Business processing system based on multiple agents

Abstract

The embodiment of the invention provides a business processing system based on multiple agents, and relates to the technical field of data processing. The system comprises an agent layer which is integrated with a plurality of professional agents, the professional agents comprise a business agent and a task scheduling agent, and the task scheduling agent is used for receiving a task request, dynamically allocating a task to a service agent corresponding to the task based on the task request, monitoring a processing state and a system load of each service agent, optimizing a task execution path, and coordinating cooperative work of a plurality of service agents to enable the service agents to process the task in parallel; the collaboration layer is used for providing information sharing and collaborative decision for a plurality of professional agents; and the resource layer is used for providing shared knowledge and tool support for a plurality of professional agents. Through one embodiment of the invention, the problem of low efficiency of the business process in the related technology is at least solved, and the effect of improving the efficiency of the business process in the related technology is further realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular, to a business processing system based on multiple agents. BACKGROUND

[0002] In the related art, the transaction receiving, compliance audit, risk assessment, fund clearing, and account service of cross-border renminbi payment business highly rely on manual intervention, and lack intelligent distributed collaborative processing mechanism. There is a lack of effective automatic connection and intelligent scheduling capability between business links, and multiple task parallel processing and end-to-end full-process closed-loop automation cannot be realized, resulting in low efficiency of business process. SUMMARY

[0003] Embodiments of the present application provide a business processing system based on multiple agents to at least solve the problem of low efficiency of business process in the related art.

[0004] According to an embodiment of the present application, a business processing system based on multiple agents is provided, comprising:

[0005] An agent layer integrated with multiple professional agents, wherein the professional agents include: a business agent, a task scheduling agent, the task scheduling agent is configured to receive a task request, dynamically allocate tasks to the business agent corresponding to the task based on the task request, monitor the processing state and system load of each business agent, optimize the task execution path, and coordinate the collaborative work of multiple business agents, so that the business agents process the tasks in parallel;

[0006] A collaboration layer configured to provide information sharing and collaborative decision-making for the multiple professional agents;

[0007] A resource layer configured to provide shared knowledge and tool support for the multiple professional agents.

[0008] In an implementation, the professional agents include:

[0009] An LLM reasoning engine module configured to receive a task prompt word, perform natural language understanding and logical reasoning processing based on the task prompt word, to generate a tool calling instruction and a decision result;

[0010] A tool calling module configured to provide the LLM reasoning engine module with a tool corresponding to the tool calling instruction based on the tool calling instruction;

[0011] A structured memory module configured to provide the LLM reasoning engine module with a working memory, so that the LLM reasoning engine module decides the task to be executed next based on the working memory;

[0012] a knowledge graph access interface module, configured to convert a query request into a graph query statement, so as to provide the LLM inference engine module with query capability of a knowledge graph based on the graph query statement; wherein the knowledge graph comprises: a business entity unit, a relationship type unit, and a knowledge updating mechanism unit, so as to provide the LLM inference engine module with shared knowledge support;

[0013] a prompt engineering module, configured to convert a business request into the task prompt word that can be understood by the LLM inference engine module.

[0014] In an embodiment, the tool calling module comprises:

[0015] an API tool set unit, encapsulating a plurality of interfaces for interacting with external systems, so as to have the capability of interacting with a plurality of external systems to exchange data;

[0016] a model tool set unit, integrating a plurality of machine learning models, so as to use the plurality of machine learning models as special tools;

[0017] a data tool set unit, integrating a database query tool, a document retrieval tool, a data statistics tool, and a format conversion tool, so as to provide data query and processing capability.

[0018] In an embodiment, the structured memory module comprises:

[0019] a short-term memory unit, configured to store context information of a current session, wherein the context information comprises: a business ID and basic information currently being processed, called tools and returned results, interaction history with other professional agents, and intermediate state of a current task;

[0020] a long-term memory unit, configured to store experience and knowledge that need to be persisted, wherein the experience and knowledge that need to be persisted comprise: historical case library data, experience summary data, user portrait data, and abnormal pattern library data;

[0021] a working memory unit, configured to store a task queue and to-do items that are currently being processed, wherein the task queue and to-do items that are currently being processed comprise: a to-do task list and priority, task dependency relationship, task execution plan and schedule, and temporarily shelved tasks and reasons.

[0022] In an embodiment, the task prompt word comprises: role definition, task description, context information, tool description, output format constraint, and thinking chain guidance.

[0023] In an embodiment, the business entity unit comprises: a customer entity sub-unit, a business entity sub-unit, a regulatory entity sub-unit, and a risk entity sub-unit.

[0024] In an embodiment, the relationship type unit comprises: a connection relationship subunit, a connection enterprise subunit, a business relationship subunit, a compliance relationship subunit, and a risk relationship subunit.

[0025] In an embodiment, the knowledge updating mechanism unit comprises:

[0026] An automatic synchronization updating subunit is configured to periodically call the API interface of the regulatory authority to automatically synchronize the latest regulatory rules and restricted list information;

[0027] A feedback updating subunit is configured to enable the plurality of professional agents to submit an updating request to the knowledge graph when discovering new entities, relationships, or rules in the process of handling the business;

[0028] A manual assisted review mechanism subunit is configured to send important knowledge updates to a manual review queue and call the review results for updating.

[0029] In an embodiment, the task scheduling agent is configured with:

[0030] A load monitoring tool is configured to monitor the task queue length, response time, and success rate of the plurality of business agents in real time;

[0031] A reinforcement learning strategy model is configured to generate task allocation and path planning decisions based on the current system state through a scheduling strategy, wherein the scheduling strategy is trained based on a proximal policy optimization algorithm;

[0032] A priority calculation tool is configured to calculate the task priority based on the business data, customer level, and business type;

[0033] A path simulation tool is configured to simulate the expected time and success rate of different execution paths based on the task allocation and path planning decisions.

[0034] In an embodiment, the proximal policy optimization algorithm comprises:

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] ;

[0040] wherein, is the current policy network parameter; old policy network parameters expected operator old policy induced state marginal distribution task scheduling time domain length time domain decay adjustment coefficient advantage function current policy action conditional probability old policy action conditional probability PPO clipping coefficient clipping function high-order KL constraint weight KL divergence series expansion order KL divergence of new and old policies KL divergence constraint threshold total number of professional agents resource weight of the th agent resource utilization rate of the th agent minimum resource utilization rate threshold total number of tasks to be processed risk weight of the th task compliance risk level of the th task maximum allowed risk threshold value network output value network parameters multi-objective reward weight task processing time consumption of the th step maximum allowed task processing time consumption decision accuracy rate of the th step action response time of the th agent at the th moment maximum allowed response time of the professional agent priority of the th task at the th moment highest priority of the task business peak period duration system compliance matching degree

[0041] Through one of the embodiments of the present application, in the intelligent agent layer, multiple business intelligent agents realize the professional division of labor in the links of transaction receiving, compliance review, etc., the task scheduling intelligent agent promotes the parallel processing of tasks of each business intelligent agent by dynamically allocating tasks, monitoring the processing state and system load, and optimizing the execution path, breaks the bottleneck of traditional serial processing; the collaboration layer provides a unified information sharing channel and collaborative decision-making mechanism for each professional intelligent agent, ensures real-time data intercommunication and efficient decision-making linkage of each link, and avoids repeated operation and information barriers; the resource layer provides instant knowledge support and tool calling capability for intelligent agents through sharing knowledge graph and tool set, reduces the resource redundancy and processing delay of each intelligent agent. The three layers support and cooperate closely, realize the full-process closed-loop automation of the bank business end to end, completely get rid of the dependence on manual intervention, greatly improve the parallelism and connection efficiency of business processing, and finally significantly improve the overall business process efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0042] The drawings described herein are used to provide further understanding of the present application, constitute a part of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute improper limitation on the present application. In the drawings:

[0043] Figure 1 is a structure diagram of a business processing system based on multiple intelligent agents according to an embodiment of the present application Figure 1 ;

[0044] Figure 2 is a structure diagram of a business processing system based on multiple intelligent agents according to an embodiment of the present application Figure 2 ;

[0045] Figure 3 is a structure diagram of a professional intelligent agent according to an embodiment of the present application;

[0046] Figure 4 is a structure diagram of a knowledge graph access interface module according to an embodiment of the present application;

[0047] Figure 5 is a signaling diagram of multiple professional intelligent agents in the scenario of communication and cooperation through the collaboration layer to complete complex business tasks according to an embodiment of the present application;

[0048] Figure 6 is a complete automatic processing flowchart of cross-border renminbi business according to an embodiment of the present application;

[0049] Figure 7 is a dynamic optimization mechanism diagram of task scheduling realized by using reinforcement learning technology according to an embodiment of the present application;

[0050] Figure 8 is a system deployment architecture diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0051] Hereinafter, the present application will be described in detail with reference to the accompanying drawings and examples. It should be noted that the examples and features in the examples of the present application can be combined with each other without conflict.

[0052] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and do not necessarily describe a specific order or sequence.

[0053] In the technical solution of the present application, the collection, storage, use, processing, transmission, provision and disclosure of financial data or user data and other information involved in the technical solution comply with relevant laws and regulations and do not violate public order and good customs.

[0054] It should be noted that in the examples of the present application, some software, components, models and other existing solutions in the industry may be mentioned, which should be considered as exemplary, and the purpose is only to illustrate the feasibility of the implementation of the technical solution of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0055] In the examples of the present application, a business processing system based on multiple agents is provided, Figure 1 is a structural diagram of the business processing system based on multiple agents according to the examples of the present application Figure 1 , Figure 2 is a structural diagram of the business processing system based on multiple agents according to the examples of the present application Figure 2 As shown in Figure 1 and Figure 2 , the system comprises:

[0056] The agent layer is integrated with multiple professional agents, wherein the professional agents include business agents and task scheduling agents. The task scheduling agent is used to receive task requests, dynamically allocate tasks to the corresponding business agents based on the task requests, monitor the processing status of each business agent and the system load, optimize the task execution path, and coordinate the collaborative work of multiple business agents to enable the business agents to process tasks in parallel.

[0057] In an exemplary embodiment, as shown in Figure 2 , the agent layer is integrated with six professional agents (five business agents and one task scheduling agent). Each professional agent is constructed based on a large language model (LLM) and has natural language understanding, reasoning and decision-making, and autonomous learning capabilities. Among them, the five business agents are:

[0058] Transaction parsing agent: responsible for receiving and parsing cross-border transaction instructions, business documents, extracting key elements such as transaction amount, currency, payee and payer information, and converting unstructured text into structured data.

[0059] Compliance audit agent: responsible for conducting compliance review of transactions according to Anti-Money Laundering (AML), Know Your Customer (KYC) and other regulatory requirements, checking the legality of customer identity, transaction background, and source of funds.

[0060] Risk assessment agent: responsible for comprehensive analysis of transaction characteristics, customer historical behavior, and correlation information, to identify potential fraud risk, credit risk and operational risk, and give risk level score and risk factor explanation.

[0061] Funds clearing agent: responsible for selecting the optimal clearing path, generating clearing instructions, completing cross-border fund transfer through CIPS (RMB cross-border payment system) or bank channel, and monitoring the clearing execution status.

[0062] Reconciliation service agent: responsible for comparing transaction records with bank flow and CIPS clearing results, identifying differences, analyzing difference causes, and performing automatic reconciliation or generating manual processing tasks.

[0063] Task scheduling agent: as a global coordinator, responsible for receiving business requests, dynamically allocating tasks to various professional agents, monitoring the processing status and system load of various agents, optimizing task execution path, and coordinating the collaborative work of multiple agents.

[0064] Collaboration layer, used to provide information sharing and collaborative decision-making for multiple professional agents;

[0065] In an exemplary embodiment, as shown in Figure 2 The collaboration layer provides communication and collaboration mechanisms between agents. The collaboration layer includes: agent communication protocol, message bus, and collaborative decision engine. Among them, the agent communication protocol defines the format and specification of message exchange between agents; the message bus is responsible for message routing and distribution; the collaborative decision engine is responsible for comprehensive analysis of multiple agent results and global decision optimization.

[0066] Resource layer, used to provide shared knowledge and tool support for multiple professional agents.

[0067] In an exemplary embodiment, as shown in Figure 3As shown, the resource layer provides shared knowledge and tool support for the agent. The resource layer includes: knowledge graph, structured memory bank, tool library, external system interface. Among them, the knowledge graph stores business entity, relationship and rule knowledge; the structured memory bank maintains the short-term and long-term memory of each professional agent; the tool library contains various API interfaces, machine learning models, data processing tools; the external system interface is responsible for the interface with the CIPS system (RMB cross-border payment system), bank system, credit investigation system and other external systems.

[0068] By adopting the above technical solutions, in the agent layer, multiple business agents realize professional division of labor in transaction receiving, compliance auditing and other links, the task scheduling agent promotes parallel processing of tasks of each business agent by dynamically allocating tasks, monitoring processing state and system load, and optimizing execution path, breaks the traditional serial processing bottleneck; the collaboration layer provides a unified information sharing channel and collaborative decision-making mechanism for each professional agent, ensures real-time data intercommunication and efficient decision-making linkage of each link, and avoids repeated operation and information barriers; the resource layer provides instant knowledge support and tool calling capability for the agent through sharing knowledge graph and tool set, reduces resource redundancy and processing delay of each agent. The three layers support and cooperate closely, realize the full-process closed-loop automation of bank business end-to-end, completely get rid of the dependence on manual intervention, greatly improve the parallelism and connection efficiency of business processing, and finally significantly improve the overall business process efficiency.

[0069] Figure 3 is a structural schematic diagram of a professional agent according to an embodiment of the present application, as Figure 4 As shown, in one embodiment, the professional agent includes:

[0070] The LLM reasoning engine module is used for receiving a task prompt word, performing natural language understanding and logical reasoning processing based on the task prompt word, to generate a tool calling instruction and a decision result.

[0071] In an exemplary embodiment, the LLM reasoning engine module is the core decision-making module of the agent, which is constructed based on a large-scale pre-trained language model (such as GPT-4, Claude, and Ernie etc.). The engine receives the task prompt word (including task description) processed by the Prompt engineering module, and performs natural language understanding, logical reasoning and decision planning. The LLM reasoning engine module has the following key capabilities: natural language understanding: capable of understanding complex business instructions and document content, extracting key information and semantic relationships. Reasoning and decision-making: based on business rules, historical experience and current context, multi-step reasoning is performed to generate a decision-making scheme. Tool orchestration: capable of determining when to call which tool, and decomposing complex tasks into multiple tool calling sequences. Self-reflection: capable of evaluating the quality of its own output and self-correcting or requesting assistance from other agents when necessary.

[0072] The tool invocation module is used to provide the LLM inference engine module with the tools corresponding to the tool invocation commands based on the tool invocation commands.

[0073] The structured memory module provides working memory for the LLM inference engine module, enabling the LLM inference engine module to determine the next task to be executed based on the working memory.

[0074] Figure 4 This is a structural diagram of the knowledge graph access interface module according to an embodiment of this application, such as... Figure 5 As shown, the knowledge graph access interface module is used to convert query requests into graph query statements, so as to provide the LLM inference engine module with the ability to query the knowledge graph based on the graph query statements; wherein, the knowledge graph includes: business entity units, relation type units, and knowledge update mechanism units, to provide shared knowledge support for the LLM inference engine module;

[0075] In one exemplary implementation, the knowledge graph access interface module provides professional agents with the ability to query a shared knowledge graph. Through this module, the professional agent can perform the following functions: query the attributes and relationships of business entities, such as "querying the related companies of this customer"; traverse the relationship graph, such as "finding all fund flows involved in this transaction"; match regulatory rules, such as "checking whether this transaction violates regulations on proper fund processing"; and infer implicit relationships, such as "inferring whether this customer has any abnormal associations". The knowledge graph access interface module supports the Cypher query language (for the Neo4j graph database) and also supports describing query intent in natural language, which is automatically converted into graph query statements by the interface.

[0076] In one implementation, the business entity unit includes: a customer entity subunit, a business entity subunit, a regulatory entity subunit, and a risk entity subunit.

[0077] In one exemplary implementation, the data stored in the customer entity sub-unit may include: Corporate customers: including attributes such as company name, unified social credit code, registered address, business scope, and registered capital. Individual customers: including attributes such as name, ID number, contact information, occupation, and income level. Account information: including attributes such as account number, bank name, account type, currency, and account opening date.

[0078] The data stored in the business entity sub-unit may include: Transaction record data: including attributes such as transaction ID, transaction time, transaction amount, currency, transaction type, and transaction status. Contract data: including attributes such as contract number, signing date, contract amount, and contract terms summary. Bill data: including attributes such as bill number, bill type, issuance date, and bill amount.

[0079] The data stored in the regulatory entity subunit may include: Regulatory rules: including attributes such as rule number, rule name, rule content, effective date, and scope of application. Compliance requirements: including attributes such as requirement type (e.g., KYC requirement, AML requirement), specific requirement description, and inspection items. Restricted lists: including attributes such as reason for inclusion, inclusion time, inclusion organization, and restricted list level.

[0080] The data stored in the risk entity sub-unit may include: Risk events: including attributes such as event ID, event type, occurrence time, risk level, and processing result. Fraud patterns: including attributes such as pattern description, feature vector, number of historical cases, and recognition accuracy.

[0081] In one implementation, the relationship type unit includes: association relationship subunit, related enterprise subunit, business relationship subunit, compliance relationship subunit, and risk relationship subunit.

[0082] In one exemplary implementation, the data stored in the relationship sub-unit may include: actual controller: the actual controller relationship of the enterprise; account opening: the account opening relationship between customers. The data stored in the related enterprise sub-unit may include: related enterprises: investment relationships, supply chain relationships, and cooperative relationships between enterprises.

[0083] The data stored in the business relationship sub-unit may include: Payer / Payee: the relationship between transaction records and account fund flows; Corresponding Contracts / Invoices: the correspondence between transaction records and contracts / invoices.

[0084] The compliance relationship sub-unit stores data that may include: Applicable rules: Regulatory rules that the transaction must comply with. Compliance requirements: Whether the transaction meets compliance requirements. Restriction list: Relationships where the client is on a restricted list.

[0085] The data stored in the risk relationship sub-unit may include: Triggering risk events: Risk events triggered by transactions. Matching fraud patterns: The matching relationship between risk events and known fraud patterns. Associated risk events: The association between a customer and their historical risk events.

[0086] In one implementation, the knowledge update mechanism unit includes:

[0087] The automatic synchronization update sub-unit is used to periodically call the API interface of regulatory agencies to automatically synchronize the latest regulatory rules and restricted list information;

[0088] In one exemplary implementation, for example, the regulatory rules entity automatically synchronizes the latest regulatory rules and restricted list information by periodically calling the regulatory agency's API interface. For instance, the list of restricted entities for handling unusual fund flows is automatically updated from the banking system every day at midnight.

[0089] The feedback update subunit is used to enable multiple specialized intelligent agents to submit update requests to the knowledge graph when they discover new entities, relationships or rules during the process of handling business.

[0090] In one exemplary implementation, for example, if a risk assessment agent discovers a new fraud pattern, it will add the pattern to the fraud pattern entity in the knowledge graph; if a reconciliation service agent discovers frequent financial transactions between two companies, it will add the related company relationship to the knowledge graph.

[0091] The manual review mechanism subunit is used to send important knowledge updates to the manual review queue and call the review results for updates.

[0092] In one exemplary implementation, for example, important knowledge updates (such as adding fraud patterns or modifying the interpretation of regulatory rules) require manual review and confirmation before they can take effect. The system submits the knowledge updates to be reviewed to a manual review queue, where business experts will conduct the review.

[0093] The Prompt module is used to convert business requests into task prompts that the LLM inference engine module can understand.

[0094] In one implementation, task prompts include: role definition, task description, contextual information, tool description, output format constraints, and thought process guidance.

[0095] In one exemplary implementation, the Prompt engineering module is responsible for translating business requests into prompts that the LLM inference engine module can understand. A complete Prompt may include: Role definition: clearly defining the agent's professional role. For example, the Prompt for a risk assessment agent might begin with: "You are a senior cross-border payment risk analysis expert, skilled in identifying fraud and compliance risks in transactions...".

[0096] Task Description: Clearly describe the task that needs to be completed. For example: "Please analyze the risk level of the following cross-border transaction: [Transaction Information]".

[0097] Contextual information: Provides relevant background information and constraints. For example: "Customer history records: [Memory retrieval results]; Relevant regulatory requirements: [Knowledge graph query results]".

[0098] Tool Description: List the available tools and how to use them. For example: "You can call the following tools: 1. xgboost_risk_score(features): Calculate the risk score; 2. query_blacklist(customer_id): Query the blacklist...".

[0099] Output format constraints: Specifies the data structure and format of the output. For example: "Please output in JSON format, including the fields: risk_level (risk level: low / medium / high), risk_score (risk score: 0-100), risk_factors (list of risk factors), and reasoning (reasoning process)."

[0100] Guided thought process chain: Guide LLM through step-by-step reasoning. For example: "Please analyze according to the following steps: 1. Extract transaction features; 2. Query historical records; 3. Call up the risk model; 4. Make a comprehensive judgment; 5. Give a conclusion."

[0101] Therefore, a well-designed Prompt can guide the LLM inference engine module to fully utilize its reasoning capabilities, accurately complete business tasks, and output structured decision results.

[0102] In one implementation, the tool invocation module includes:

[0103] The API toolset unit encapsulates various interfaces for interacting with external systems, enabling it to exchange data with a variety of external systems.

[0104] In one exemplary implementation, interfaces for interacting with external systems include: a CIPS system interface for submitting RMB cross-border payment instructions and querying clearing status; a banking system interface for querying account information and executing fund transfers; a credit reporting system interface for querying credit records of enterprises and individuals; and a regulatory reporting interface for reporting compliance data to regulatory agencies.

[0105] The model toolset unit integrates multiple machine learning models to serve as dedicated tools for various machine learning models;

[0106] In one exemplary implementation, for example, the machine learning model includes: an XGBoost risk scoring model: a gradient boosting tree model trained on historical transaction data, which takes a transaction feature vector as input and outputs a risk score between 0 and 1. A rules engine: stores predefined business rules and regulatory rules, supporting rapid rule matching and execution. An anomaly detection model: identifies abnormal transaction patterns based on statistical analysis or the Isolation Forest algorithm. A document OCR tool: performs text recognition on uploaded image files such as invoices and contracts.

[0107] The data toolset unit integrates database query tools, document retrieval tools, data statistics tools, and format conversion tools to provide data query and processing capabilities.

[0108] In one exemplary implementation, the system includes: a database query tool that supports SQL queries to retrieve structured data such as customer information and transaction history from a relational database; a document retrieval tool that retrieves relevant business documents and case records from a document library; a data statistics tool that calculates metrics such as transaction amount statistics and frequency analysis; and a format conversion tool that converts data between different formats (JSON, XML, CSV, etc.).

[0109] By adopting the above technical solution, the tool invocation module uses the Function Calling mechanism. The LLM inference engine module generates tool invocation instructions (including tool name and parameters) according to the task requirements. After the tool is executed, it returns the result to the LLM inference engine module. The LLM inference engine module continues inference or performs the next tool invocation based on the returned result.

[0110] In one implementation, the structured memory module includes:

[0111] The short-term memory unit is used to store the context information of the current session. The context information includes: the business ID and basic information of the currently processed business, the tools that have been called and their return results, the interaction history with other professional intelligent agents, and the intermediate state of the current task.

[0112] In one exemplary implementation, for example, the short-term memory unit stores context information of the current session, including: the transaction ID and basic information of the currently processed transaction, the tools invoked and their return results, the interaction history with other agents, and the intermediate state of the current task. Short-term memory data is cleared or archived to long-term memory after the task is completed. Short-term memory data is stored using a key-value pair structure and is read and written quickly using an in-memory database such as Redis.

[0113] Long-term memory units are used to store experiences and knowledge that need to be persisted. These experiences and knowledge include: historical case database data, experience summary data, user profile data, and abnormal pattern database data.

[0114] In one exemplary implementation, for example, the long-term memory unit stores experiences and knowledge that need to be persisted, including: a historical case library: transaction cases that were successfully or unsuccessfully processed and their processing procedures; experience summaries: processing patterns and decision-making rules summarized from multiple cases; user profiles: customer trading habits, risk characteristics, and preference settings; and an anomaly pattern library: identified fraud patterns and abnormal behavioral characteristics. The long-term memory is organized in a hierarchical structure, supporting similarity-based case retrieval and rule-based exact matching. The long-term memory data is stored in a NoSQL database such as MongoDB, supporting flexible data structures and efficient querying.

[0115] The working memory unit is used to store the currently processed task queue and to-do items. The currently processed task queue and to-do items include: a list of to-do tasks and their priorities, task dependencies, task execution plans and schedules, and temporarily shelved tasks and their reasons.

[0116] In one exemplary implementation, for example, the system stores a queue of tasks and to-do items currently being processed by the specialized agent, including: a list of tasks to be processed and their priorities, task dependencies (some tasks need to wait for other tasks to complete), task execution plans and schedules, and temporarily shelved tasks and their reasons. The working memory unit acts as a "workbench" for each specialized agent, which determines the next task to execute based on the working memory. The working memory supports dynamic updates, automatically adjusting when new tasks arrive or tasks are completed.

[0117] The memory read / write update mechanism is as follows:

[0118] Retrieval: Before making a decision, the LLM inference engine module retrieves relevant information from memory through natural language queries. For example: "Retrieve this customer's transaction records for the past 3 months" or "Find similar risk cases".

[0119] Write: After processing, LLM writes important information into memory. For example: "Mark this transaction as high risk and record the risk factors", "Add the processing result to the case library".

[0120] Update: When the agent learns new knowledge or discovers errors, it updates the information in its memory. For example, "update the client's risk level to medium" or "correct the previous judgment on this trading pattern".

[0121] In one implementation, to provide a shared memory storage system for each specialized intelligent agent, the structured memory module can be a structured memory bank. The structured memory bank employs a hierarchical and categorized organizational approach.

[0122] (1) Memory organization form:

[0123] Key-value pair structure: used to store simple state information, such as:

[0124] Transaction ID → Current processing status (Pending parsing, parsing completed, under review, risk assessment in progress, clearing in progress, completed);

[0125] Customer ID → Last transaction time;

[0126] Agent ID → Current load level (idle, normal, busy);

[0127] Hierarchical structure: Used to store information with hierarchical relationships, for example:

[0128] Customer A;

[0129] Basic information;

[0130] Company Name;

[0131] Registration address;

[0132] Credit rating;

[0133] Transaction history;

[0134] Transaction in 2024;

[0135] Transaction 001;

[0136] Transaction 002;

[0137] Transaction in 2025;

[0138] Risk events;

[0139] Risk Event 001;

[0140] Risk event 002.

[0141] Graph structure: used to store information about related networks, such as customer relationship networks (related to other customers and businesses) and transaction fund flow networks (the path of funds flowing between multiple accounts).

[0142] (2) Memory retrieval mechanism:

[0143] Similarity-based case retrieval: When the agent needs to refer to historical cases, vector similarity retrieval is used. The specific steps are as follows:

[0144] S1. Encode the characteristics of the current transaction (transaction amount, currency, customer type, transaction time, etc.) into a feature vector;

[0145] S2. Calculate the cosine similarity in the case library of long-term memory and find the top-K most similar historical cases;

[0146] S3. Provide these cases to the LLM for reference;

[0147] S4. Rule-based exact matching: When an agent needs to find records based on specific conditions, it uses exact matching. For example: query "all high-risk transactions of customer A in the last 30 days"; query "transactions with a risk score greater than 80 that involve overseas accounts".

[0148] (3) Memory decay and update strategies:

[0149] To prevent the memory bank from expanding indefinitely while ensuring that important information is not forgotten, a memory decay mechanism is employed:

[0150] Time-based decay: Short-term memory is automatically cleared 24 hours after the task is completed; cases older than 3 years and with low access frequency in long-term memory are archived to cold storage. Importance-based retention: Memory with high business value is marked as "important" and permanently retained, including: records of handling major risk events, cases of successfully identifying fraud, transaction records related to regulatory penalties, and access frequency-based eviction (using an LRU (Least Recently Used) strategy to downgrade or delete memories that have not been accessed for a long time).

[0151] The memory update adopts a version control mechanism, creating a new version each time the memory is updated, retaining historical versions for 3 months, and supporting backtracking and auditing.

[0152] In one implementation, the task scheduling agent is configured with:

[0153] Load monitoring tools are used to monitor the task queue length, response time, and success rate of multiple business intelligence agents in real time.

[0154] A reinforcement learning policy model is used to generate task allocation and path planning decisions based on the current system state through a scheduling policy; wherein the scheduling policy is trained based on a near-end policy optimization algorithm.

[0155] Priority calculation tool, used to calculate task priority based on business data, customer level, and business type;

[0156] Path simulation tools are used to simulate the expected time and success rate of different execution paths based on task allocation and path planning decisions.

[0157] In one implementation, the near-end policy optimization algorithm includes:

[0158] ;

[0159] ;

[0160] ;

[0161] ;

[0162] ;

[0163] in, These are the network parameters for the current policy. For old strategy network parameters; For expectation operators; The marginal distribution of states induced by the old strategy; The time domain length for task scheduling; This is the time-domain attenuation adjustment coefficient; The dominant function; The conditional probability of the action of the current policy; The conditional probability of the action in the old strategy; The cutoff factor for PPO; This is a truncation function; These are higher-order KL constraint weights; Let KL divergence series be the order of its expansion. The KL divergence between the old and new strategies; The threshold for KL divergence constraint; The total number of specialized intelligent agents; For the first Resource weights for each agent; For the first Resource utilization rate of each intelligent agent; The minimum resource utilization threshold; This represents the total number of tasks pending processing. For the first Risk weights for each task; For the first The compliance risk level of each task; This represents the maximum permissible risk threshold. Output to the value network; For value network parameters; Weighting for multi-objective rewards; For the first Step-by-step task processing time; This represents the maximum allowed processing time for the task. For the first The accuracy of decision-making for each step; For the first An intelligent agent in Response time at any given moment; This represents the maximum permissible response time for a specialized intelligent agent. for The task is in Time priority; This is the highest priority task. This refers to the duration of peak business hours; For system compliance matching degree.

[0164] In this near-end policy optimization algorithm, the policy parameters are used to optimize the near-end policy. and Constraint fits, higher-order KL divergence series (K) and constraint thresholds The regulation ensures smooth and stable policy updates, avoiding business interruptions caused by scheduling abrupt changes during multi-agent collaboration; leveraging the time domain length of task scheduling Time-domain attenuation adjustment coefficient and the duration of peak business periods It accurately adapts to the entire process and peak-valley fluctuations of cross-border payments, enhancing real-time scheduling and response; and utilizes multi-objective reward weighting. Integration of the first Step-by-step task processing time ,decision making Efficiency indicators, combined with the resource utilization rate of specialized intelligent agents Minimum resource utilization threshold and the priority of pending tasks This achieves a dynamic balance between agent load balancing and task priority adaptation; simultaneously, it integrates the third... Compliance risks of each task With the maximum permissible risk threshold System compliance matching degree These parameters optimize scheduling efficiency while meeting the regulatory requirements for cross-border payments. Ultimately, through the synergistic effect of these parameters, the task scheduling agent is provided with an optimal scheduling strategy that balances efficiency, stability, resource utilization, and compliance. This effectively supports the parallel collaboration of multiple professional agents and the automation of the entire process, further improving the processing efficiency and reliability of cross-border RMB payment business.

[0165] The following examples illustrate the specific configurations of the six specialized intelligent agents mentioned above.

[0166] 1. Regarding the transaction parsing agent:

[0167] 1.1 Role Definition Prompt: "You are an experienced cross-border payment transaction document analysis expert, skilled at accurately extracting key information from various transaction instructions, contracts, invoices, and other documents. You need to identify elements such as transaction amount, currency, payer, payee, transaction purpose, and related attachments, and convert the information into a structured format."

[0168] 1.2 Configured toolset: OCR document recognition tool: performs text recognition on uploaded PDF and image format invoices and contracts; Database query tool: queries customer basic information and account information; Format verification tool: verifies the legality of amount format, date format, and account number format; Currency conversion tool: performs exchange rate query and currency conversion.

[0169] 1.3 Output format: JSON structured data, containing the following fields:

[0170] {"transaction_id":"Unique identifier for the transaction",

[0171] "amount":"Transaction amount (number)",

[0172] "currency": "currency code (e.g., CNY, USD)",

[0173] "payer":{"name":"Payer's Name",

[0174] "account":"Payment account",

[0175] "bank":"Payment Bank"},

[0176] "payee":{"name":"Payee's Name",

[0177] "account":"payment account",

[0178] "bank":"receiving bank"},

[0179] "purpose":"Description of the purpose of the transaction",

[0180] "attachments":["List of attachments"],

[0181] "parse_confidence":"parse confidence (0-1)",

[0182] "ambiguities": ["ambiguities or uncertainties"]}.

[0183] 2. Regarding the compliance audit intelligent agent:

[0184] 2.1 Role Definition Prompt: "You are a professional cross-border payment compliance review expert, proficient in anti-abnormal fund handling regulations, foreign exchange management regulations, and international sanctions policies. You need to review whether transactions comply with regulatory requirements, check the authenticity of customer identities, the reasonableness of transaction backgrounds, and the legality of fund sources, identify suspicious transactions, and provide professional advice."

[0185] 2.2 Configured toolset: Knowledge graph query tool: queries regulatory rules, compliance requirements, and restricted list information; External credit reporting API: calls bank credit reporting interfaces and enterprise credit information disclosure systems; Rule engine: executes predefined compliance check rules (such as large transaction reporting thresholds and suspicious transaction feature matching); Sanctions list check tool: matches with sanctions lists such as OFAC (Office of Foreign Assets Control) and UN (United Nations).

[0186] 2.3 Output Format:

[0187] {"compliance_result":"Compliant / Non-compliant / Requires manual review",

[0188] "risk_level":"Compliance risk level: None / Low / Medium / High",

[0189] "check_items":[{"item":"Check item name (e.g., KYC check, restricted list check)",

[0190] "result":"Pass / Fail",

[0191] "details":"Detailed explanation"},

[0192] "violations":["List of violations"],

[0193] "recommendations": "Suggestions for handling"

[0194] "reasoning":"examine the reasoning process"}.

[0195] 3. Regarding the risk assessment agent:

[0196] 3.1 Role Definition Prompt: "You are a senior cross-border payment risk analysis expert, skilled in identifying fraud risk, credit risk, and operational risk in transactions. You need to comprehensively analyze multi-dimensional information such as transaction characteristics, customer historical behavior, and relationship networks, and use statistical models and machine learning tools to provide accurate risk assessment conclusions."

[0197] 3.2 Configured Toolset: XGBoost Risk Scoring Model: Input feature vector (containing 20+ dimensions of features such as transaction amount, frequency, time, and customer type), output risk probability between 0 and 1; Knowledge Graph Query Tool: Query customer association network and historical risk events; Anomaly Detection Tool: Detect anomalies in transaction patterns based on the Isolation Forest algorithm; Statistical Analysis Tool: Calculate customer transaction amount distribution, frequency statistics, year-on-year and month-on-month growth rates, and other indicators; Memory Retrieval Tool: Retrieve similar historical cases.

[0198] 3.3 Output Format:

[0199] {"risk_level":"Risk level: Low / Medium / High / Extremely High",

[0200] "risk_score":"Quantitative risk score (0-100)",

[0201] "risk_factors":[{"factor":"risk factor name",

[0202] "severity": "Severity level (1-5)"

[0203] "description":"Function Description"}],

[0204] "model_scores":{"xgboost_score":"XGBoost model score",

[0205] "anomaly_score":"anomaly detection score"},

[0206] "similar_cases":["List of similar historical case IDs"],

[0207] "recommendations": "risk response recommendations",

[0208] "reasoning":"risk analysis reasoning process"}.

[0209] 4. Regarding the intelligent fund clearing agent:

[0210] 4.1 Role Definition Prompt: "You are an experienced cross-border funds clearing expert, familiar with the CIPS system, SWIFT network, and the clearing processes of various banks. You need to select the optimal clearing path based on transaction information, generate correct clearing instructions, monitor the clearing execution status, and handle clearing anomalies."

[0211] 4.2 Configured toolset: CIPS system interface: submit RMB cross-border payment instructions and query settlement status; Bank API interface: call the settlement interface of partner banks; Exchange rate query tool: obtain real-time exchange rate information; Path planning algorithm: select the optimal settlement path based on factors such as historical success rate, settlement timeliness, and handling fees; Abnormal retry tool: automatically retry or switch paths when settlement fails.

[0212] 4.3 Output Format:

[0213] {"clearing_plan":{"channel":"Clearing channel (CIPS / direct bank connection)",

[0214] "route":"Specific liquidation path",

[0215] "estimated_time":"Estimated arrival time",

[0216] "fee":"Clearing fee"},

[0217] "instruction":{"instruction_id":"liquidation instruction ID",

[0218] "instruction_content":"content of the liquidation instruction"},

[0219] "execution_status":"Execution status: Pending execution / Executing / Success / Failure",

[0220] "status_details":"Status details",

[0221] "exception_handling": "Exception handling solution (if any)",

[0222] "reasoning":"path selection reasoning process"}.

[0223] 5. Regarding the reconciliation service intelligent agent:

[0224] 5.1 Role Definition Prompt: "You are a professional financial reconciliation and analysis expert, skilled at identifying discrepancies between transaction records and bank statements, analyzing the reasons for these discrepancies, and developing reconciliation plans. You need to accurately compare large amounts of data, identify issues such as differences in amount, time, and status, and provide professional handling suggestions."

[0225] 5.2 Configured toolset: Data comparison algorithm: performs batch comparison of system transaction records with bank statements and CIPS clearing records; Difference analysis tool: classifies identified differences (amount differences, time differences, missing records, etc.) and analyzes the causes; Automatic reconciliation rule engine: automatically performs reconciliation for simple differences (such as temporary differences caused by time delays); Report generation tool: generates reconciliation reports and difference details.

[0226] 5.3 Output Format:

[0227] {"reconciliation_result":"Reconciliation result: Consistent / Discrepancies exist",

[0228] "matched_count":"Number of matching records",

[0229] "discrepancy_count":"Number of discrepancy records",

[0230] "discrepancies":[{"transaction_id":"transaction ID",

[0231] "discrepancy_type":"discrepancy type",

[0232] "system_record":"System record",

[0233] "bank_record":"bank records",

[0234] "difference":"difference content",

[0235] "possible_cause":"Possible cause",

[0236] "auto_adjustable": "Whether automatic adjustment of accounts is possible",

[0237] "adjustment_plan":"adjustment plan"},

[0238] "report_link":"Reconciliation report download link",

[0239] "reasoning":"the process of reasoning through difference analysis"}.

[0240] 6. Regarding task scheduling agents:

[0241] 6.1 Role Definition Prompt: "You are a global system coordination expert, responsible for task scheduling and optimization of the entire cross-border payment business process. You need to allocate tasks reasonably based on business priorities, system load, and agent capabilities, plan the optimal execution path, and coordinate the collaborative work of multiple agents to ensure that the business is completed efficiently and accurately."

[0242] 6.2 Configured toolset: Load monitoring tool: Real-time monitoring of task queue length, response time, and success rate for each agent; Reinforcement learning policy model: Scheduling policy trained based on the PPO (Proximal Policy Optimization) algorithm, taking the current system state as input and outputting task allocation and path planning decisions; Priority calculation tool: Calculates task priority based on transaction amount, customer level, and business type; Path simulation tool: Simulates the expected time and success rate of different execution paths.

[0243] 6.3 Output Format:

[0244] {"task_assignment":[{"agent":"agent name",

[0245] "task": "Assigned task",

[0246] "priority":"priority",

[0247] "deadline":"expected completion time"}],

[0248] "execution_path":"Task execution path (serial / parallel / hybrid)",

[0249] "coordination_plan":"coordination plan",

[0250] "expected_completion_time":"Expected completion time",

[0251] "optimization_reasoning":"Scheduling optimization reasoning process"}.

[0252] Figure 5 This is a signaling diagram illustrating a scenario where multiple specialized intelligent agents communicate and collaborate through a collaboration layer to complete complex business tasks, according to embodiments of this application. Figure 6 As shown in the signaling diagram:

[0253] 1. Communication protocol design:

[0254] 1.1 A unified message communication protocol is used between intelligent agents, and the message format adopts a JSON structure:

[0255] {"message_id":"Unique identifier for the message",

[0256] "timestamp":"Message timestamp",

[0257] "from_agent":"Sender Agent ID",

[0258] "to_agent":"Receiving agent ID (or 'broadcast' for broadcast)",

[0259] "message_type":"Message type",

[0260] "priority":"Priority (1-5)",

[0261] "content":{ / / The specific content of the message, which varies depending on the message type},

[0262] "correlation_id":"The original task ID associated with it",

[0263] "reply_to":"The message ID of the reply (if it is a response message)"}.

[0264] 1.2 Message types include: TASK_REQUEST (task request): The scheduling agent assigns a task to a specialist agent; TASK_RESPONSE (task response): The specialist agent returns the result after completing the task; INFO_SHARE (information sharing): An agent actively shares information with other agents; CONSULTATION (negotiation request): An agent requests advice or assistance from other agents; NOTIFICATION (notification): System event notifications, such as status changes and anomaly alarms. Messages are routed and distributed through a message bus. The message bus is implemented based on a publish-subscribe pattern, where agents can subscribe to message types of interest and automatically receive relevant messages when they become available.

[0265] 2. Collaborative decision-making process:

[0266] 2.1 Information Collection Phase: After completing their respective analysis tasks, each specialized intelligent agent writes the results into the shared area of ​​structured memory. For example, the compliance audit agent writes the audit conclusion and violations, while the risk assessment agent writes the risk level and risk factors.

[0267] 2.2 Consultation Mechanism: When multiple agents disagree (e.g., compliance review deems it acceptable, but risk assessment considers it high-risk), a consultation mechanism is initiated: The task scheduling agent retrieves the analysis results and reasoning processes of each agent from its memory; the scheduling agent sends a CONSULTATION message to the relevant agents, requesting more detailed evidence; each agent provides supplementary explanations, for example, the compliance agent explains "although it complies with the rules, the rules may have loopholes," and the risk agent explains "highly similar to known fraud patterns"; the scheduling agent synthesizes the opinions of all parties, uses the reasoning capabilities of LLM to weigh the options, and makes a final decision.

[0268] 2.3 Conflict Resolution: When negotiations fail to reach an agreement, the following priority rules shall be adopted: Compliance issues take precedence: If there is a clear compliance violation, the transaction shall be rejected directly; Risk level threshold: If the risk score exceeds the preset high-risk threshold (e.g., 85 points), manual review is required even if the transaction is compliant; Manual intervention: For cases with low decision confidence, the process shall be transferred to manual handling.

[0269] For example, a typical normal transaction processing flow is as follows:

[0270] Task reception: Users or business systems submit cross-border transaction requests to the task scheduling agent.

[0271] Create a record: The scheduling agent creates a task record in structured memory, recording the initial state.

[0272] Transaction parsing: The scheduling agent assigns tasks to the transaction parsing agent, which extracts transaction information and writes it into memory.

[0273] Parallel review: The scheduling agent simultaneously triggers the compliance review agent and the risk assessment agent, which work in parallel: the compliance agent queries the regulatory rules and restricted lists in the knowledge graph and calls the rule engine to perform compliance checks; the risk agent queries the customer's historical records and related networks and calls the XGBoost model to calculate the risk score.

[0274] Collaborative decision-making: The scheduling agent reads the analysis results of the two agents and makes a comprehensive judgment.

[0275] Execution of liquidation: If the risk is deemed low and compliant, the scheduling agent will assign the task to the fund liquidation agent for execution.

[0276] Reconciliation Confirmation: After settlement is completed, the reconciliation service agent performs reconciliation, confirms the successful transaction, and updates the experience to the knowledge graph.

[0277] Return result: The scheduling agent returns the transaction processing result to the user.

[0278] Therefore, for high-risk or non-compliant transactions, the process terminates at the collaborative decision-making stage, the scheduling agent directly returns a rejection result or transfers it to manual review, and records the reason for rejection in memory for subsequent analysis and learning.

[0279] Figure 6 This is a complete automated processing flowchart of cross-border RMB business according to an embodiment of this application, such as... Figure 7 As shown, the complete business process includes the following key steps:

[0280] 1. Transaction Reception and Parsing: The system receives cross-border transaction requests from users or business systems. The task scheduling agent creates a task record and assigns it to the transaction parsing agent. The transaction parsing agent uses an OCR tool to recognize the uploaded document and employs NLP technology to extract key information such as transaction amount, currency, payee, and payee, performing format verification and data completion. If parsing fails (e.g., the document is blurry or key information is missing), a parsing failure error is returned directly. Upon successful parsing, the structured transaction information is written to the shared area of ​​structured memory.

[0281] 2. Parallel Compliance Review and Risk Assessment: The task scheduling agent simultaneously triggers both the compliance review agent and the risk assessment agent, allowing them to work in parallel to improve efficiency.

[0282] 2.1 Compliance Audit Process:

[0283] The system queries the knowledge graph for applicable regulatory rules (such as regulations on handling abnormal funds and foreign exchange management); checks whether the payer and payee are on restricted lists (including domestic regulatory restricted lists and international sanctions lists); uses a rule engine to match suspicious transaction characteristics (such as cash-intensive enterprises, frequent small transactions, transactions inconsistent with business scope, etc.); calls external credit reporting interfaces to verify the authenticity of customer information; and outputs the audit results: compliant / non-compliant, along with detailed inspection items and explanations of violations.

[0284] 2.2 Risk Assessment Process: The process involves: querying the customer's historical transaction records and risk event records from the structured memory; querying the customer's related enterprise network and actual controller relationships from the knowledge graph; extracting risk features (including 20+ dimensions of features such as transaction amount, transaction frequency, transaction time, customer type, and historical risk status); calling the XGBoost risk scoring model to calculate the risk probability; using anomaly detection tools to determine whether the transaction pattern is abnormal; retrieving similar historical risk cases from the memory; and outputting the risk assessment results: risk level (low / medium / high / extremely high), risk score (0-100), and a list of risk factors. After both agents complete their work, they write their respective analysis results into the structured memory.

[0285] 3. Collaborative Decision-Making: The task scheduling agent reads the results of compliance review and risk assessment, and makes a comprehensive judgment:

[0286] 3.1 Low-risk and compliant: Directly enters the clearing process. This type of transaction accounts for approximately 70-80% and is the main scenario for automated processing by the system.

[0287] 3.2 Medium risk and compliant: Requires manual review. Manual reviewers can view the complete analysis report (including compliance check details, risk factor descriptions, and similar case references) and make a decision to approve or reject. If manually approved, the liquidation process continues.

[0288] 3.3 High-risk or non-compliant: Transaction rejected outright, reason for rejection recorded, and result returned requiring manual processing. These transactions require in-depth investigation by business personnel, and reporting to regulatory agencies if necessary.

[0289] The decision-making process is completed by the LLM inference engine module, which can comprehensively consider multiple factors and provide explainable reasons for the decision. For example: "Although the transaction meets the formal compliance requirements, the risk assessment found that the client has a connection with a known fraud group, and the transaction amount has increased abnormally and significantly. Based on the overall assessment, it is considered high risk, and it is recommended to reject it or transfer it to a human for in-depth review."

[0290] 4. Funds Clearing Execution: For transactions that have passed the review, the task scheduling agent will assign the task to the funds clearing agent:

[0291] 4.1 The clearing agent selects the optimal clearing path (direct connection to CIPS or through a partner bank) based on information such as transaction amount, currency, and receiving and paying banks.

[0292] 4.2 Generate clearing instructions conforming to the CIPS system format;

[0293] 4.3 Submit clearing instructions by calling the CIPS system interface; monitor the clearing execution status and wait for the CIPS system to return the clearing results;

[0294] 4.4 If the settlement fails (e.g., insufficient account balance, abnormal receiving account, system timeout), the settlement agent will initiate the exception handling process;

[0295] 4.5 Analyze the reasons for failure and determine whether to retry or switch paths: For recoverable errors (such as network timeout, system busy), automatically retry, up to 3 times; for unrecoverable errors (such as account freeze, insufficient balance), record the reasons for failure and transfer to manual processing.

[0296] 4.6 After successful settlement, update the transaction status in the memory to "settlement completed".

[0297] 5. Reconciliation and Confirmation: The reconciliation service agent executes the reconciliation process:

[0298] 5.1 Retrieve transaction records from the business system database;

[0299] 5.2 Obtain bank statements from the banking system;

[0300] 5.3 Obtain clearing records from the CIPS system;

[0301] 5.4 Use data comparison algorithms to perform three-party data matching and check whether the amount, time, and status are consistent;

[0302] 5.5 If the accounts match, the transaction is confirmed as successful. The transaction case and processing experience are then updated into the knowledge graph and long-term memory for future learning.

[0303] 5.6 If discrepancies are found, the accounting agent will perform discrepancy analysis: Discrepancies that can be automatically adjusted: such as temporary discrepancies caused by time delays (bank statements arriving 1 to 2 days late), the system will automatically perform adjustments without manual intervention; Discrepancies that require manual processing: such as serious discrepancies like inconsistent amounts or missing transactions, will generate manual processing tasks for finance personnel to conduct in-depth investigations and adjustments.

[0304] 5.7 Finally, the system returns the transaction processing result to the user or business system: success, failure, or manual processing required.

[0305] Figure 7 This is a diagram illustrating a dynamic optimization mechanism for task scheduling using reinforcement learning techniques, according to an embodiment of this application. Figure 8 As shown, the steps of task scheduling may include:

[0306] 1. State Space Definition. The state space of reinforcement learning contains information in the following dimensions:

[0307] 1.1 Business Status: Number of tasks currently pending; Distribution of task types (ratio of ordinary transactions, large transactions, and high-risk transactions); Distribution of task urgency; Current time period (working hours, non-working hours, peak hours).

[0308] 1.2 Agent Status: Current load (idle, normal, busy) of each specialized agent; task queue length of each agent; average response time of each agent; historical success rate of each agent;

[0309] 1.3 System Resource Status: Server CPU and memory utilization; number of available database connections; message queue backlog; remaining LLMAPI call quota.

[0310] The state vector is a fixed-length numerical vector (e.g., 50-dimensional), with each dimension corresponding to a state feature. The state vector is input into the policy network of reinforcement learning.

[0311] 2. Action Space. The actions that a reinforcement learning agent can take include:

[0312] 2.1 Task allocation decision: Assign a newly arrived task to which agent instance (if multiple agent copies with the same role are deployed); adjust task priority (increase or decrease the priority of a task); decide whether to postpone the processing of some low-priority tasks.

[0313] 2.2 Path Selection Decision: Choose between a sequential or parallel path (e.g., for small transactions by low-risk clients, risk assessment can be skipped and clearing can proceed directly); decide whether a certain step is necessary (e.g., for long-term, high-quality clients, the compliance review process can be simplified); choose different clearing channels (CIPS direct connection or bank intermediary).

[0314] 2.3 Resource allocation decision: Dynamically adjust the number of agent instances (automatically expand or shrink based on load); adjust the LLMAPI call strategy (use a faster but more expensive model, or use a more economical but slightly slower model); the action space can be represented as discrete actions (such as "assign to agent A", "assign to agent B") or continuous actions (such as priority adjustment range).

[0315] 3. Reward Function Design. The reward function comprehensively considers multiple business objectives and is designed as follows:

[0316] 3.1 Positive Rewards: Processing Speed ​​Rewards: The shorter the processing time, the higher the reward. Accuracy bonus: Earn rewards for completing tasks accurately; deduct rewards for errors. Resource utilization reward: A reward is given when resource utilization is close to the target range (e.g., 70-80%). User satisfaction reward: Based on user ratings of the service (such as whether transactions were completed promptly and whether multiple manual interventions were required). Among these, This is a weighting coefficient, which can be adjusted according to business priorities.

[0317] 3.2 Negative Penalties: Error Penalty: A significant negative reward is given when a decision leads to an error in transaction processing (e.g., incorrectly rejecting a normal transaction or incorrectly approving a high-risk transaction). Timeout Penalty: A negative reward is given when the task processing time exceeds the time limit specified in the SLA (Service Level Agreement). Human Intervention Penalty: A slight negative reward is given when frequent human intervention is required (indicating a poor automation strategy).

[0318] Therefore, the total reward function is:

[0319] .

[0320] 4. Learning and Update Mechanism. This application's embodiment employs the PPO (Proximal Policy Optimization) algorithm for reinforcement learning:

[0321] 4.1 Online Learning. The system continuously learns during actual business operations: after each transaction task is completed, it collects status, action, and reward information; stores the experience in the experience replay buffer; when enough experience samples are accumulated (such as 100 transactions), it triggers a strategy update; and it uses the PPO algorithm to update the strategy network parameters and optimize task allocation and path selection strategies.

[0322] 4.2 Offline Training. Offline training is performed using historical business data: historical transaction records, processing procedures, and processing results are used as training data; supervised learning methods are used to pre-train the policy network to learn the decision-making patterns of human scheduling experts; and then online learning is used for fine-tuning and optimization.

[0323] 4.3 Policy Update Constraints. The PPO algorithm avoids performance degradation caused by policy mutations by limiting the policy update magnitude: it calculates the Kullback-Leibler divergence between the new and old policies; if the KL divergence exceeds a threshold, the update magnitude is limited; this ensures the stability of the policy during the optimization process.

[0324] 4.4 Multi-objective balancing: By adjusting the weight coefficients of the reward function, a balance can be achieved between different business objectives: during peak business periods, increase the weight of Rspeed to prioritize processing speed; during periods of high risk, increase the weight of Raccuracy to prioritize accuracy; and when resources are scarce, increase the weight of Rresource to optimize resource utilization.

[0325] Through continuous optimization using reinforcement learning, the decision-making ability of the task scheduling agent is constantly improved, enabling it to dynamically adjust strategies based on real-time business status and achieve globally optimal task scheduling and path planning.

[0326] Figure 8 This is a schematic diagram of the system deployment architecture according to an embodiment of this application, such as... ​ As shown, in the system deployment architecture:

[0327] 1. Agent Runtime Environment. Each agent runs as an independent microservice in a Docker container. Containerized deployment offers the following advantages: Environment Isolation: Each agent has an independent runtime environment and dependency libraries, avoiding version conflicts; Elastic Scaling: The number of agent instances can be automatically increased or decreased based on load; Rapid Deployment: New versions of agents can be quickly deployed and rolled back; Fault Recovery: Containers automatically restart after crashes, ensuring service availability; Container orchestration uses Kubernetes to achieve automatic scheduling, load balancing, and health checks of agent instances.

[0328] 2. LLM Inference Service. The LLM inference service provides inference capabilities for large language models and has two deployment methods: Cloud API method: calling the API interfaces of commercial large models such as Claude, GPT-4, and Wenxin Yiyan. The advantages are strong model performance and no need to maintain infrastructure; the disadvantages are higher cost and data privacy issues.

[0329] Local deployment: Deploy large open-source models (such as LLaMA, ChatGLM, etc.) on a local GPU server. Advantages include controllable data security and low long-term costs; disadvantages include the need for GPU hardware investment and potential performance limitations compared to commercial models.

[0330] In actual deployment, a hybrid approach can be adopted: use the locally deployed model for ordinary tasks, and use the high-performance model of the cloud API for complex tasks.

[0331] 3. Knowledge Graph Service. The knowledge graph is stored using the Neo4j graph database, a professional graph database that supports efficient graph traversal and relationship queries. The knowledge graph service provides the following interfaces: Entity Query: Query entity information based on entity ID or attribute; Relationship Traversal: Find relationship paths between entities; Pattern Matching: Match specific graph patterns (such as triangular associations, circular cash flows); Knowledge Reasoning: Perform knowledge reasoning based on rules; The knowledge graph supports read / write separation, with read operations distributed to multiple slave nodes to improve query performance; write operations are executed on the master node to ensure data consistency.

[0332] 4. Structured Memory Service. The structured memory service consists of two layers of storage: a Redis cache layer, which stores short-term memory and frequently accessed long-term memory, providing millisecond-level read and write performance; and a MongoDB persistence layer, which stores complete long-term memory, providing flexible document structures and efficient queries. Redis and MongoDB maintain data consistency through a cache update strategy: write operations update both Redis and MongoDB simultaneously; read operations prioritize reading from Redis, and if the cache is not found, then load from MongoDB and update the cache.

[0333] 5. Message Queues. Message queues are implemented using Kafka and are responsible for asynchronous communication between agents. Kafka features high throughput, persistence, and distributed processing, enabling it to handle large message volumes. Agents communicate using a publish-subscribe pattern: each agent acts as both a message producer and consumer; different message types are published to different topics; agents subscribe to topics of interest and receive relevant messages; Kafka guarantees message ordering and at least-once delivery, ensuring reliable communication between agents.

[0334] 6. External System Interfaces. The external system interface layer encapsulates interactions with external systems such as CIPS, banking systems, and credit reporting systems. This layer provides a unified interface specification, shielding the protocol differences between different external systems. The interface layer implements the following functions: Protocol Conversion: Converts internal JSON format to XML, fixed-length message, and other formats required by external systems; Connection Pool Management: Maintains a connection pool with external systems to improve connection reuse; Retry and Circuit Breaker: Automatically retryes external system calls when they fail, and triggers a circuit breaker mechanism when the failure rate is too high to avoid a cascading failure effect; Traffic Limiting: Controls the frequency of calls to external systems to avoid exceeding limits; Logging and Monitoring: Records all requests and responses to external system calls for auditing and troubleshooting.

[0335] Through the above device design, this application embodiment realizes a highly available, high-performance, and scalable multi-agent cross-border payment automation system. The system can support thousands of transactions per second while ensuring millisecond-level response time and availability.

[0336] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0337] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. A business processing system based on multi-agent intelligence, characterized in that, include: The intelligent agent layer integrates multiple specialized intelligent agents, including: a business intelligent agent and a task scheduling intelligent agent. The task scheduling intelligent agent is used to receive task requests, dynamically allocate tasks to the corresponding business intelligent agents based on the task requests, monitor the processing status and system load of each business intelligent agent, optimize task execution paths, and coordinate the collaborative work of multiple business intelligent agents so that the business intelligent agents can process the tasks in parallel. The collaboration layer is used to provide information sharing and collaborative decision-making for the multiple specialized intelligent agents. The resource layer is used to provide shared knowledge and tool support for the multiple specialized intelligent agents.

2. The system according to claim 1, characterized in that, The specialized intelligent agents include: The LLM inference engine module is used to receive task prompt words, perform natural language understanding and logical reasoning processing based on the task prompt words, and generate tool call instructions and decision results. The tool invocation module is used to provide the LLM inference engine module with the tool corresponding to the tool invocation instruction based on the tool invocation instruction; A structured memory module is used to provide working memory for the LLM inference engine module, so that the LLM inference engine module can determine the next task to be executed based on the working memory; The knowledge graph access interface module is used to convert query requests into graph query statements, so as to provide the LLM inference engine module with the ability to query the knowledge graph based on the graph query statements; wherein, the knowledge graph includes: business entity units, relation type units, and knowledge update mechanism units, so as to provide shared knowledge support for the LLM inference engine module; The Prompt module is used to convert business requests into task prompts that the LLM inference engine module can understand.

3. The system according to claim 2, characterized in that, The tool invocation module includes: The API toolset unit encapsulates various interfaces for interacting with external systems, enabling it to exchange data with a variety of external systems. The model toolset unit integrates multiple machine learning models to serve as dedicated tools. The data toolset unit integrates database query tools, document retrieval tools, data statistics tools, and format conversion tools to provide data query and processing capabilities.

4. The system according to claim 2, characterized in that, The structured memory module includes: A short-term memory unit is used to store the context information of the current session, wherein the context information includes: the currently processed business ID and basic information, the tools invoked and their return results, the interaction history with other professional intelligent agents, and the intermediate state of the current task; Long-term memory units are used to store experiences and knowledge that need to be persisted, wherein the experiences and knowledge that need to be persisted include: historical case database data, experience summary data, user profile data, and abnormal pattern database data; The working memory unit is used to store the currently processed task queue and to-do items, wherein the currently processed task queue and to-do items include: a list of to-do tasks and their priorities, task dependencies, task execution plans and schedules, and temporarily shelved tasks and their reasons.

5. The system according to claim 2, characterized in that, The task prompts include: role definition, task description, context information, tool instructions, output format constraints, and thought process guidance.

6. The system according to claim 2, characterized in that, The business entity units include: customer entity subunit, business entity subunit, regulatory entity subunit, and risk entity subunit.

7. The system according to claim 2, characterized in that, The relationship type units include: association relationship subunit, related enterprise subunit, business relationship subunit, compliance relationship subunit, and risk relationship subunit.

8. The system according to claim 2, characterized in that, The knowledge update mechanism unit includes: The automatic synchronization update sub-unit is used to periodically call the API interface of regulatory agencies to automatically synchronize the latest regulatory rules and restricted list information; The feedback update subunit is used to enable multiple specialized intelligent agents to submit update requests to the knowledge graph when they discover new entities, relationships or rules during business processing. The manual review mechanism subunit is used to send important knowledge updates to the manual review queue and call the review results for updates.

9. The system according to claim 1, characterized in that, The task scheduling agent is configured with: A load monitoring tool is used to monitor the task queue length, response time, and success rate of multiple business intelligence agents in real time. A reinforcement learning strategy model is used to generate task allocation and path planning decisions based on the current system state through a scheduling strategy; wherein the scheduling strategy is trained based on a near-end policy optimization algorithm. Priority calculation tool, used to calculate task priority based on business data, customer level, and business type; A path simulation tool is used to simulate the expected time and success rate of different execution paths based on the task allocation and path planning decisions.

10. The system according to claim 9, characterized in that, The near-end strategy optimization algorithm includes: ； ； ； ； ； in, These are the network parameters for the current policy. For old strategy network parameters; For expectation operators; The marginal distribution of states induced by the old strategy; The time domain length for task scheduling; This is the time-domain attenuation adjustment coefficient; The dominant function; The conditional probability of the action of the current policy; The conditional probability of the action in the old strategy; The cutoff factor for PPO; This is a truncation function; These are higher-order KL constraint weights; Let KL divergence series be the order of its expansion. The KL divergence between the old and new strategies; The threshold for KL divergence constraint; The total number of specialized intelligent agents; For the first Resource weights for each agent; For the first Resource utilization rate of each intelligent agent; The minimum resource utilization threshold; This represents the total number of tasks pending processing. For the first Risk weights for each task; For the first The compliance risk level of each task; This represents the maximum permissible risk threshold. Output for value networks; For value network parameters; Weighting for multi-objective rewards; For the first Step-by-step task processing time; This represents the maximum allowed processing time for the task. For the first The accuracy of decision-making for each step; For the first An intelligent agent in Response time at any given moment; This represents the maximum permissible response time for a specialized intelligent agent. for The task is in Time priority; This is the highest priority task. This refers to the duration of peak business hours; For system compliance matching degree.

Images