Search

CN-122021703-A - Multi-agent contract cost evaluation integrated treatment system and method based on mode priority

CN122021703ACN 122021703 ACN122021703 ACN 122021703ACN-122021703-A

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a multi-intelligent system control and management technology. Aiming at the technical problems that a multi-agent system lacks a unified contract, cost control and evaluation mechanism under a resource limited scene, so that behavior is unpredictable, cost is out of control and audit is difficult, the multi-agent contract cost evaluation integrated treatment system and method based on mode priority are provided. The system comprises a mode definition module, a budget perception management module, a structural evaluation module and a control closed loop, wherein the mode definition module is used for defining the identity, progress and evaluation modes of an intelligent agent to standardize a data structure, contract verification service is used as a management valve to execute structure, authority and budget verification, the budget perception management module generates cost constraint in a decision through multi-layer budget constraint modeling and optimization, and the structural evaluation module generates quality scores and risk labels to drive rework or approval decisions to realize management closed loop. The system is suitable for resource-constrained organizations, and realizes high autonomy, cost controllability and auditable management of the multi-agent system.

Inventors

  • XU KAITAO

Assignees

  • 徐凯韬

Dates

Publication Date
20260512
Application Date
20260130

Claims (10)

  1. 1. A multi-agent contract cost evaluation integrated abatement system based on mode prioritization, comprising: (1) The system comprises a mode definition module, an agent progress mode, an agent evaluation mode, a data processing module and a data processing module, wherein the mode definition module is used for defining at least three types of unified data modes, namely an agent identity mode, an agent progress mode, an agent evaluation mode and an agent evaluation mode, wherein the agent identity mode is used for describing identity information, roles, authority levels and calculated upper limit parameters of an agent; (2) The contract verification service is used as a governance valve before the input and output of all the agents are written into the system state and is configured to execute the structural verification based on the data mode, the permission and override detection combined with the identity mode of the agents, the budget rationality verification aiming at the cost-related field, the circular dependency detection aiming at the workflow topology; (3) The budget perception treatment module comprises a budget monitoring submodule and a decision-making agent, wherein the budget monitoring submodule is used for aggregating and calculating the budget usage and the budget usage of a task level, a project level and a company level based on an agent progress mode, and the decision-making agent is used for generating agent allocation, model gear selection, task priority and decision whether manual approval is needed or not based on a budget state and an agent evaluation mode; (4) The system comprises a structured evaluation module, a decision-making agent and a budget monitoring sub-module, wherein the structured evaluation module is used for calling an evaluation agent to perform multidimensional evaluation on intermediate results and final results, generating a structured evaluation report conforming to an agent evaluation mode, and feeding back the evaluation results to the decision-making agent and the budget monitoring sub-module to form closed loop management, wherein the system is deployed in a multi-agent application scene with limited resources, and the cost controllability, the quality assurance and the auditable behavior of the multi-agent system are realized through the cooperation of the mode definition module, the contract verification service, the budget perception management module and the structured evaluation module.
  2. 2. The abatement system of claim 1, wherein the budget aware abatement module further forms the multi-agent collaboration process into a constrained markov decision process model, comprising: (1) The state space is defined to be composed of progress, cost and risk information of each task and project, which are obtained by aggregation according to the progress mode of the agent, task level, project level and company level budget using state, which are obtained by calculation according to the budget monitoring submodule, agent role, authority level and preset upper limit abstract, which are obtained by compression according to the identity mode of the agent; (2) Defining an action space to comprise at least one action of selecting and executing an allocation action of an agent for a task to be processed, selecting an action of calling a model gear for the selected agent, adjusting the priority of a task queue, and deciding whether to require manual approval or not; (3) Defining a reward function to consist of an overall quality score and a cost normalization term in the agent assessment mode to balance quality and cost; (4) Taking the hyperbranched amounts of the task level budget, the project level budget and the company level budget as a plurality of constraint functions in the constrained Markov decision process respectively; (5) And internalizing each constraint function into an objective function by adopting a Lagrange relaxation method, and online adjusting Lagrange multipliers corresponding to each constraint by adopting an approximate gradient updating mode based on budget use conditions, so that long-term accumulated rewards are optimized on the premise of meeting preset budget constraint targets.
  3. 3. The abatement system of claim 1 or 2, wherein the contract validation service is specifically configured to perform the steps of: (1) Receiving an object to be verified from a target intelligent agent and a corresponding data pattern type identifier, wherein the object to be verified comprises at least one of a document, a configuration, a workflow description or a progress record; (2) Performing structural verification on the object to be verified based on the data pattern, and generating a rejection result and returning an error reason when a field is missing, the type is not matched or the pattern constraint is violated; (3) Inquiring a corresponding identity mode of an intelligent agent according to the identity of a target intelligent agent, judging whether the authority set of the intelligent agent allows the intelligent agent to generate or modify resources or fields related in the object to be verified, and generating a refusing result and returning authority error information when an override operation is detected; (4) When the object to be verified is a progress record or a budget report, budget rationality verification is carried out on the predicted cost field according to the calculated upper limit of the corresponding task, project or intelligent agent, and when the predicted cost exceeds a preset proportion threshold value, a reject result is generated and budget exception information is returned; (5) When the object to be verified is a workflow description, performing topology analysis on nodes and dependency relations of the workflow, and if a circularly dependent or impermissible call path is found, generating a refusing result and returning workflow error information; (6) Only when the structure check, the authority check, the budget rationality check and the workflow topology check pass, an acceptance result is generated and the object to be verified is allowed to be written into the system state or the subsequent execution is triggered.
  4. 4. The abatement system of claim 1 or 2, wherein the workflow of the structured assessment module comprises: (1) Selecting a preset evaluation dimension set for different types of output objects, including but not limited to a requirement document, architecture design, source code, test report and online decision document; (2) Based on the evaluation dimension construction evaluation prompt information, invoking an evaluation agent to score the performance of the output object in each evaluation dimension and generating a text comment: (3) Packaging the scores of all evaluation dimensions and the corresponding comments into a structured evaluation report conforming to the evaluation mode of the agent, and calculating an overall quality score field; (4) Extracting risk labeling information from the structured assessment report, marking high risk items related to data security, legal compliance, business interruption and the like, and generating corresponding improvement suggestion fields; (5) When the overall quality score is lower than a preset threshold value and/or high risk labeling exists, a reworking or manual approval suggestion is output to the decision-making agent so as to trigger a reworking process or a manual approval process of the related task; (6) The overall quality score is used as a part of a reward function of the constrained Markov decision process, and risk labeling and reworking information are fed back to a budget monitoring sub-module and a decision-making agent so as to dynamically adjust budget allocation, model gear selection and manual intervention strategies in subsequent tasks.
  5. 5. The abatement system of claim 1, further comprising a three-tier organization architecture: (1) The human decision layer is used for setting a global budget target, a quality target and a risk preference and carrying out final approval on a high-risk or high-influence task; (2) The management agent layer comprises the budget monitoring sub-module, the decision agent, the contract verification service and the structural evaluation module, and is used for scheduling and controlling the execution agent layer under unified contract and budget constraint; (3) The intelligent agent management system comprises an intelligent agent management layer, a human decision layer and a decision layer, wherein the intelligent agent management layer comprises a plurality of intelligent agent management layers for carrying specific tasks such as demand analysis, architecture design, code realization, test execution, operation and maintenance, marketing and the like, and the intelligent agent management layer is used for carrying out unified management and audit on task allocation, resource use and output quality of the intelligent agent management layer based on roles, rights and preset upper limits recorded in an intelligent agent identity mode, and the human decision layer is only used for carrying out intervention at preset high-risk nodes through an artificial approval interface initiated by the decision intelligent agent.
  6. 6. The abatement system of claim 1, wherein the agent identity model comprises at least the following fields of agent identity, role type, belonging hierarchy, set of rights, security level, pre-calculated upper limit parameters, and delegation relationship identity with an upper level agent or human decision maker.
  7. 7. The abatement system of claim 1, wherein the agent progress pattern comprises at least a task identification, an item identification, an execution agent identification, an execution status, an actual cost field, an expected cost field, a risk level field, an approval status field, and an association identification with a corresponding assessment report.
  8. 8. The abatement system of claim 1, wherein the agent evaluation pattern comprises at least the following fields of an evaluation object type, an evaluation dimension list, each evaluation dimension score, an overall quality score, a risk annotation list, an improvement suggestion field, and a recommendation flag that triggers rework or manual approval.
  9. 9. The abatement system of claim 1 or 2, wherein the decision-making agent dynamically adjusts the model gear and task priority of the selected execution agent according to the budget usage level and the risk marking level, wherein when the budget usage of any level exceeds the corresponding pre-warning threshold, the model gear of the non-critical task is preferentially downgraded and the task priority thereof is reduced, and when the risk marking level in the assessment report is high or extremely high, the corresponding task is forcedly marked as requiring manual approval and automatic online is prohibited.
  10. 10. The abatement system of claim 1 or 2, wherein the budget monitor submodule calculates in real-time task-, project-and company-level budget usage and generates a hierarchical alert signal, when the budget usage exceeds a first threshold, a second threshold and a third threshold, respectively, generating normal, early warning and severe warning levels, respectively, the decision agent limiting or adjusting the reception, queuing, model gear selection and whether new projects are allowed for subsequent tasks based on the warning levels.

Description

Multi-agent contract cost evaluation integrated treatment system and method based on mode priority Technical Field The invention relates to the technical field of artificial intelligence, in particular to a multi-intelligent system management technology, and more particularly relates to a management system and a management method for realizing contract management, cost constraint and quality assessment closed loop integration in a multi-intelligent system based on a mode priority mechanism, which are suitable for software development and operation scenes with limited resources, such as a software company, a lightweight software team and the like. Background With the popularization of large language models and multi-agent orchestration frameworks, more and more software development and operation work is completed by multiple agent cooperations. Under the extreme resource limited scene of a software company and the like, an creator needs to maintain the iteration speed, the delivery quality and the compliance safety of the product under the limited large language model budget and the single task cost constraint. In such a scenario, the multi-intelligent system generally faces the following prominent technical problems: And (one) lack of unified formalized contracts. Most of the existing multi-agent systems describe agent responsibilities through free text prompt and loose role setting, and restrict multi-class objects such as a requirement document, a framework description, codes, a test report, a budget report, an evaluation result and the like, which lack a unified structured data pattern family. The fields, structures and semantics of the outputs of different agents are inconsistent, a large number of redundant prompts are needed by the downstream agents to carry out fault-tolerant inference and repair, and comprehensive audit and responsibility tracing on the system behaviors are difficult. And (two) cost and budget are not inherent in the decision making process. The large language model call cost directly affects cash flow, belonging to hard constraint. However, the existing method mostly regards the cost as a post-monitoring index instead of a core element in strategy planning, lacks a multi-layer budget unified modeling mechanism from a task level to a company level, and lacks a technical means for internally considering budget limitation and cost fluctuation in a decision stage through a constraint decision process and other methods. And (III) the quality and risk assessment lacks a structured closed loop. The multi-agent collaboration process often includes multiple stages such as demand analysis, architecture design, code implementation, test verification, and online decision making. The existing system either relies on manual step-by-step inspection or adopts coarse-granularity test pass/fail signals, lacks a unified structured evaluation mode to bear soft indexes such as user journey coverage, structural definition, risk exposure points and the like, and does not link an evaluation result with task reworking, authority adjustment, model gear selection and manual approval to form a closed loop. (IV) limitations of the related art. For example, the multi-agent system token consumption optimization method mainly performs early stop optimization at the round level based on the task success probability and token consumption relation, belongs to local speaking round number optimization, does not establish unified contracts from the system level, does not perform multi-level modeling and management on task level, project level and company level budgets, the multi-agent based adaptive budget report generation method mainly faces to a budget report scene, generates a budget report through multi-agent cooperation and performs quality scoring on the report model, does not bring actual large language model calling cost into a multi-agent decision process, and the multi-agent based alarm management method focuses on alarm question-answering and knowledge retrieval workflow optimization, lacks unified data mode constraint, and does not relate to closed-loop management of cost constraint and structured quality assessment. In summary, the prior art cannot meet the comprehensive requirements of a multi-agent system in terms of contract consistency, biochemistry within budget constraints and quality assessment closed-loop treatment under resource-limited scenes such as a software company. Disclosure of Invention The invention mainly aims to solve the following technical problems of the existing multi-intelligent system: 1. the input and output of multiple agents lack of uniform formalized contracts, which results in inconsistent structure and difficult audit of semantics; 2. cost and budget constraints are not built in multi-agent decisions, so that the budget constraint satisfaction rate is low and the cost fluctuation is large; 3. the quality and risk assessment lacks structural expression and closed-loop use, a