grouping, annotation-based review, debate-based generation, and collaborative execution scheduling. The business adaptation layer, connected to the aforementioned agent container and orchestration layer, is used to encapsulate multi-agent collaboration capabilities into standardized securities business interfaces, enabling integration with existing business systems. The intelligent agent container and orchestration layer include: An agent container is used to encapsulate and run multiple agents with different role labels in a containerized manner. The message bus employs a high-throughput message queue system to transmit messages and events between intelligent agents; The agent scheduler, connected to the message bus and agent container, is used to execute centralized scheduling strategies according to the task graph and role grouping mapping, and to manage the lifecycle of agents and task orchestration. The semantic context management module, built on the model context protocol, is connected to the message bus and is used to maintain a globally shared session context.

12. The system for large-scale multi-agent collaborative processing of securities business according to claim 11, characterized in that, The basic model and tool layer, the intelligent agent container and orchestration layer, and the business adaptation layer constitute a three-layer composite mapping system, which is represented by the following formula: ； in, Based on the basic model and tools layer, For intelligent agent containers and orchestration layers, This is a business adaptation layer.

13. The system for large-scale multi-agent collaborative processing of securities business as described in claim 11, characterized in that, The aforementioned basic model and tool layer includes a large model base, a set of tool plugins, and a vector knowledge base. The large model base is based on a general financial model, combined with N scenario-based fine-tuning sub-models. The set of tool plugins encapsulates financial data retrieval, computing engines, and communication interfaces. The vector knowledge base stores research reports, legal provisions, historical Q&A, and market data.

14. The system for large-scale multi-agent collaborative processing of securities business models according to claim 11, characterized in that, The aforementioned agent scheduler uniformly manages the lifecycle and task orchestration of agents, achieving load balancing and automatic fault recovery; when an agent fails to execute, it automatically restarts the instance or assigns another agent to take over the task. The built-in tracking subsystem monitors the execution chain and resource usage of each agent in real time.

15. A device for implementing large-scale multi-agent collaborative processing in securities business, characterized in that, The device includes: A processor is configured to execute computer-executable instructions; The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the method for large-scale multi-agent collaborative processing of securities business as described in any one of claims 1 to 10.

16. A processor for implementing large-scale multi-agent collaborative processing in securities business, characterized in that, The processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the method for large-scale multi-agent collaborative processing of securities business as described in any one of claims 1 to 10.

17. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method for large-scale multi-agent collaborative processing of securities business as described in any one of claims 1 to 10.

Images