Agent-based tendering business workflow automation arrangement method and system

Abstract

The present application relates to the technical field of intelligent process automation, in particular to a bidding business workflow automation arrangement method and system based on an intelligent agent. The present application realizes accurate conversion of bidding business targets and effective analysis of natural language rules by constructing an intention-driven intelligent agent role model, solving the problem that traditional workflows are difficult to adapt to personalized needs. A cross-system dynamic negotiation framework uses federated learning and consensus algorithms to enable heterogeneous system agents to efficiently collaborate while protecting data privacy and quickly respond to process abnormalities. A digital twin sandbox and reinforcement learning optimize the process path, ensuring efficiency and controllable risk. Blockchain technology ensures compliance traceability and improves process transparency. Overall, the bidding business automation level, efficiency and compliance are improved, and manual intervention is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent process automation technology, specifically to a method and system for automating the orchestration of bidding business workflows based on intelligent agents. Background Technology

[0002] Bidding processes involve multiple parties and complex workflows, such as issuing tender notices, reviewing tender documents, and signing contracts. Traditional workflows typically rely on manual operation and fixed rules, making it difficult to dynamically adapt to the personalized needs of different projects. In recent years, agent technology has provided new ideas for workflow automation through autonomous decision-making and task collaboration capabilities. However, when introducing agent technology into bidding processes, achieving dynamic orchestration across systems and multiple roles remains a technical challenge.

[0003] While existing workflow engines can execute tasks according to preset rules, they lack the proactive scheduling capability for real-time interactions between heterogeneous systems. When sudden changes or abnormal events occur in the bidding process, such as temporary adjustments to evaluation criteria or disputes over bidder qualifications, existing systems can only rely on manual intervention to reconfigure the process, and cannot quickly generate adaptive solutions through negotiation mechanisms between intelligent agents. This passive response mode leads to inefficiency and makes it difficult to ensure the compliance and consistency of process execution, thus limiting the practical application of bidding automation. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an automated orchestration method and system for bidding business workflows based on intelligent agents, which solves the problems of reliance on manual labor, difficulty in dynamically adapting to personalized needs, slow response to anomalies, and weak compliance traceability in existing technologies.

[0005] To achieve the above objectives, the present invention implements an automated orchestration method for bidding business workflow based on intelligent agents through the following technical solution, wherein the automated orchestration method for bidding business workflow includes:

[0006] Constructing an intent-driven agent role model, including:

[0007] The business objectives of each participant in the bidding process are transformed into machine-understandable intent vectors, which are distinct from traditional workflow nodes.

[0008] The intelligent agent dynamically parses natural language rules containing tender document clauses by fine-tuning a domain-specific large language model;

[0009] Generate an executable intent policy tree, which enables real-time encoding of unstructured requirements;

[0010] Design a cross-system dynamic negotiation framework, including:

[0011] Design a distributed negotiation protocol based on federated learning, which allows heterogeneous system agents to engage in multi-round games through encrypted demand proposals and constraint tokens without exposing raw data;

[0012] When process abnormalities occur, such as changes in evaluation criteria or disputes over bidder qualifications, the evaluation intelligence agent generates proposals that include timeliness weights and compliance boundaries.

[0013] The contract management agent responds to the proposal and generates a counter-proposal containing risk warnings regarding legal terms;

[0014] An adaptive solution is reached through a consensus algorithm, and the adaptive solution is used to handle process anomalies.

[0015] A dynamic verification mechanism for building a digital twin sandbox includes:

[0016] Before the process changes take effect, a digital twin sandbox is constructed using ultra-real-time simulation technology. The sandbox is used to simulate the bidding process.

[0017] Historical anomalous data is injected into the sandbox to train adversarial examples;

[0018] Simulate the debate process between the judge's intelligent agent and the auditing intelligent agent on controversial decisions;

[0019] By dynamically optimizing the process path through reinforcement learning, the path ensures that the dual goals of improving efficiency and keeping risks under control are met simultaneously.

[0020] Establish a blockchain-based compliance traceability network, including:

[0021] The key steps of negotiation between intelligent agents are transformed into lightweight zero-knowledge proofs and anchored to a private blockchain. These key steps include the exchange of qualification review opinions and the reasons for score adjustment.

[0022] The process jump is based on the compliance verification and signature verification of the preceding node;

[0023] A timestamped decision chain is automatically generated through smart contracts, and this decision chain is used for audit traceability.

[0024] Furthermore, the construction of the intent-driven agent role model also includes:

[0025] The intent vector is generated in the following way: based on the business objectives of the bidding participants, a predefined intent classification system is used to map the business objectives into numerical intent vectors;

[0026] The rules for the intelligent agent to parse natural language include:

[0027] Semantic analysis of unstructured text rules in bidding documents is performed using the aforementioned domain-specific large language model;

[0028] Identify the business constraints and conditions in the unstructured text rules;

[0029] The business constraints and conditions are transformed into machine-executable logical expressions.

[0030] Furthermore, the design of the cross-system dynamic negotiation framework also includes:

[0031] The evaluation intelligence agent dynamically calculates the timeliness weight and compliance boundary parameters of the proposal based on the current evaluation progress and preset compliance strategy;

[0032] The contract management agent responds to the proposal by conducting a risk assessment of the legal clauses involved and generating a counter-proposal with risk warnings for the legal clauses based on preset rules.

[0033] The consensus algorithm includes a multi-party secure computation protocol, which iteratively negotiates among the heterogeneous system agents until an adaptive solution that satisfies the constraints of all parties is reached.

[0034] Furthermore, the dynamic verification mechanism for constructing the digital twin sandbox also includes:

[0035] Historical anomaly data includes data related to process interruptions, rule conflicts, and decision disputes;

[0036] The training adversarial examples include constructing simulated scenarios using the historical anomalous data, and the scenarios are used to test the robustness of the agent's decision-making.

[0037] The reinforcement learning dynamic optimization process path includes: setting process efficiency and compliance risk as reward function objectives, and adjusting the agent's decision-making strategy through iterative simulation and agent interaction until the reward function reaches its optimum.

[0038] Furthermore, the establishment of a blockchain-based compliance traceability network also includes:

[0039] The zero-knowledge proof is used to prove the validity and compliance of key negotiation steps without disclosing business content;

[0040] A private blockchain is a permissioned blockchain network, in which only authorized smart agent nodes can participate in block writing and verification;

[0041] The deployment of the smart contract is used to automate the execution of compliance verification signatures and the generation of the decision chain;

[0042] The decision chain records all important decision points, participating agent identifiers, and timestamps during the negotiation process.

[0043] Furthermore, the generation of the executable intent policy tree includes:

[0044] Based on the parsed natural language rules, a hierarchical logical structure is constructed, which represents the decision path of the intelligent agent;

[0045] The logical structure nodes are associated with business operations and condition judgments;

[0046] By employing rule engines or decision graph technology, the logical structure is transformed into a set of code or instructions that can be directly executed by the intelligent agent.

[0047] Furthermore, the distributed negotiation protocol design includes:

[0048] Initialize the global model, which includes negotiation strategy parameters;

[0049] Each heterogeneous system agent independently trains its local model based on its own local data and privacy policy;

[0050] The local model only uploads model parameters to the central aggregator, which then aggregates them to form a new global model.

[0051] The new global model is distributed to each heterogeneous system agent, and the agent updates its local model and continues to negotiate and iterate.

[0052] Furthermore, the debate process between the simulated judge agent and the auditing agent regarding controversial decisions includes:

[0053] The digital twin sandbox constructs simulated dispute scenarios, which include conflicting bid evaluation results or contract terms.

[0054] The judge and audit agents conduct multiple rounds of dialogue and argumentation in the simulated scenario according to their preset rules and intentions.

[0055] Key interactive data from the dialogue and arguments are recorded, and this data is used for decision pattern analysis.

[0056] Furthermore, the compliance verification signature check for the process jump includes:

[0057] After the preceding intelligent agent node completes the business processing, it generates a digital signature containing a business data digest and a timestamp.

[0058] The digital signature is encapsulated using zero-knowledge proof technology to ensure that the information is not leaked;

[0059] After receiving the digital signature, the target node of the process jumps to initiate a verification request to the private chain;

[0060] The consensus mechanism of the private blockchain verifies the validity of the signature, and the validity is based on preset compliance rules.

[0061] The present invention also provides an agent-based automated orchestration system for bidding workflows, applied to any of the agent-based automated orchestration methods for bidding workflows described above, the system comprising:

[0062] The intelligent agent role modeling module is used to transform the business objectives of each participant in the bidding into machine-understandable intent vectors, and generate an executable intent policy tree by parsing natural language rules based on a domain-wide large language model.

[0063] The dynamic negotiation module is used to design a distributed negotiation protocol based on federated learning, which allows heterogeneous system agents to engage in multiple rounds of game-playing through encrypted demand proposals and constraint tokens without exposing the original data, and to reach an adaptive solution through a consensus algorithm when the process is abnormal.

[0064] The dynamic verification module is used to build a digital twin sandbox using ultra-real-time simulation technology before process changes take effect, inject historical abnormal data to train adversarial samples, simulate the process of dispute debate, and dynamically optimize the process path through reinforcement learning.

[0065] The compliance traceability module is used to transform key negotiation steps between intelligent agents into lightweight zero-knowledge proofs and anchor them to the private chain, verify the compliance verification signature of process jumps, and automatically generate a decision chain with timestamps through smart contracts.

[0066] Beneficial Effects: This invention, by constructing an intent-driven intelligent agent role model, achieves precise conversion of bidding business objectives and effective parsing of natural language rules, solving the problem that traditional workflows struggle to adapt to personalized needs. The cross-system dynamic negotiation framework, leveraging federated learning and consensus algorithms, enables heterogeneous system agents to collaborate efficiently while protecting data privacy and quickly responding to process anomalies. Digital twin sandboxes and reinforcement learning optimize process paths, ensuring efficiency and risk control. Blockchain technology guarantees compliance traceability and enhances process transparency. Overall, this invention improves the automation level, efficiency, and compliance of bidding processes, reducing manual intervention. Attached Figure Description

[0067] Figure 1 This is a flowchart of the method of the present invention;

[0068] Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Please see Figure 1 This invention provides an automated orchestration method for bidding business workflows based on intelligent agents. The automated orchestration method for bidding business workflows includes:

[0071] Constructing an intent-driven agent role model, including:

[0072] The business objectives of each participant in the bidding process are transformed into machine-understandable intent vectors, which are distinct from traditional workflow nodes.

[0073] The intelligent agent dynamically parses natural language rules containing tender document clauses by fine-tuning a domain-specific large language model;

[0074] Generate an executable intent policy tree, which enables real-time encoding of unstructured requirements;

[0075] Design a cross-system dynamic negotiation framework, including:

[0076] Design a distributed negotiation protocol based on federated learning, which allows heterogeneous system agents to engage in multi-round games through encrypted demand proposals and constraint tokens without exposing raw data;

[0077] When there are procedural anomalies such as changes in evaluation criteria or disputes over bidder qualifications, the evaluation AI agent generates proposals that include timeliness weights and compliance boundaries.

[0078] The contract management agent responds to proposals and generates counter-proposals that include risk warnings regarding legal terms;

[0079] An adaptive solution is reached through a consensus algorithm, and this adaptive solution is used to handle process anomalies.

[0080] A dynamic verification mechanism for building a digital twin sandbox includes:

[0081] Before the process changes take effect, a digital twin sandbox is built using ultra-real-time simulation technology. The sandbox is used to simulate the bidding process.

[0082] Injecting historical anomalous data into the sandbox to train adversarial examples;

[0083] Simulate the debate process between the judge's intelligent agent and the auditing intelligent agent on controversial decisions;

[0084] By dynamically optimizing process paths through reinforcement learning, the paths ensure that the dual goals of improving efficiency and keeping risks under control are achieved simultaneously.

[0085] Establish a blockchain-based compliance traceability network, including:

[0086] The key steps of negotiation between intelligent agents are transformed into lightweight zero-knowledge proofs and anchored to a private blockchain. The key steps include the exchange of qualification review opinions and the reasons for score adjustment.

[0087] The process jump is based on the compliance verification and signature verification of the preceding node;

[0088] A decision chain with timestamps is automatically generated through smart contracts, and this decision chain is used for audit traceability.

[0089] Specifically, this implementation method addresses the problems of traditional bidding workflows, such as reliance on manual operation, difficulty in dynamically adapting to personalized needs, and low efficiency in handling exceptions. Detailed implementation steps are provided as follows:

[0090] When constructing an intent-driven agent role model, the first step is to identify the business objectives of the bidding participants. For example, the bidding party's objective is to select the bidder with the highest cost-effectiveness, the bidder's objective is to successfully win the bid, and the evaluation party's objective is to objectively and fairly review the bid documents. Using a predefined intent classification system, these business objectives are mapped to numerical intent vectors. Each dimension corresponds to a specific business objective parameter, such as price, quality, or timeframe. Intent vectors can more accurately reflect the needs of the participants, unlike the fixed nodes in traditional workflows.

[0091] The intelligent agent dynamically parses natural language rules containing tender document clauses by fine-tuning a domain-specific large language model. A suitable basic large language model for processing legal and business texts is selected and fine-tuned using historical tender documents and review rules data from the tendering domain. This enables the model to understand the specialized terminology and rule logic within the tendering domain. For example, when parsing the clause "bidders must have at least 3 years of experience in similar projects," the model can identify the definition criteria for "similar project experience" and the time constraint of "3 years," and transform them into a machine-understandable logical expression.

[0092] When generating an executable intent policy tree, a hierarchical logical structure is constructed based on parsed natural language rules. Upper-level nodes represent the main business objectives, while lower-level nodes represent the specific steps and conditional judgments for achieving those objectives. For example, for the upper-level node "determine the winning bidder," lower-level nodes could include "qualification review," "technical review," and "business review," etc. Each lower-level node is associated with specific operations and judgment conditions. For instance, the "qualification review" node is associated with operations such as "check the validity of the business license" and "view project experience certificates," as well as the judgment condition "pass if all conditions are met, otherwise fail." Rule engine technology is used to transform this logical structure into code that the agent can directly execute.

[0093] When designing a cross-system dynamic negotiation framework, a distributed negotiation protocol based on federated learning is designed. A global model containing negotiation policy parameters is initialized. Agents from different heterogeneous systems, such as those in a bidding system and a contract management system, independently train their local models based on their local data and privacy policies. Local models only upload their parameter updates to the central aggregator, which then aggregates data to form a new global model using the following formula:

[0094] ,

[0095] in, These are the new global model parameters. These are the global model parameters for round t. Where N is the learning rate, and N is the number of agents. These are the local model parameters of the i-th agent in round t. After the new global model is distributed to each agent, the agents update their local models and continue to negotiate and iterate.

[0096] When a process anomaly occurs due to changes in evaluation criteria, the evaluation agent dynamically calculates the timeliness weight and compliance boundary parameters of the proposal based on the current evaluation progress and preset compliance strategies. For example, if the evaluation deadline is approaching, the timeliness weight is higher. The contract management agent responds to the proposal, conducts a risk assessment of the relevant legal clauses, such as determining whether the clauses comply with the relevant provisions of the "Tendering and Bidding Law," and generates a counter-proposal containing risk warnings regarding the legal clauses. Using a multi-party secure computation protocol as a consensus algorithm, iterative negotiation takes place among the heterogeneous system agents until an adaptive solution that satisfies the constraints of all parties is reached.

[0097] When constructing a dynamic verification mechanism for a digital twin sandbox, a digital twin sandbox is built using ultra-real-time simulation technology before process changes take effect. This sandbox simulates the various stages and behaviors of participants in the bidding process. Historical abnormal data, such as past process interruptions, rule conflicts, and decision disputes, are injected into the sandbox to construct simulated scenarios for training adversarial examples and testing the robustness of the agent's decision-making.

[0098] The simulation demonstrates the debate process between a judge's AI agent and an auditing AI agent on controversial decisions, such as whether a bidder's qualifications meet the requirements. The judge's AI agent presents its viewpoint based on the review rules, while the auditing AI agent raises questions based on compliance standards. The dialogue and argumentation data of both parties are recorded. The process path is dynamically optimized through reinforcement learning, with process efficiency and compliance risk set as reward function objectives. The reward function formula is as follows:

[0099] ,

[0100] Where R is the reward value, and E is a weighting coefficient, and E is a process efficiency indicator, such as process completion time. These are compliance risk indicators, such as the probability of violation. The agent's decision-making strategy is adjusted through iterative simulations and interactions until the reward function reaches its optimal state.

[0101] When establishing a blockchain-based compliance traceability network, key negotiation steps between intelligent agents, such as the exchange of qualification review opinions and the reasons for score adjustments, are transformed into lightweight zero-knowledge proofs and anchored to a private blockchain. The private blockchain adopts a permissioned blockchain network, where only authorized intelligent agent nodes can participate in block writing and verification. Process transitions are verified by the compliance signature of the preceding node, and a timestamped decision chain is automatically generated through smart contracts for subsequent audit traceability.

[0102] The above implementation methods enable automated orchestration of the bidding process, dynamically adapt to the personalized needs of different projects, quickly generate adaptive solutions when anomalies occur, improve work efficiency, and ensure the compliance and consistency of process execution.

[0103] In this embodiment, constructing the intent-driven agent role model further includes:

[0104] Intent vectors are generated as follows: based on the business objectives of the bidding participants, a predefined intent classification system is used to map the business objectives into numerical intent vectors;

[0105] The rules for intelligent agents to parse natural language include:

[0106] Semantic analysis of unstructured text rules in bidding documents is performed using a domain-specific large language model;

[0107] Identify business constraints and conditions in unstructured text rules;

[0108] Transform business constraints and conditions into machine-executable logical expressions.

[0109] Specifically, when constructing an intent-driven agent role model, intent vectors are generated as follows: First, a predefined intent classification system is established, covering various business objectives that may be involved in the bidding process, such as cost control, quality assurance, and schedule management. Then, based on the specific business objectives of the bidding participants, the corresponding categories are selected from the intent classification system and mapped to numerical intent vectors. For example, if the bidder's business objective is "to win the bid with the lowest price," then the dimension corresponding to cost control in the intent vector will have a value of 1, while other dimensions will have values ​​of 0.

[0110] When the intelligent agent parses natural language rules, it uses a domain-specific large language model to perform semantic analysis on the unstructured text rules in the bidding documents. After fine-tuning with the bidding domain corpus, the domain-specific large language model can identify technical terms and grammatical structures in the text. For example, for the rule "Bidding documents must be submitted before the deadline, otherwise they will be considered invalid bids," the model can analyze that "submitting before the deadline" is the condition, and "considered invalid bids" is the result.

[0111] After identifying the business constraints and conditions in the unstructured text rules, they are transformed into machine-executable logical expressions. For example, the above rule is transformed into the logical expression "IF bid submission time > deadline THEN bid invalid", enabling the agent to make decisions and take actions based on this expression.

[0112] This implementation method can accurately translate the business objectives of the bidding participants into a form that the intelligent agent can understand, while also enabling effective parsing of natural language rules, laying the foundation for subsequent automated workflow orchestration.

[0113] In this embodiment, the design of the cross-system dynamic negotiation framework also includes:

[0114] The evaluation intelligence agent dynamically calculates the timeliness weight and compliance boundary parameters of the proposals based on the current evaluation progress and preset compliance strategies.

[0115] The contract management agent responds to proposals, conducts risk assessments on the legal clauses involved, and generates counter-proposals with risk warnings for the legal clauses based on preset rules;

[0116] Consensus algorithms include multi-party secure computation protocols, which iteratively negotiate among agents in heterogeneous systems until an adaptive solution that satisfies the constraints of all parties is reached.

[0117] Specifically, when designing a cross-system dynamic negotiation framework, the evaluation agent dynamically calculates the timeliness weight and compliance boundary parameters of proposals based on the current evaluation progress and preset compliance strategies. For example, if the evaluation work is nearing completion, the timeliness weight can be set to 0.8 to prioritize ensuring the timely completion of the process; if the preset compliance strategy stipulates that bids must not be evaluated below cost, then the lower limit of cost price in the compliance boundary parameters is the estimated cost of the project.

[0118] When responding to a proposal, the contract management agent conducts a risk assessment of the relevant legal clauses. For example, if the payment term clause in the proposal may not conform to industry practice, the contract management agent will assess the potential legal risks of the clause by searching relevant legal provisions and past cases, and, in conjunction with preset risk avoidance rules, generate a counter-proposal that includes risk warnings regarding the legal clauses, such as suggesting that the payment term be adjusted to the industry-standard 30 days.

[0119] The consensus algorithm employs a multi-party secure computation protocol, which involves iterative negotiation among agents in a heterogeneous system. In each round of negotiation, each agent evaluates and provides feedback on the proposal based on its own constraints. The protocol encrypts the feedback from each agent to ensure information security. Through multiple rounds of iteration, an adaptive solution that satisfies the constraints of all parties is reached.

[0120] This implementation enables cross-system agents to fully consider timeliness and compliance during the negotiation process, reach a reasonable adaptive solution through an effective consensus mechanism, and improve the efficiency and accuracy of anomaly handling.

[0121] In this embodiment, the dynamic verification mechanism for constructing a digital twin sandbox also includes:

[0122] Historical anomaly data includes data related to process interruptions, rule conflicts, and decision disputes;

[0123] Training adversarial examples involves constructing simulated scenarios using historical anomalous data, which are then used to test the robustness of the agent's decision-making.

[0124] The reinforcement learning dynamic optimization process path includes: setting process efficiency and compliance risk as reward function objectives, and adjusting the agent's decision-making strategy through iterative simulation and agent interaction until the reward function reaches its optimum.

[0125] Specifically, when constructing a dynamic verification mechanism for a digital twin sandbox, historical anomaly data includes data on process interruptions, rule conflicts, and decision disputes. Examples include records of process interruptions due to system failures during a bidding process, rule conflict cases where different review rules contradict each other, and decision disputes where judges give significantly different scores to the same bidder.

[0126] When training adversarial examples, these historical anomaly data are used to construct simulated scenarios. For example, based on historical data of process interruptions, a simulated scenario of sudden system crash is constructed; based on rule conflict data, a scenario with two contradictory evaluation criteria is constructed. These scenarios are used to test the robustness of the agent's decision-making in the face of anomalies.

[0127] When using reinforcement learning to dynamically optimize process paths, process efficiency and compliance risk are set as the objectives of the reward function. Process efficiency can be measured by the process completion time, and compliance risk can be represented by the probability of violation. The agent's decision-making strategy is adjusted through iterative simulation and agent interaction. For example, in a simulated scenario, the agent tries different process paths, calculates the reward value based on the process completion time and violation probability under each path, and continuously adjusts the strategy to optimize the reward function.

[0128] This implementation method can effectively test the decision-making ability of intelligent agents under abnormal conditions and optimize process paths to ensure that the process is both efficient and compliant.

[0129] In this embodiment, establishing a blockchain-based compliance traceability network further includes:

[0130] Zero-knowledge proofs are used to demonstrate the validity and compliance of key negotiation steps without revealing business details.

[0131] A private blockchain is a permissioned blockchain network, in which only authorized smart agent nodes can participate in block writing and verification;

[0132] The deployment of smart contracts is used to automate the execution of compliance verification signatures and the generation of decision chains;

[0133] The decision chain records all important decision points, participating agent identifiers, and timestamps during the negotiation process.

[0134] Specifically, when establishing a blockchain-based compliance traceability network, zero-knowledge proofs are used to demonstrate the validity and compliance of key negotiation steps without disclosing business details. For example, in the exchange of qualification review opinions, zero-knowledge proofs can demonstrate that the qualification review has been conducted in accordance with the prescribed procedures and standards, without having to disclose specific review details.

[0135] Private blockchains employ permissioned blockchain networks, where only authorized smart agent nodes can participate in block writing and verification. For example, only authorized nodes such as bidding smart agents, evaluation smart agents, and auditing smart agents can perform data writing and verification operations on the blockchain, ensuring the security and reliability of blockchain data.

[0136] The deployment of smart contracts is used to automate the execution of compliance verification signatures and the generation of decision chains. When a smart agent completes a task at a certain stage, the smart contract automatically verifies the generated signature. Upon successful verification, the decision information for that stage is written into the decision chain. The decision chain records all important decision points in the negotiation process, the identifiers of participating smart agents, and timestamps. For example, the decision point of a bidder's qualification review being approved includes the identifiers of the participating evaluation smart agents and the time of approval.

[0137] This implementation method enables traceability of the bidding process while protecting the privacy of the business content and ensuring the authenticity and integrity of the traceability information.

[0138] In this embodiment, generating an executable intent policy tree includes:

[0139] Based on parsed natural language rules, a hierarchical logical structure is constructed, which represents the decision path of the intelligent agent;

[0140] Logical structure nodes are associated with business operations and condition judgments;

[0141] By employing rule engines or decision graph technology, the logical structure is transformed into a set of code or instructions that the agent can directly execute.

[0142] Specifically, when generating an executable intent policy tree, a hierarchical logical structure is constructed based on parsed natural language rules. This logical structure represents the agent's decision-making path, with upper-level nodes representing the main business objectives and lower-level nodes representing the specific operations and conditional judgments required to achieve those objectives. For example, the upper-level node might be "determine the winning bidder," and lower-level nodes might include "qualification review," "technical review," and "business review," with the "qualification review" node further divided into sub-nodes such as "check business license" and "view project experience."

[0143] Logical structure nodes are associated with business operations and conditional judgments. For example, the business operation associated with the "Check Business License" node is to query business license information, and the conditional judgment is whether the business license is valid.

[0144] By employing rule engines or decision graph technology, logical structures are transformed into code or instruction sets that can be directly executed by the agent. For example, a rule engine can transform the logic "If the business license is valid, the qualification review is passed; otherwise, it is not passed" into corresponding code, which the agent can execute to complete the decision-making process in the qualification review stage.

[0145] The intent policy tree generated in this way enables the agent to clearly execute various business operations according to the decision path, improving the accuracy and efficiency of workflow execution.

[0146] In this embodiment, the distributed negotiation protocol design includes:

[0147] Initialize the global model, which contains negotiation strategy parameters;

[0148] Each heterogeneous system agent independently trains its local model based on its own local data and privacy policy;

[0149] Local models only upload model parameters to the central aggregator, which then aggregates them to form a new global model;

[0150] The new global model is distributed to agents in each heterogeneous system, and the agents update their local models and continue to negotiate and iterate.

[0151] Specifically, when designing a distributed negotiation protocol, the first step is to initialize a global model, which contains negotiation strategy parameters, such as the weight of each agent in the negotiation and the priority of the negotiation.

[0152] Each heterogeneous system agent independently trains its local model based on its own local data and privacy policies. For example, the evaluation agent uses its own historical evaluation data and pre-set evaluation rules to train a local model without disclosing its specific local evaluation data during the training process.

[0153] Local models only upload parameter updates to the central aggregator. The aggregator processes the parameter updates from all local models using an aggregation algorithm to form a new global model. For example, the aggregator uses a weighted average method, assigning different weights based on the importance of each agent, to aggregate the parameter updates.

[0154] The new global model is distributed to agents in various heterogeneous systems, which then update their local models and continue negotiating and iterating. Through multiple rounds of model training and aggregation, the negotiation strategies of each agent are continuously optimized, improving the efficiency and effectiveness of the negotiation.

[0155] This implementation enables intelligent agents in heterogeneous systems to conduct effective collaborative negotiation while protecting data privacy, thereby enhancing the ability to collaborate across systems.

[0156] In this embodiment, the debate process between the simulating judge agent and the auditing agent on controversial decisions includes:

[0157] Digital twin sandboxes are used to simulate dispute scenarios, including conflicting bid evaluation results or contract terms.

[0158] The judge and audit agents conduct multiple rounds of dialogue and argumentation in a simulated scenario based on their preset rules and intentions.

[0159] Record key interactive data from dialogues and arguments; this data is used for decision pattern analysis.

[0160] Specifically, when simulating the debate process between the judges' intelligent agents and the auditing intelligent agents on controversial decisions, the digital twin sandbox constructs simulated dispute scenarios. For example, the scenario is set as follows: a bidder's technical solution partially meets the requirements and partially does not, leading to a disagreement among the judges' intelligent agents in the evaluation results, or there are ambiguities in the contract terms regarding liability for breach of contract.

[0161] The judging agent and the auditing agent engage in multiple rounds of dialogue and argumentation within a simulated scenario, based on their pre-set rules and intentions. The judging agent explains its scoring rationale for the bidders according to the evaluation criteria, while the auditing agent raises questions and challenges to the judges' scores based on compliance requirements. For example, the judging agent might believe that the bidder's technological innovation is highly advanced and deserves a high score; the auditing agent, however, might point out that this technological innovation is not explicitly reflected in the relevant documents and does not comply with the evaluation process.

[0162] Record key interactive data from dialogues and arguments, such as the arguments, evidence, and rules of citation by both parties. This data is used to analyze the decision-making patterns of the intelligent agent and provide a basis for optimizing the agent's decision-making capabilities.

[0163] By simulating the debate process, we can identify potential problems in the agent's decision-making, further improve the agent's decision-making logic, and enhance the rationality and accuracy of the decisions.

[0164] In this embodiment, the compliance verification signature check for process jump includes:

[0165] After the preceding intelligent agent node completes the business processing, it generates a digital signature containing a business data summary and a timestamp.

[0166] Digital signatures are encapsulated using zero-knowledge proof technology to ensure that information is not leaked;

[0167] After receiving the digital signature, the target node of the process jumps to initiate a verification request to the private blockchain;

[0168] The consensus mechanism of a private blockchain verifies the validity of a signature, and the validity is based on preset compliance rules.

[0169] Specifically, during the compliance verification signature check for process jumps, after the preceding intelligent agent node completes its business processing, it generates a digital signature containing a business data summary and a timestamp. For example, after the qualification review intelligent agent completes the qualification review of a bidder, it signs the data summary of the review result and the review completion time.

[0170] Digital signatures are encapsulated using zero-knowledge proof technology to ensure that specific information about the business content is not disclosed during transmission and verification; they can only prove the validity of the signature.

[0171] Upon receiving the digital signature, the target node in the process jumps to the private blockchain and initiates a verification request. The private blockchain's consensus mechanism verifies the validity of the signature based on preset compliance rules, such as whether the signature was generated by an authorized node and whether the timestamp is within a valid range.

[0172] Only verified signatures are allowed to redirect the process. This implementation method ensures the compliance and security of process redirection and prevents illegal operations and data tampering.

[0173] Please see Figure 2 The present invention also provides an agent-based automated orchestration system for bidding workflows, applicable to any of the above-mentioned agent-based automated orchestration methods for bidding workflows, the system comprising:

[0174] The intelligent agent role modeling module is used to transform the business objectives of each participant in the bidding into machine-understandable intent vectors, and generate an executable intent policy tree by parsing natural language rules based on a domain-wide large language model.

[0175] The dynamic negotiation module is used to design a distributed negotiation protocol based on federated learning, which allows heterogeneous system agents to engage in multiple rounds of game-playing through encrypted demand proposals and constraint tokens without exposing the original data, and to reach an adaptive solution through a consensus algorithm when the process is abnormal.

[0176] The dynamic verification module is used to build a digital twin sandbox using ultra-real-time simulation technology before process changes take effect, inject historical abnormal data to train adversarial samples, simulate the process of dispute debate, and dynamically optimize the process path through reinforcement learning.

[0177] The compliance traceability module is used to transform key negotiation steps between intelligent agents into lightweight zero-knowledge proofs and anchor them to the private chain, verify the compliance verification signature of process jumps, and automatically generate a decision chain with timestamps through smart contracts.

[0178] Specifically, the agent role modeling module transforms the business objectives of each participant in the bidding process into machine-understandable intent vectors using a pre-defined intent classification system. Simultaneously, it utilizes a domain-specific large language model to parse the natural language rules in the bidding documents, extracting business constraints and conditions to generate an executable intent policy tree, providing a basis for the agent's decision-making.

[0179] The dynamic negotiation module is responsible for designing a distributed negotiation protocol based on federated learning. This protocol allows agents from heterogeneous systems to engage in multi-round game theory using encrypted demand proposals and constraint tokens without exposing raw data. When anomalies occur in the process, such as changes in evaluation criteria, this module prompts the relevant agents to reach an adaptive solution through a consensus algorithm, ensuring the smooth progress of the process.

[0180] The dynamic verification module constructs a digital twin sandbox using ultra-real-time simulation technology before process changes take effect. Historical anomaly data is injected into the sandbox to train adversarial examples, simulating the debate process between a judge agent and an audit agent on controversial decisions. Through reinforcement learning, the process path is dynamically optimized, ensuring that the process meets both efficiency improvement requirements and risk control.

[0181] The compliance traceability module transforms key negotiation steps between intelligent agents into lightweight zero-knowledge proofs and anchors them to a private blockchain. During process transitions, it verifies the compliance signatures of preceding nodes and automatically generates a timestamped decision chain through smart contracts, providing a reliable basis for auditing and tracing bidding processes.

[0182] The modules work together to automate the workflow of bidding, improving the efficiency and compliance of bidding processes.

[0183] In summary, this invention, by constructing an intent-driven intelligent agent role model, achieves precise conversion of bidding business objectives and effective parsing of natural language rules, solving the problem that traditional workflows struggle to adapt to personalized needs. The cross-system dynamic negotiation framework, leveraging federated learning and consensus algorithms, enables heterogeneous system agents to collaborate efficiently while protecting data privacy and quickly responding to process anomalies. Digital twin sandboxes and reinforcement learning optimize process paths, ensuring efficiency and risk control. Blockchain technology guarantees compliance traceability and enhances process transparency. Overall, this invention improves the automation level, efficiency, and compliance of bidding processes, reducing manual intervention.

[0184] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0185] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automated orchestration method for bidding business workflow based on intelligent agents, characterized in that, The automated orchestration method for the bidding workflow includes: Constructing an intent-driven agent role model, including: The business objectives of each participant in the bidding process are transformed into machine-understandable intent vectors, which are distinct from traditional workflow nodes. The intelligent agent dynamically parses natural language rules containing tender document clauses by fine-tuning a domain-specific large language model; Generate an executable intent policy tree, which enables real-time encoding of unstructured requirements; Design a cross-system dynamic negotiation framework, including: Design a distributed negotiation protocol based on federated learning, which allows heterogeneous system agents to engage in multi-round games through encrypted demand proposals and constraint tokens without exposing raw data; When process abnormalities occur, such as changes in evaluation criteria or disputes over bidder qualifications, the evaluation intelligence agent generates proposals that include timeliness weights and compliance boundaries. The contract management agent responds to the proposal and generates a counter-proposal containing risk warnings regarding legal terms; An adaptive solution is reached through a consensus algorithm, and the adaptive solution is used to handle process anomalies. A dynamic verification mechanism for building a digital twin sandbox includes: Before the process changes take effect, a digital twin sandbox is constructed using ultra-real-time simulation technology. The sandbox is used to simulate the bidding process. Historical anomalous data is injected into the sandbox to train adversarial examples; Simulate the debate process between the judge's intelligent agent and the auditing intelligent agent on controversial decisions; By dynamically optimizing the process path through reinforcement learning, the path ensures that the dual goals of improving efficiency and keeping risks under control are met simultaneously. Establish a blockchain-based compliance traceability network, including: The key steps of negotiation between intelligent agents are transformed into lightweight zero-knowledge proofs and anchored to a private blockchain. These key steps include the exchange of qualification review opinions and the reasons for score adjustment. The process jump is based on the compliance verification and signature verification of the preceding node; A timestamped decision chain is automatically generated through smart contracts, and this decision chain is used for audit traceability.

2. The automated orchestration method for bidding business workflow based on intelligent agents according to claim 1, characterized in that, The construction of the intent-driven agent role model also includes: The intent vector is generated in the following way: based on the business objectives of the bidding participants, a predefined intent classification system is used to map the business objectives into numerical intent vectors; The rules for the intelligent agent to parse natural language include: Semantic analysis of unstructured text rules in bidding documents is performed using the aforementioned domain-specific large language model; Identify the business constraints and conditions in the unstructured text rules; The business constraints and conditions are transformed into machine-executable logical expressions.

3. The method for automated orchestration of bidding workflow based on intelligent agents according to claim 1, characterized in that, The cross-system dynamic negotiation framework also includes: The evaluation intelligence agent dynamically calculates the timeliness weight and compliance boundary parameters of the proposal based on the current evaluation progress and preset compliance strategy; The contract management agent responds to the proposal by conducting a risk assessment of the legal clauses involved and generating a counter-proposal with risk warnings for the legal clauses based on preset rules. The consensus algorithm includes a multi-party secure computation protocol, which iteratively negotiates among the heterogeneous system agents until an adaptive solution that satisfies the constraints of all parties is reached.

4. The automated orchestration method for bidding business workflow based on intelligent agents according to claim 1, characterized in that, The dynamic verification mechanism for constructing the digital twin sandbox also includes: Historical anomaly data includes data related to process interruptions, rule conflicts, and decision disputes; The training adversarial examples include constructing simulated scenarios using the historical anomalous data, and the scenarios are used to test the robustness of the agent's decision-making. The reinforcement learning dynamic optimization process path includes: setting process efficiency and compliance risk as reward function objectives, and adjusting the agent's decision-making strategy through iterative simulation and agent interaction until the reward function reaches its optimum.

5. The automated orchestration method for bidding business workflow based on intelligent agents according to claim 1, characterized in that, The establishment of a blockchain-based compliance traceability network also includes: The zero-knowledge proof is used to prove the validity and compliance of key negotiation steps without disclosing business content; A private blockchain is a permissioned blockchain network, in which only authorized smart agent nodes can participate in block writing and verification; The deployment of the smart contract is used to automate the execution of compliance verification signatures and the generation of the decision chain; The decision chain records all important decision points, participating agent identifiers, and timestamps during the negotiation process.

6. The automated orchestration method for bidding business workflow based on intelligent agents according to claim 2, characterized in that, The generation of the executable intent policy tree includes: Based on the parsed natural language rules, a hierarchical logical structure is constructed, which represents the decision path of the intelligent agent; The logical structure nodes are associated with business operations and condition judgments; By employing rule engines or decision graph technology, the logical structure is transformed into a set of code or instructions that can be directly executed by the intelligent agent.

7. The method for automated orchestration of bidding workflow based on intelligent agents according to claim 3, characterized in that, The distributed negotiation protocol design includes: Initialize the global model, which includes negotiation strategy parameters; Each heterogeneous system agent independently trains its local model based on its own local data and privacy policy; The local model only uploads model parameters to the central aggregator, which then aggregates them to form a new global model. The new global model is distributed to each heterogeneous system agent, and the agent updates its local model and continues to negotiate and iterate.

8. The automated orchestration method for bidding business workflow based on intelligent agents according to claim 4, characterized in that, The debate process between the simulated judge agent and the auditing agent on controversial decisions includes: The digital twin sandbox constructs simulated dispute scenarios, which include conflicting bid evaluation results or contract terms. The judge and audit agents conduct multiple rounds of dialogue and argumentation in the simulated scenario according to their preset rules and intentions. Key interactive data from the dialogue and arguments are recorded, and this data is used for decision pattern analysis.

9. The method for automated orchestration of bidding workflow based on intelligent agents according to claim 5, characterized in that, The compliance verification signature check for the process jump includes: After the preceding intelligent agent node completes the business processing, it generates a digital signature containing a business data digest and a timestamp. The digital signature is encapsulated using zero-knowledge proof technology to ensure that the information is not leaked; After receiving the digital signature, the target node of the process jumps to initiate a verification request to the private chain; The consensus mechanism of the private blockchain verifies the validity of the signature, and the validity is based on preset compliance rules.

10. An agent-based automated orchestration system for bidding workflows, applied to the agent-based automated orchestration method for bidding workflows as described in any one of claims 1-9, characterized in that, The system includes: The intelligent agent role modeling module is used to transform the business objectives of each participant in the bidding into machine-understandable intent vectors, and generate an executable intent policy tree by parsing natural language rules based on a domain-wide large language model. The dynamic negotiation module is used to design a distributed negotiation protocol based on federated learning, which allows heterogeneous system agents to engage in multiple rounds of game-playing through encrypted demand proposals and constraint tokens without exposing the original data, and to reach an adaptive solution through a consensus algorithm when the process is abnormal. The dynamic verification module is used to build a digital twin sandbox using ultra-real-time simulation technology before process changes take effect, inject historical abnormal data to train adversarial samples, simulate the process of dispute debate, and dynamically optimize the process path through reinforcement learning. The compliance traceability module is used to transform key negotiation steps between intelligent agents into lightweight zero-knowledge proofs and anchor them to the private chain, verify the compliance verification signature of process jumps, and automatically generate a decision chain with timestamps through smart contracts.

Images