Agent development and collaborative management system and execution method based on proxy arrangement

Abstract

The invention discloses an agent development and collaborative management system based on proxy arrangement, which comprises an application layer, a capability layer and a base layer, and is characterized in that the application layer is used for bearing agent applications oriented to different service scenes; the capability layer comprises a knowledge base subsystem which is used for constructing a document knowledge base and a question and answer knowledge base and converting an unstructured document into a structured knowledge graph; the workflow subsystem is internally provided with a self-adaptive arrangement engine and a visual process arrangement interface and is used for carrying out graphical design and operation on the multi-agent cooperation process; the agent subsystem is used for creating a plurality of agents based on natural language configuration, model integration and tool calling; and the base layer is used for providing infrastructure and management functions for system operation. The method is used for solving the problems that in the prior art, a multi-agent system is high in development complexity, low in cooperation efficiency, lack of a unified management platform, opaque in cooperation process, poor in adaptability and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an agent development and collaborative management system based on proxy orchestration and an execution method. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, agents (Agents) as intelligent entities capable of perceiving the environment, making decisions and performing actions have shown great application potential in automation, intelligent customer service, robotic process automation (RPA) and other fields. However, a single agent often has limited functions and is difficult to cope with complex and changing real-world tasks. Therefore, multi-agent systems (Multi-Agent System, MAS) have emerged.

[0003] However, the prior art still has the following deficiencies in the development and management of multi-agent systems: 1) The traditional development mode based on code programming has the problem of high development threshold: the existing development of multi-agent systems mainly relies on traditional programming methods, such as using Python, Java, etc. programming language combined with specific multi-agent framework (such as JADE, SPADE, etc.) for development. This development mode requires developers not only to master programming languages, but also to deeply understand complex technologies such as distributed systems, artificial intelligence algorithms, communication protocols, etc. Developers need to manually write a large number of agent communication code, coordination logic and error handling mechanisms, with long development cycle and high maintenance cost. In addition, business personnel are difficult to directly participate in the development process, resulting in a large communication cost and understanding deviation between business requirements and technical implementation.

[0004] 2) Lack of unified agent orchestration mechanism leads to low collaboration efficiency: in the prior art, the collaboration between agents is mainly realized through point-to-point communication or simple message passing mechanism, lacking a unified, dynamic orchestration and scheduling mechanism. The execution order of agents, resource allocation, task decomposition, etc. mostly adopt static, pre-defined workflow mode, which is prone to resource conflicts, deadlocks, repeated execution of tasks, etc. When the system scale expands and the task complexity increases, the interaction relationship between agents becomes extremely complex, making it difficult to effectively optimize and coordinate globally. In addition, the existing scheme lacks dynamic adjustment capability and cannot automatically adjust the agent orchestration strategy according to the real-time system state and task demand, resulting in low overall efficiency of the system and poor robustness in the face of node failure or performance bottleneck.

[0005] 3) Lack of enterprise-level unified management and monitoring capabilities: Most existing agent development systems focus on the construction of single agents and simple multi-agent interaction, and lack enterprise-level unified management and monitoring capabilities. In actual applications, enterprises often need to deploy and manage hundreds or even thousands of agents, and existing technologies are difficult to provide effective life cycle management, version control, performance monitoring, fault diagnosis and other functions. The deployment, upgrade, rollback and other operations of the agent need manual intervention, and the operation and maintenance cost is high. At the same time, there is a lack of unified monitoring interface and alarm mechanism, it is difficult to discover and handle system abnormalities in time, and it affects the continuity and stability of business. In addition, the collaboration process is not transparent, and there is a lack of end-to-end traceability and explainability, making it difficult to meet the requirements of enterprises in audit and compliance. SUMMARY

[0006] The purpose of the present application is to provide an agent-based orchestration intelligent agent development and collaborative management system and execution method to solve the problems of high complexity, low collaboration efficiency, lack of unified management platform, non-transparent collaboration process, and poor adaptability of existing multi-agent system development.

[0007] To achieve the above purpose, the following technical solutions are adopted: An agent-based orchestration intelligent agent development and collaborative management system, comprising an application layer, a capability layer, and a foundation layer, wherein, The application layer is used to carry intelligent agent applications for different business scenarios; The capability layer comprises A knowledge base subsystem is used to build a document knowledge base and a question and answer knowledge base, and to convert unstructured documents into structured knowledge graphs; A workflow subsystem with a built-in adaptive orchestration engine and a visual process orchestration interface is used to graphically design and run multi-agent collaboration processes; An agent subsystem is used to create multiple agents based on natural language configuration, model integration and tool invocation; A tool subsystem is used to manage built-in tools and custom tools, and provides external service calling interfaces; An MCP service subsystem is used to integrate external business service interfaces; A semantic interoperability layer is used for semantic alignment; The foundation layer is used to provide infrastructure and management functions for system operation.

[0008] Further, the adaptive orchestration engine comprises a monitor unit, a policy unit, an executor unit and a loop unit; the monitor unit is configured to collect the performance indicators of each agent node and tool node and compare them with preset multi-objective optimization constraints; the policy unit is configured to generate or adjust the agent orchestration strategy based on the orchestration experience base, the task intent and the current system state; the executor unit is configured to drive the target agent, tool and model channel to perform the task according to the agent orchestration strategy; and the loop unit is configured to trigger online re-orchestration of a local subgraph when the performance indicators exceed a threshold, isolate the abnormal node and generate an alternative execution path, so as to complete adaptive adjustment and self-healing of the process without interrupting the overall task.

[0009] Further, the document knowledge base in the knowledge base subsystem is configured to: slice and segment manage the documents according to at least one of a paragraph, a semantic and a fixed length; construct a knowledge graph based on entity recognition and relation extraction to support complex knowledge reasoning and correlation query; the question and answer knowledge base in the knowledge base subsystem is configured to support batch import and online creation of question and answer pairs, and match and answer user questions based on semantic similarity matching technology; The tool subsystem includes a candidate channel pool and an adaptive routing module; the candidate channel pool is configured to store a plurality of model channels and tool channels; the adaptive routing module dynamically balances between cost and quality in the candidate channel pool based on an online multi-armed bandit algorithm according to a preset strategy to select a target model channel or tool channel, wherein for a task set as a key decision node, the adaptive orchestration engine is configured to call multiple model channels in parallel, and after obtaining multiple outputs, perform consistency voting, when the consistency is lower than a preset threshold, trigger self-reflection and knowledge base review process to modify the initial output result.

[0010] Further, the semantic interoperation layer comprises an agent capability ontology module configured to uniformly model the capabilities of each domain agent based on a role-intent-capability-constraint four-element model; a message semantic contract module configured to add header information to the messages transmitted between agents and between agents and tools, wherein the header information includes a task intent label, a dependent resource identifier, a security level and an expected performance indicator; a cross-domain alignment module configured to establish a term mapping relationship between different business domains to realize semantic alignment and interoperation of multi-domain knowledge.

[0011] Further, the basic layer comprises a model management module for integrating multiple AI models and performing default model setting and multi-channel management; a fine-tuning module for task-specific fine-tuning of target models based on a preset dataset; a user and permission management module for managing users, roles, and user groups, and performing permission control at four granularities of agents, tools, knowledge items, and data fields; a system operation support module for providing system configuration, log recording, mobile support, and interfacing capabilities, and auditing and monitoring the execution process; a heterogeneous scheduling and self-healing module for selecting deployment domains for each target agent among the cloud, edge, and terminal according to data localization requirements, network latency, and computing resource conditions, implementing cloud-edge-terminal affinity scheduling, and when detecting failure or performance degradation of a target node, performing at least one of the following self-healing actions: automatic fuse, downgrade to a preset template, switch to an alternative agent.

[0012] Further, the heterogeneous scheduling and self-healing module is further configured to allocate computing and invocation resources to different tenants based on queue weights and resource throttling mechanisms in a multi-tenant environment, and trigger corresponding degradation strategies when the throttling threshold is reached.

[0013] Further, the system operation support module includes an audit log submodule and an interpretable trajectory graph submodule; the audit log submodule is used to sign and record audit entries for each agent orchestration decision and message passing; the interpretable trajectory graph submodule is used to generate an interpretable trajectory graph based on a visualized orchestration graph, and when a user selects any node or edge in the flowchart, an explanation information containing at least one of the following information is popped up: the knowledge items and tool outputs relied on by the node when making decisions; the model type, prompt content, and key parameters invoked by the node; the input data and output data summary of the node; the user identification and timestamp triggering the execution of the node, and the corresponding audit log entry identification.

[0014] Further, the application layer includes the following business modules: an AI finance module for at least one of the following functions: reimbursement review, invoice review, single return reconciliation, financial report generation, and tax filing assistance; an AI consultation and question answering module for at least one of the following functions: knowledge base-based intelligent question answering and contract intelligent review; an AI content review module for performing industry content analysis and review; an AI report generation module for automatically generating business reports based on data and templates; An AI recruitment module for performing intelligent resume screening.

[0015] An agent orchestration execution method based on agent orchestration, using the above system, comprising the following steps: S1: The workflow subsystem receives the multi-agent workflow task created by the user through the visual process orchestration interface, parses the task, identifies the agent nodes, tool nodes and knowledge base calling nodes in the workflow, and determines the initial logical connection relationship; S2: The policy unit generates an initial agent orchestration strategy containing the agent execution sequence, parallelism and data transmission path based on the initial logical connection relationship, the overall intention of the task and the pre-set orchestration experience library, and during the execution of the agent orchestration strategy, when a model or tool needs to be called, the policy unit selects the most suitable model instance from the candidate channel pool based on the online multi-armed bandit algorithm; S3: According to the initial agent orchestration strategy, the target agent is instantiated, and the semantic contract header field information is attached to each agent and its interaction message through the semantic interoperation layer, and the semantic alignment between agents in different fields is realized by using the capability ontology and cross-domain term mapping; S4: The executor unit drives each agent node and tool node to execute the task according to the agent orchestration strategy, and the monitor unit collects the performance indicators of each node in real time and compares them with the threshold values in the pre-set multi-objective optimization constraints; S5: When any performance indicator exceeds the corresponding threshold value, the loop unit triggers local subgraph online re-orchestration, isolates the target node that appears abnormal, and requires the policy unit to generate an alternative execution path for the corresponding subtask, wherein the alternative execution path includes at least one of switching to an alternative model channel, an alternative tool channel or an alternative agent; S6: After completing the execution of all nodes, the multi-agent execution results are summarized, and for the tasks marked as key decision nodes, multiple model channels are used for parallel reasoning and consistent voting is performed, and when the consistency of the voting results is lower than the pre-set threshold, a self-reflection and knowledge base review process is triggered to correct the initial results; S7: After the task execution is completed, the agent orchestration strategy and the corresponding running results that meet the pre-set performance requirements are written into the orchestration experience library, and the orchestration decisions and message transmission in the execution process are generated with signed audit logs, which are used for subsequent explainable trajectory graph generation and compliance audit.

[0016] A computer readable storage medium stores a computer program, the computer program is suitable for being loaded and executed by a processor, so that a computer device with the processor executes the above method.

[0017] With the above scheme, the application has the following advantages: 1) The application innovatively proposes a visual drag-and-drop intelligent agent arrangement development mode, and combines an adaptive "agent arrangement" algorithm. Through a graphical user interface, a user can drag and drop intelligent agent components with different functions to a canvas, and define the cooperation relationship and data flow direction between intelligent agents through a connection line. Unlike a static workflow, the arrangement engine has an online rearrangement and multi-objective optimization mechanism. Through a "monitor-strategist-executor-circuit" four-unit closed loop, the arrangement engine collects performance indicators (such as delay, cost, and accuracy) of each node in real time, and dynamically selects an optimal execution path based on a multi-objective scheduler (using a heuristic and reinforcement learning hybrid strategy). When the performance exceeds a preset threshold (for example, the timeout time is greater than 1.5 times the historical P95 value), the system can trigger online rewriting and replacement of a local subgraph, realize incremental rollback, and ensure the robustness and efficiency of the process without interrupting the overall task. This development mode greatly reduces the development threshold of a multi-agent system, and realizes a leap from "static preset" to "dynamic adaptation" through algorithm innovation. 2) The application innovatively proposes an agent cooperation mechanism based on agent arrangement, and ensures the depth and accuracy of cooperation through an independent "semantic interoperability layer". The arrangement engine adopts a layered agent architecture (global arrangement agent, local coordination agent, and execution agent), and the semantic interoperability layer provides a unified "language" for cooperation. The layer includes three cores: an agent capability ontology, which unifies a four-element model of "role-intent-capability-restriction"; a message semantic contract, which adds a header field containing "task intent label, dependent resource, security level, and expected index" to each message, and a cross-domain alignment mechanism, which breaks through the terminology barriers between different fields such as finance, human resources, and maps through ontology mapping. This design enables multi-agent cooperation to be upgraded from simple "speaking" (protocol layer communication) to "understanding the same semantics" (semantic layer understanding), which is a key difference from traditional frameworks, and significantly improves the cooperation efficiency and accuracy of complex cross-domain tasks. 3) The application innovatively constructs an enterprise-level agent full-life-cycle management system, and integrates a trusted collaboration and explainability mechanism; the system provides full-process management capabilities from agent development, testing, deployment, operation to maintenance, supports enterprise-level functions such as version management, gray release, automatic scaling, etc.; further, the application establishes an end-to-end traceable system, signs and records audit logs for each orchestration decision and message passing, ensures the security of the system through four-level fine-grained permission authorization of "Agent / tool / knowledge item / data field", and generates a visual "explainable trajectory graph" to clearly show "which Agent made the decision based on what knowledge, used what tools, and with what parameters", greatly improving the transparency and credibility of the collaboration process and meeting the strict requirements of enterprises for compliance and audit; 4) In the use of models and tools, the application proposes an adaptive routing mechanism to achieve dynamic balance between cost and quality; the system constructs a candidate channel pool containing different large models, different inference parameters, different suppliers and corresponding cost / delay characteristics; during task execution, the system uses an online Multi-Armed Bandit selection strategy to dynamically allocate traffic and retry according to preset rules such as "fast first, accurate later" or "cheap first, fine later", and for key decision nodes, the system also enables a consensus voting mechanism, i.e. parallel invocation of multiple models, only when the consistency of the outputs of multiple models is higher than a certain threshold (θ) will it pass, otherwise it will trigger a "Self-Refine" and knowledge base review process, this innovation enables the system to optimize costs or maximize quality under the premise of guaranteeing service level agreement (SLA), providing quantifiable technical and economic benefits for enterprises; 5) To adapt to complex enterprise IT environments, the application integrates support for heterogeneous environments and multi-tenant scenarios in the system architecture; the system has cloud-edge-end affinity scheduling capabilities, can intelligently determine the deployment domain of the agent according to factors such as data's territorial requirements and delay sensitivity, and at the same time, the system establishes a standardized fault self-healing closed loop process, including automatic fusing, downgrading to a preset template, switching to an alternative Agent, and finally gray recovery, for multi-tenant environments, the system realizes quality of service (QoS) guarantee through queue weight and resource throttling mechanisms, ensuring that operations between tenants do not interfere with each other, these designs and the system's original automatic scaling, unified monitoring, etc. functions resonate at the algorithm level of scheduling and self-healing, greatly improving the stability, reliability and resource utilization efficiency of the system. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of the method of the application. DETAILED DESCRIPTION

[0019] The application will be described in detail below in combination with the drawings and specific embodiments.

[0020] The application provides an agent-based intelligent agent development and collaborative management system, which comprises an application layer, a capability layer and a foundation layer in an embodiment, wherein The application layer is used for carrying intelligent agent applications for different business scenarios. The capability layer comprises A knowledge base subsystem is used for constructing a document knowledge base and a question and answer knowledge base, and converting unstructured documents into a structured knowledge graph. A workflow subsystem is used for graphically designing and running a multi-agent collaboration process, and comprises an adaptive arrangement engine and a visual process arrangement interface. An intelligent agent subsystem is used for creating multiple intelligent agents based on natural language configuration, model integration and tool calling. A tool subsystem is used for managing built-in tools and custom tools, and providing an external service calling interface. An MCP service subsystem is used for integrating external business service interfaces. A semantic interoperability layer is used for semantic alignment. The foundation layer is used for providing infrastructure and management functions for system running.

[0021] In the embodiment, the system is a complete system composed of multiple specialized subsystems and working collaboratively, which not only comprises three-layer architecture of the application layer, the capability layer and the foundation layer, but also introduces an innovative mechanism between the capability layer and the foundation layer to realize dynamic, reliable and efficient intelligent agent collaboration, and the architecture design is as follows: For the application layer, the following business modules are included: An AI finance module is used for realizing at least one function of reimbursement auditing, invoice auditing, single return reconciliation, financial report generation and tax filing assistance. An AI consulting and question answering module is used for realizing at least one function of knowledge base-based intelligent question answering and contract intelligent review. An AI content auditing module is used for performing industry content analysis and auditing. An AI report generation module is used for automatically generating business reports according to data and templates. An AI recruitment module is used for performing intelligent resume screening.

[0022] In this embodiment, the application layer is a user-oriented service application layer, which provides special intelligent agent applications according to different service scenarios. The AI financial module integrates the core functions of enterprise financial management, including intelligent reimbursement auditing, invoice auditing, single return reconciliation, financial statement generation, and tax assistance, and automatically processes a large number of repetitive financial work through intelligent agents; the AI consulting and answering module provides intelligent question and answer services and contract intelligent review functions based on a knowledge base, helping enterprises improve consulting service efficiency; the AI content auditing module has industry analysis capabilities and can intelligently audit and analyze various types of content; the AI report generation module can automatically generate various types of business reports according to data and templates; and the AI recruitment module provides a resume intelligent screening function to improve human resource management efficiency.

[0023] As for the capability layer, it is the core technology layer of the platform, which provides strong technical support capabilities for the application layer, specifically: The document knowledge base in the knowledge base subsystem is configured to: slice and segment manage the documents according to at least one of a paragraph, a semantic, and a fixed length slicing strategy; construct a knowledge graph based on entity recognition and relationship extraction to support complex knowledge reasoning and associated queries; the question and answer knowledge base in the knowledge base subsystem is configured to support batch import and online creation of question and answer pairs, and match and answer user questions based on semantic similarity matching technology.

[0024] The knowledge base subsystem adopts a double-layer knowledge base architecture, with the document knowledge base and the QA knowledge base (question and answer knowledge base) complementing each other. The document knowledge base adopts advanced slicing strategies, supporting multiple slicing methods such as by paragraph, by semantic, and by fixed length, ensuring moderate granularity of knowledge; the segment management function allows users to structure the documents and establish chapter hierarchy relationships; the knowledge graph module converts unstructured documents into structured knowledge representations through entity recognition, relationship extraction, and other technologies, supporting complex knowledge reasoning and associated queries; the QA knowledge base supports batch import and online creation of question and answer pairs, and realizes accurate question and answer matching through semantic similarity matching technology.

[0025] The workflow subsystem (adaptive orchestration engine) is the orchestration core of the platform, provides a visual process design interface, and is embedded with adaptive orchestration algorithms, supports version management, tool calling, knowledge base calling, prompt word management and other functions, especially its multi-agent collaboration capability and MCP (Model Context Protocol) calling capability, under the support of semantic interoperation layer, realizes the automation and intelligent orchestration of complex business processes, preferably, the adaptive orchestration engine includes a monitor unit, a policy unit, an executor unit and a loop unit; the monitor unit is used to collect the running performance indicators of each intelligent agent node and tool node and compare them with the preset multi-objective optimization constraints; the policy unit is used to generate or adjust intelligent agent orchestration strategies based on an orchestration experience base, a task intent and a current system state; the executor unit is used to drive target intelligent agents, tools and model channels to perform tasks according to the intelligent agent orchestration strategies; the loop unit is used to trigger online re-orchestration of a local subgraph when the performance indicators exceed a threshold, isolate nodes with abnormalities and generate an alternative execution path, so as to complete adaptive adjustment and self-healing of the process without interrupting the overall task.

[0026] The intelligent agent subsystem supports creating intelligent agents through natural language, integrates open AI capabilities, and provides rich tool calling and MCP calling interfaces.

[0027] The tool subsystem includes a candidate channel pool and an adaptive routing module; the candidate channel pool is used to store various model channels and tool channels; the adaptive routing module performs dynamic trade-off between cost and quality in the candidate channel pool according to a preset strategy based on an online multi-armed bandit algorithm, to select a target model channel or tool channel, wherein for a task set as a key decision node, the adaptive orchestration engine is configured to call multiple model channels in parallel, and after obtaining multiple outputs, perform consistency voting, when the consistency is lower than a preset threshold, trigger self-reflection and knowledge base review processes to modify the initial output result.

[0028] The tool subsystem is divided into built-in tools and custom tools, the built-in tools include OCR recognition, text-to-speech, network search, enterprise WeChat docking, email pushing, query and other commonly used functions; the custom tool supports users to add and configure special tools according to needs, and its core is an adaptive routing mechanism, which can dynamically select the optimal model or tool channel.

[0029] The MCP service subsystem provides external service integration capabilities, including a shared center, related software, a third-party service interface of a treasury supervision, and a newly added semantic interoperability layer, which is a key to realizing high-level collaboration. The semantic interoperability layer ensures that all agents and tools interact under a unified semantic framework through capability ontology, semantic contract, and cross-domain alignment. The semantic interoperability layer includes An agent capability ontology module, configured to uniformly model capabilities of agents in different fields based on a role-intent-capability-restriction four-element model; A message semantic contract module, configured to add header information to messages transmitted between agents and between agents and tools, wherein the header information includes a task intent label, a dependent resource identifier, a security level, and an expected performance indicator; A cross-domain alignment module, configured to establish a term mapping relationship between different business domains to realize semantic alignment and interoperability of multi-field knowledge.

[0030] The semantic interoperability layer is used to realize capability alignment. It defines a unified agent capability ontology, specifies a message semantic contract, and supports cross-domain term alignment to ensure that agents in different fields can effectively communicate at the same semantic level.

[0031] In an embodiment, for the basic layer, it includes A model management module, configured to integrate multiple AI models and perform default model setting and multi-channel management; A fine-tuning module, configured to perform task customization fine-tuning on a target model based on a preset data set; A user and permission management module, configured to manage users, roles, and user groups, and perform permission control at four granularities of agents, tools, knowledge entries, and data fields; A system running support module, configured to provide system configuration, log recording, mobile terminal support, and interfacing capabilities, and to audit and monitor the execution process; A heterogeneous scheduling and self-healing module, configured to select a deployment domain for each target agent among the cloud, edge, and terminal according to data locality requirements, network delay, and computing resource conditions, to realize cloud-edge-terminal affinity scheduling, and to perform at least one of the following self-healing actions when detecting that a target node fails or performance degrades: automatic fuse, downgrade to a preset template, and switch to an alternative agent.

[0032] In this embodiment, the base layer provides infrastructure and management functions for stable operation of the platform, wherein the model management module is responsible for system integration, default model setting and multi-channel management of the AI model, ensuring reasonable calling of the model in different business scenarios; the fine-tuning module provides dataset management, fine-tuning task management and model fine-tuning functions, and supports customized training of the model according to specific business needs; the user and permission management module provides complete user lifecycle management and realizes fine-grained permission control; the system running support module provides infrastructure functions such as mobile terminal adaptation, system configuration, system interfacing and running log, and integrates trusted collaboration and self-healing mechanisms, including end-to-end traceable audit logs, interpretable trajectory graph generation, and intelligent scheduling and fault self-healing supporting heterogeneous environments, to ensure the stability, reliability, safety and compliance of the platform.

[0033] Further, the heterogeneous scheduling and self-healing module is further configured to allocate computing and calling resources for different tenants based on queue weights and resource throttling mechanisms in a multi-tenant environment, and trigger corresponding degradation strategies when the throttling threshold is reached; the system running support module includes an audit log submodule and an interpretable trajectory graph submodule; the audit log submodule is used to sign and record audit entries for each agent orchestration decision and message passing; the interpretable trajectory graph submodule is used to generate an interpretable trajectory graph based on a visualized orchestration graph, and when a user selects any node or edge in the flowchart, an explanation information containing at least one of the following information is popped up: knowledge entries and tool outputs relied on by the node when making decisions; model type, prompt content and key parameters called by the node; summary of input data and output data of the node; user identification and timestamp triggering the execution of the node and corresponding audit log identification.

[0034] In this embodiment, in addition to the conventional mobile terminal support, system configuration, interfacing and log functions, the system running support module also includes end-to-end traceable audit logs (signed for each operation) and the ability to generate interpretable trajectory graphs. At the same time, it has heterogeneous / multi-tenant scheduling and self-healing mechanisms, supports cloud-edge-end affinity scheduling and fault self-healing closed loop (fusing, degradation, replacement, recovery), and guarantees the quality of service of multi-tenant through queue weights and throttling.

[0035] In another embodiment, an agent orchestration execution method based on agent orchestration is also provided, which adopts the above system and includes the following steps: S1: The workflow subsystem receives a multi-agent workflow task created by a user through a visual flow orchestration interface, parses the task, identifies agent nodes, tool nodes and knowledge base calling nodes in the workflow, and determines the initial logical connection relationship; S2: The policy unit generates an initial agent orchestration strategy containing the execution order of agents, parallelism, and data transmission path based on the initial logical connection relationship, task overall intention, and pre-set orchestration experience library. When a model or tool needs to be called during the execution of the agent orchestration strategy, the policy unit selects the most suitable model instance from the candidate channel pool based on the online multi-armed bandit algorithm.

[0036] In this step, the policy unit of the orchestration engine intervenes. It first analyzes the overall intention of the task (by analyzing its structure or with the help of NLP models) and searches for an efficient scheduling strategy for similar tasks from the "orchestration experience library" for hot start. If there is no match, an initial execution strategy is generated based on the user-designed flowchart and node connection relationship, including the execution order of agents, parallelism, and preliminary data transmission path. At the same time, the policy unit will use an adaptive routing mechanism at each step of execution. For example, when a large model needs to be called, it will select the most suitable model instance (considering cost, delay, and expected accuracy) from the candidate channel pool according to the online multi-armed bandit strategy. This process is dynamic, not static binding.

[0037] S3: According to the initial agent orchestration strategy, instantiate the target agent, and through the semantic interoperability layer, attach semantic contract header field information to each agent and its interaction message. Use the capability ontology and cross-domain term mapping to achieve semantic alignment between agents in different domains.

[0038] In this step, according to the orchestration strategy, the system instantiates the relevant agents. At this time, the semantic interoperability layer plays a role, attaching semantic contracts to each Agent and message to ensure that they communicate under a unified capability ontology, and the MCP invocation interface is configured to enable seamless collaboration across domains (such as finance and human resources).

[0039] S4: The executor unit drives each agent node and tool node to execute the task according to the agent orchestration strategy, while the monitor unit collects performance indicators of each node in real time and compares them with the threshold values in the pre-set multi-objective optimization constraints.

[0040] In this step, the executor unit is responsible for driving the agents to execute the tasks as planned. At the same time, the monitor unit collects performance indicators of each node in real time, such as delay, success rate, Token consumption, cost, accuracy score, etc., and compares them with the pre-set threshold values (for example, timeout threshold T = 1.5 x historical P95 delay, failure rollback number ≤ 2).

[0041] S5: When any performance indicator exceeds the corresponding threshold, the loop unit triggers the local subgraph online rearrangement, isolates the target node that appears abnormal, and requires the policy unit to generate an alternative execution path for the corresponding subtask, wherein the alternative execution path includes at least one of switching to an alternative model channel, an alternative tool channel, or an alternative agent.

[0042] When the monitor finds that a certain indicator is out of limit (for example, a certain tool call continuously fails or the delay is too high), the loop unit is triggered immediately. The loop performs the "local subgraph online rewriting" logic: it isolates the problem node and requires the policy to re-generate an alternative path for the subtask (for example, switch to a backup tool, or use a more reliable but slower model). This replacement is incremental and does not affect other branches that are executing normally. The entire process realizes the rapid self-healing of faults.

[0043] S6: After completing the execution of all nodes, the multi-agent execution results are summarized, for tasks marked as key decision nodes, multiple model channels are used for parallel inference and consistent voting is performed, when the consistency of the voting result is lower than the preset threshold, the self-reflection and knowledge base review process is triggered to correct the initial result.

[0044] After the task execution is completed, the system summarizes the execution results of each agent. For key nodes, parallel execution of N>1 models may be enabled, at which time consistent voting is performed, and if the result consistency is lower than the threshold θ, the "self-refine" mechanism is triggered to allow the agent to combine the knowledge base for self-correction, and the final confirmed result is pushed to the user or external system through the external release function.

[0045] S7: After the task execution is completed, the agent scheduling strategy that meets the preset performance requirement and the corresponding running result are written into the scheduling experience library, and signed audit logs are generated for each scheduling decision and message passing in the execution process, for subsequent explainable trajectory graph generation and compliance audit.

[0046] If the entire process is completed efficiently and stably, the successful scheduling strategy (including path selection, parallelism setting, etc.) will be stored in the "scheduling experience library" for future reuse of similar tasks. Finally, the system performs version management records, releases the occupied resources, and performs related cleaning work.

[0047] In addition, another embodiment further provides a computer readable storage medium storing a computer program, and the computer program is suitable for being loaded and executed by a processor, so as to enable a computer device having the processor to execute the above method. The computer readable storage medium can be an internal storage unit of the computer device, for example, a hard disk or a memory of the computer device. The computer readable storage medium can also be an external storage device of the computer device, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card and the like equipped on the computer device. Further, the computer readable storage medium can include both the internal storage unit and the external storage device of the computer device. The computer readable storage medium is used to store the computer program and other programs and data required by the computer device. The computer readable storage medium can also be used to temporarily store data that has been output or will be output.

[0048] The above only is the preferred embodiment of the present application, and is not used to limit the present application, and any modification, equivalent replacement and improvement etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A smart agent development and collaborative management system based on agent-based orchestration, characterized in that, It includes the application layer, capability layer, and infrastructure layer, among which, The application layer is used to support intelligent agent applications for different business scenarios; The capability layer includes The knowledge base subsystem is used to build document knowledge bases and question-answering knowledge bases, and to convert unstructured documents into structured knowledge graphs; The workflow subsystem has a built-in adaptive orchestration engine and a visual process orchestration interface, which is used for the graphical design and operation of multi-agent collaborative processes. The intelligent agent subsystem is used to create multiple intelligent agents based on natural language configuration, model integration, and tool invocation. The tools subsystem manages built-in and custom tools and provides external service call interfaces. The MCP service subsystem is used to integrate external business service interfaces; The semantic interoperability layer is used for semantic alignment. The base layer is used to provide the infrastructure and management functions for system operation.

2. The agent development and collaborative management system based on agent-based orchestration according to claim 1, characterized in that, The adaptive orchestration engine includes a monitor unit, a strategy unit, an executor unit, and a looper unit. The monitor unit is used to collect the operational performance indicators of each agent node and tool node and compare them with preset multi-objective optimization constraints. The strategy unit is used to generate or adjust agent orchestration strategies based on the orchestration experience base, task intent, and current system state. The actuator unit is used to drive the target agent, tools, and model channels to perform tasks according to the agent orchestration strategy. The looper unit is used to trigger online reordering of the local subgraph when the performance index exceeds the threshold, isolate abnormal nodes and generate alternative execution paths, thereby completing the adaptive adjustment and self-healing of the process without interrupting the overall task.

3. The agent development and collaborative management system based on agent-based orchestration according to claim 2, characterized in that, The document knowledge base in the knowledge base subsystem is configured as follows: Slice and segment the document using at least one of the following slicing strategies: by paragraph, by semantics, and by fixed length; Knowledge graphs are constructed based on entity recognition and relation extraction to support complex knowledge reasoning and related queries; The question-and-answer knowledge base in the knowledge base subsystem is configured to support batch import and online creation of question-and-answer pairs, and to match and answer user questions based on semantic similarity matching technology. The tool subsystem includes a candidate channel pool and an adaptive routing module. The candidate channel pool stores various model channels and tool channels. The adaptive routing module, based on an online multi-armed slot machine algorithm, dynamically weighs cost and quality in the candidate channel pool according to a preset strategy to select a target model channel or tool channel. For tasks designated as key decision nodes, the adaptive orchestration engine is configured to call multiple model channels in parallel and perform a consensus vote after obtaining multiple outputs. When the consensus is lower than a preset threshold, a self-reflection and knowledge base review process is triggered to correct the initial output results.

4. The agent development and collaborative management system based on agent-based orchestration according to claim 1, characterized in that, The semantic interoperability layer includes The Agent Capability Ontology module is used to uniformly model the capabilities of intelligent agents in various domains based on the four-element model of role-intent-capability-constraint. The message semantic contract module is used to add header information to messages transmitted between agents and between agents and tools. The header information includes task intent tags, dependent resource identifiers, confidentiality levels, and expected performance metrics. The cross-domain alignment module is used to establish terminology mapping relationships between different business domains to achieve semantic alignment and interoperability of multi-domain knowledge.

5. The agent development and collaborative management system based on agent-based orchestration according to claim 1, characterized in that, The base layer includes The model management module is used to integrate multiple AI models and perform default model settings and multi-channel management; The fine-tuning module is used to perform task-specific fine-tuning of the target model based on a preset dataset; The User and Permission Management module is used to manage users, roles, and user groups, and to control permissions at four granularities: agents, tools, knowledge entries, and data fields. The system operation support module is used to provide system configuration, logging, mobile terminal support and integration capabilities, and to audit and monitor the execution process; The heterogeneous scheduling and self-healing module is used to select deployment domains for each target intelligent agent among the cloud, edge and terminal based on data location requirements, network latency and computing resources, to achieve cloud-edge-terminal affinity scheduling, and to perform at least one of the following self-healing actions when a target node fails or its performance degrades: automatic circuit breaking, downgrading to a preset template, or switching to an alternative intelligent agent.

6. The agent development and collaborative management system based on agent-based orchestration according to claim 5, characterized in that, The heterogeneous scheduling and self-healing module is also configured to allocate computing and calling resources to different tenants in a multi-tenant environment based on queue weights and resource rate limiting mechanisms, and to trigger corresponding degradation strategies when the rate limiting threshold is reached.

7. The agent development and collaborative management system based on agent-based orchestration according to claim 6, characterized in that, The system operation support module includes an audit log submodule and an interpretable trajectory graph submodule. The audit log submodule is used to sign and record audit entries for each intelligent agent orchestration decision and message transmission. The interpretable trajectory graph submodule is used to generate an interpretable trajectory graph based on the visual orchestration graph. When the user selects any node or edge in the flowchart, an explanation message containing at least one of the following information pops up: The knowledge items and tool outputs that nodes rely on when making decisions; The model type, prompt words, and key parameters invoked by the node; A summary of the node's input and output data; The user identifier and timestamp that triggered the execution of this node, as well as the corresponding audit log entry identifier.

8. The agent development and collaborative management system based on agent-based orchestration according to claim 1, characterized in that, The application layer includes the following business modules: The AI ​​finance module is used to perform at least one of the following functions: expense reimbursement review, invoice review, receipt reconciliation, financial statement generation, and tax filing assistance. The AI ​​consultation and question-answering module is used to realize at least one of the functions of knowledge base-based intelligent question answering and intelligent contract review; The AI ​​content moderation module is used to perform industry content analysis and moderation. The AI ​​report generation module is used to automatically generate business reports based on data and templates; The AI ​​recruitment module is used to perform intelligent resume screening.

9. A method for orchestrating and executing intelligent agents based on proxy-based orchestration, employing the system described in any one of claims 2 to 8, characterized in that, Includes the following steps: S1: The workflow subsystem receives multi-agent workflow tasks created by users through a visual process orchestration interface, parses the tasks, identifies agent nodes, tool nodes and knowledge base call nodes in the workflow, and determines the initial logical connection relationship. S2: Based on the initial logical connection relationship, the overall task intent, and the pre-set orchestration experience library, the strategy unit generates an initial agent orchestration strategy that includes the agent execution order, parallelism, and data transmission path. During the execution of the agent orchestration strategy, when it is necessary to call a model or tool, the strategy unit selects the most suitable model instance from the candidate channel pool based on the online multi-armed slot machine algorithm. S3: Instantiate the target agent according to the initial agent orchestration strategy, and attach semantic contract header information to each agent and its interaction messages through the semantic interoperability layer. Use capability ontology and cross-domain term mapping to achieve semantic alignment between agents in different domains. S4: While the actuator unit drives each agent node and tool node to perform tasks according to the agent orchestration strategy, the monitor unit collects the performance indicators of each node in real time and compares them with the thresholds in the preset multi-objective optimization constraints. S5: When any performance indicator exceeds the corresponding threshold, the looper unit triggers online re-arrangement of the local subgraph, isolates the abnormal target node, and requires the strategy unit to generate an alternative execution path for the corresponding subtask. The alternative execution path includes switching to at least one of the alternative model channel, alternative tool channel, or alternative agent. S6: After all nodes have been executed, the execution results of the multi-agent system are summarized. For tasks marked as key decision nodes, multiple model channels are used for parallel reasoning and consensus voting. When the consensus of the voting results is lower than the preset threshold, a self-reflection and knowledge base review process is triggered to correct the initial results. S7: After the task is completed, the intelligent agent orchestration strategy that meets the preset performance requirements and the corresponding running results are written into the orchestration experience base. Signed audit logs are generated for each orchestration decision and message passing during the execution process for subsequent interpretable trajectory graph generation and compliance audit.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded and executed by a processor, such that a computer device having the processor performs the method of claim 9.

Images