Interruption step arrangement intelligence analysis multi-agent cooperation method and system

Abstract

The present application relates to a kind of interruptable step scheduling intelligence analysis multi-agent cooperation method, system and storage medium, belong to computer information processing technical field, the method includes: according to task initial object generation cooperation task step chart;Each execution step is converted into standardization step description;Build execution context;For each execution step, set step state visual control unit;After each step is executed, intermediate result is obtained and intermediate product package is generated;When artificial intervention occurs, determine the affected step set;The execution step in the affected step set is re-executed;Based on intermediate product package, output final task result.The present application makes upstream result by intermediate product package, which can be directly reused by downstream step, while supporting to view and modify data after pausing at any step, determine which downstream steps are affected by current modification, and only re-run the affected steps, rather than re-execute the entire task, thereby significantly reducing resource consumption and improving execution efficiency.

Description

Technical Field

[0001] This invention belongs to the field of computer information processing technology. Specifically, this invention relates to a multi-agent collaborative method, system, and storage medium for intelligence analysis with interruptible step scheduling. Background Technology

[0002] In intelligence analysis, analysts typically need to complete a full workflow: first, define the objectives and timeframe; then, collect relevant data; extract and filter the data; finally, form an evidence set; and lastly, generate intelligence briefings, thematic analyses, or question-and-answer conclusions. With the continuous enrichment of intelligent agent models such as large-scale models, retrieval engines, graph engines, data collection engines, and report generation engines, each step in the workflow can be undertaken by different intelligent agent models.

[0003] The current system primarily employs two implementation methods. The first is the "function menu-based" approach, where data collection, extraction, retrieval, question answering, and reporting each exist as independent intelligent agent models. Users need to manually switch between multiple pages, progressively copying the results of each step to the next. The second is the "black box intelligent agent" approach, where after the user inputs a single command, the system automatically runs multiple internal agents, but the intermediate processes are invisible, cannot be stopped, and cannot be modified; the user can only passively wait for the final result.

[0004] The closest existing technologies to this invention are mostly workflow engines or agent orchestration systems. While these solutions support chaining multiple steps together, they generally suffer from the following problems:

[0005] First, the system lacks sufficient interruptibility at the step level. If a user discovers that the suggestion is inappropriate in the integration step, that the recall scope is too broad in the retrieval step, or that the template selection is incorrect in the reporting step, the existing system often has no choice but to stop the entire task and start running from scratch, unable to retain the confirmed upstream results.

[0006] Second, there is a lack of a unified intermediate product structure. The output formats of different steps vary, making it difficult to accurately determine which downstream steps are affected by the current changes. Therefore, it is difficult to achieve local re-execution by "only re-running the affected steps".

[0007] Third, there is a lack of a unified execution state machine. Many systems only record the overall state of a task (such as running, completed, or failed), without fine-grained management of the states of each step, such as pending execution, running, paused, completed, invalidated, or pending rerun.

[0008] Fourth, insufficient auditability. The system failed to organize the step diagram version, parameter snapshot, prompt word version, agent model version, and input / output references into a unified playback object, making it difficult to locate result deviations.

[0009] Fifth, existing business capabilities are easily duplicated. The intelligent agent models for data collection, extraction, retrieval, and synthesis are implemented again at the collaboration layer, resulting in high maintenance costs.

[0010] Further analysis reveals that most existing technologies only focus on "linking together multiple agent models," without addressing "how to define the step structure, how to switch step states, how to save intermediate results, how to modify after the user pauses, which steps need to be rerun, and how to replay the historical process" as a complete technical problem. As a result, they cannot truly achieve collaborative analysis that is interruptible, intervention-friendly, and partially re-executed. Summary of the Invention

[0011] To address the problems existing in the prior art, this invention proposes a multi-agent collaborative intelligence analysis method with interruptible step orchestration. Without redundantly building existing business capabilities such as collection, extraction, retrieval, synthesis, and reporting, this method organizes these capabilities into collaborative tasks through step diagrams and intermediate product packages. During task execution, it supports step-level pauses, manual modifications, partial re-execution, and audit replay, thereby improving the flexibility, resource utilization, result interpretability, and system maintainability of collaborative tasks.

[0012] A first aspect of the present invention provides a multi-agent collaborative method for intelligence analysis with interruptible step orchestration, comprising:

[0013] Step S1: Receive and parse the collaborative task request to obtain the initial task object;

[0014] Step S2: Generate a collaborative task step diagram based on the initial task object and preset capability mapping rules;

[0015] Step S3: Transform each execution step in the collaborative task step diagram into a standardized step description;

[0016] Step S4: Based on the collaborative task step diagram and the step description, construct the execution context;

[0017] Step S5: Set up a separate step status visualization control unit for each execution step to provide a step execution status display and manual intervention interface;

[0018] Step S6: Based on the execution context and the collaborative task step diagram, execute each execution step;

[0019] Step S7: After each execution step is completed, an intermediate result is obtained, and an intermediate product package is generated based on the intermediate result and the execution context;

[0020] Step S8: When manual intervention occurs, determine the set of affected steps based on the dependencies in the collaborative task step diagram, the output contract of the currently intervened execution step, and the reference relationship between the downstream execution steps and the upstream output results.

[0021] Step S9: Retain the intermediate results and intermediate product packages corresponding to the execution steps that did not enter the affected step set, and re-execute only the execution steps in the affected step set; generate updated intermediate results for each re-executed execution step, and generate corresponding intermediate product packages based on the updated intermediate results and the execution context;

[0022] Step S10: Output the final task result based on the intermediate product package and generate an audit playback object.

[0023] Furthermore, it also includes: step S11, which encapsulates the collaborative task step diagram, prompts, parameters and templates that have successfully executed the intelligence analysis task into a collaborative template for reuse.

[0024] Furthermore, the initial objects of the task include: task subject, task objective, time range, source range, output format, priority, confidentiality level, constraints and / or user identity identifier.

[0025] Furthermore, the execution steps include: a collection step, an extraction step, a retrieval step, a synthesis step, and a reporting step.

[0026] Furthermore, the execution steps also include: an export step, a translation step, a mapping step, a trend analysis step, and / or a verification step;

[0027] The intermediate product package includes: task identifier, step identifier, input information reference list, output result reference list, parameter snapshot, prompt word version, agent version, running time, status code and timestamp, replay pointer, exception information, log, upstream step version number and current step version number;

[0028] The intermediate product package is used to support the reuse of results between steps, users to view intermediate processes, playback of historical results, and team sharing.

[0029] Furthermore, the step status visualization control unit is used to display in real time the step name, step status, step task parameters, running time, input summary, output summary, current progress and execution status, intermediate results and status control buttons for a single execution step;

[0030] The step status includes: pending execution, in execution, paused, resumed, completed, failed, invalidated, and pending rerun.

[0031] The status control button is used to trigger pause, resume execution, stop, or restart the execution of steps.

[0032] Furthermore, the affected step set is a set of steps determined after human intervention, based on the step dependencies in the collaborative task step diagram, the changes in the output contract of the intervened step, and the reference relationships between downstream steps and the output results of the intervened step; the affected step set includes at least the currently intervened step and downstream steps that directly or indirectly depend on its output results.

[0033] A second aspect of the present invention provides an intelligence analysis multi-agent collaborative system with interruptible step orchestration, comprising:

[0034] The task receiving and parsing module is used to receive collaborative task requests and parse them to obtain the initial task object.

[0035] The step diagram generation module is used to generate a collaborative task step diagram based on the initial task object and preset capability mapping rules;

[0036] The standardized step description module is used to convert each execution step in the collaborative task step diagram into a standardized step description.

[0037] The execution context management module is used to construct an execution context based on the collaborative task step diagram and the step description;

[0038] The state machine management module is used to set up a separate step status visualization and control unit for each execution step, providing a step execution status display and manual intervention interface, and maintaining step status management;

[0039] The step scheduling module is used to call each agent model to execute each execution step based on the execution context and the collaborative task step diagram; and to re-execute the execution steps in the affected step set.

[0040] The intermediate artifact package management module is used to obtain intermediate results after each execution step is completed, generate intermediate artifact packages based on the intermediate results and the execution context, and manage the intermediate artifact package version.

[0041] The affected step set determination module is used to determine the affected step set when human intervention occurs;

[0042] The final result module is used to output the final task result based on the intermediate product package.

[0043] Furthermore, it also includes:

[0044] The audit replay module is used to generate audit replay objects and reproduce historical execution processes;

[0045] The intelligent agent model is used to execute steps by calling the interface of the intelligent agent model, sending input information to the intelligent agent model, and obtaining result information fed back from the intelligent agent model;

[0046] The intelligent agent model includes: a data acquisition intelligent agent model, an information extraction intelligent agent model, a retrieval or question answering intelligent agent model, a comprehensive generation intelligent agent model, a report template filling intelligent agent model, an export intelligent agent model, a translation intelligent agent model, a graph analysis intelligent agent model, a trend analysis intelligent agent model, and / or a verification intelligent agent model.

[0047] A third aspect of the present invention provides a processor-readable storage medium storing a computer program, wherein when the processor executes the computer program, it implements the interruptible step orchestration method for multi-agent collaborative intelligence analysis described in the first aspect of the present invention.

[0048] The advantages of this invention compared to the prior art are:

[0049] First, it avoids redundant investment of intelligent large-scale model resources in the collaboration layer. By unifying the step diagrams and step descriptions, it achieves standardized orchestration of existing intelligent large-scale model capabilities interfaces such as collection, extraction, retrieval, synthesis, and reporting. This allows the technical focus of multi-agent collaboration to be on step orchestration and state control. When a step fails or the user wants to make adjustments, only local large-scale model resources need to be called, without having to repeatedly call the entire intelligent large-scale model's business capabilities, thereby significantly reducing computational resource consumption and improving interaction flexibility.

[0050] Second, it enables the reuse of step results and manual intervention. By standardizing intermediate product packages and unifying the input and output formats of each step, upstream results can be directly reused downstream. It also supports users viewing and modifying intermediate products after pausing at any step, and supports result reuse between steps, user viewing of intermediate processes, historical result playback, and team sharing, improving the flexibility and controllability of collaborative tasks. By defining standardized step descriptions for each step and establishing shared execution contexts and fine-grained step states, it enables precise management of the input, output, parameters, version, and status of individual steps, providing a technical foundation for pausing, resuming, and re-execution.

[0051] Third, it supports step-level pauses and partial re-execution. Through the step status visualization control unit and the calculation of the affected step set, the system can accurately determine which downstream steps are affected by the current modification, and only rerun the affected steps, rather than re-executing the entire task, thereby significantly reducing resource consumption and improving execution efficiency.

[0052] Fourth, enhance the interpretability of results and the team's collaborative reuse capabilities. By fully retaining the complete process information such as the version of the step diagram, parameter snapshots, prompt word versions, agent model versions, and input / output references in the audit replay object, the execution process is traceable, and result deviations can be located, which facilitates team review, auditing, and knowledge reuse. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the steps of a multi-agent collaborative intelligence analysis method with interruptible step arrangement provided in an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of an intelligence analysis multi-agent collaborative system with interruptible step arrangement provided by an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0056] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0057] Method Example 1

[0058] This invention proposes a multi-agent collaborative intelligence analysis method with interruptible step orchestration. Without duplicating existing business capabilities such as collection, extraction, retrieval, synthesis, and reporting, it organizes these capabilities into collaborative tasks through step diagrams and intermediate product packages. During task execution, it supports step-level pauses, manual modifications, partial re-execution, and audit replay, thereby improving the flexibility, resource utilization, result interpretability, and system maintainability of collaborative tasks.

[0059] In a first aspect, the present invention provides a multi-agent collaborative method for intelligence analysis with interruptible step scheduling, the flowchart of which is attached. Figure 1 As shown, it specifically includes:

[0060] Step S1: Receive and parse the collaborative task request to obtain the initial task object.

[0061] The collaborative task request is an intelligence analysis task described in natural language input by the user and / or an intelligence analysis task selected through task template parameters. It includes task information such as task topic, task objective, time range, source range, output format, priority, confidentiality level, constraints and / or user identification.

[0062] By parsing the text of the collaborative task request, the initial task object is obtained. The initial task object includes: task topic, task objective, time range, source range, output format, priority, confidentiality level, constraints, and / or user identification.

[0063] Step S2: Generate a collaborative task step diagram based on the initial task object and preset capability mapping rules.

[0064] The collaborative task step graph is a directed graph structure composed of several execution steps and directed edges connecting each execution step. The directed edges describe the logical topological relationship between each execution step.

[0065] The logical topology relationships include execution order, dependency relationships, call order, parallel relationships, and / or branching relationships.

[0066] The execution steps include a collection step, an extraction step, a retrieval step, a synthesis step, and a reporting step.

[0067] Optionally, the execution steps may further include: an export step, a translation step, a mapping step, a trend analysis step, and / or a verification step.

[0068] In a workflow system, an execution step represents a specific and well-defined operational stage in a business process. The steps of collection, extraction, retrieval, synthesis, and reporting are precisely these five independent yet interconnected stages, collectively forming a complete intelligence analysis chain. The purpose of dividing the execution steps is to ensure that the technical focus of multi-agent collaboration lies in step orchestration and state control, rather than rerunning the entire task.

[0069] The capability mapping rule is a set of preset conversion logic that automatically transforms an initial task object into a collaborative task step diagram and the specific configuration of each step. Its core function is that after the system receives an initial task object that has undergone text parsing, it does not require manual specification of which agent model to call and what parameters to pass for each step; instead, this rule automatically maps the "task intent" to the "step arrangement." Specifically, the capability mapping rule's working mechanism includes three steps:

[0070] Step T1: Determine the execution step type. The initial task object contains multiple fields such as task topic, task objective, time range, source range, and output format. The mapping rules determine which execution steps the task needs to use based on the combination of these fields. For example, if the source range specifies that data needs to be obtained from publicly available external information, the mapping rules will determine that a "data collection step" must be included; if the output format requires the generation of a structured report, then a "reporting step" must be included in the step diagram. In other words, each key field in the initial task object corresponds to a set of triggering conditions for execution steps in the rules.

[0071] Step T2 arranges the sequence and dependencies of steps. Intelligence analysis inherently involves stage dependencies—data collection must precede information extraction; only then can the extracted structured data be retrieved and synthesized; and only the synthesized evidence set ultimately drives report generation. Capability mapping rules solidify these business common sense principles into directed dependencies between steps, ensuring that the generated step diagram is topologically sound. For example, the data collection step is always upstream of the extraction step, and the synthesis step is always upstream of the reporting step. This constraint is automatically guaranteed by the rules, rather than requiring manual connections by the user.

[0072] Step T3: Automatically populate step parameters. Fields in the initial task object, such as time range, source range, confidentiality level, and priority, not only determine which steps are enabled but also directly translate into the running parameters for each step. For example, "time range" is mapped to the time filtering conditions for the collection step and the time window for the retrieval step; "source range" is mapped to the data source list for the collection step and the index range for the retrieval step; "step type" and "confidentiality level" affect the selection of the agent model and the anonymization strategy in the reporting step. The mapping rules essentially maintain a lookup table from task fields to step parameters, ensuring that each step carries a running configuration matching the task objective upon generation.

[0073] Capability mapping rules can be summarized as a set of pre-defined transformation rules that take the initial task object as input and a collaborative task step diagram as output. It maps fields such as task topic, objective, time range, source range, output format, priority, confidentiality level, constraints, and / or user identification to predefined logical selections, dependencies, and parameter configurations for execution steps such as collection, extraction, retrieval, synthesis, and reporting. This achieves automated orchestration from a natural task description to an executable collaborative task step diagram without redundantly building business capabilities.

[0074] Step S3: Transform each execution step in the collaborative task step diagram into a standardized step description.

[0075] The step description refers to a standardized definition of a single execution step.

[0076] The step description includes: step identifier, step type, input contract, output contract, agent model interface, parameter set, prompt word reference, error handling strategy, failure strategy, maximum number of retries, manual intervention permission, local re-execution strategy, and global re-execution strategy.

[0077] The intelligent agent model interface is a calling interface for collecting intelligent agent models, extracting information intelligent agent models, retrieving or answering intelligent agent models, comprehensively generating intelligent agent models, filling report templates with intelligent agent models, exporting intelligent agent models, translating intelligent agent models, graph analysis intelligent agent models, trend analysis intelligent agent models, or verifying intelligent agent models.

[0078] The corresponding agent model interfaces are also different for different types of execution steps:

[0079] The data acquisition steps correspond to the data acquisition agent model interface;

[0080] The extraction step corresponds to the information extraction intelligent agent model interface;

[0081] The retrieval steps correspond to the retrieval or question-answering intelligent agent model interface;

[0082] The synthesis step corresponds to the interface for synthesizing and generating intelligent agent models;

[0083] The reporting step corresponds to the interface for filling the report template into the agent model;

[0084] The export step corresponds to the interface for exporting intelligent agent models;

[0085] The translation steps correspond to the translation agent model interface;

[0086] The graph step corresponds to the graph analysis intelligent agent model interface;

[0087] The trend analysis steps correspond to the trend analysis intelligent agent model interface;

[0088] The verification step corresponds to the verification agent model interface.

[0089] The input contract is used to specify the acceptable input data structure and validation rules for the current execution step, including: input field name, field type, field meaning, field source, whether it is required, allowed value range, data encoding method, data storage location or location pointer, version requirements, and security level;

[0090] The output contract is used to specify the output data structure and external provision method generated by the current execution step, including: output field name, field type, field meaning, output format, version number generation rules, data storage location or location pointer, summary information, interface fields for downstream steps to read, and exception return format.

[0091] Preferably, neither the input contract nor the output contract is limited to specific business content, but rather serves as a unified format constraint for data transfer across steps, supporting automatic connection between steps, result reuse, and calculation of the affected step set.

[0092] Step S4: Based on the collaborative task step diagram and the step description, construct the execution context.

[0093] The execution context is a set of state information shared by all execution steps during the operation of a collaborative task, and it maintains version continuity when paused, resumed, or re-executed.

[0094] The execution context is a shared state container that accompanies the entire lifecycle of a collaborative task, responsible for recording and transmitting all runtime information of an intelligence analysis task from initiation to completion.

[0095] The execution context includes the following elements:

[0096] Identity and Target Classes: Task Identifier, Task Target Summary, User Identity Identifier;

[0097] Progress control categories: current step cursor, task-level status (e.g., running, paused, completed), start time;

[0098] Step arrangement category: List of completed steps, list of steps to be executed;

[0099] Data sharing classes: global variable dictionary, product index table;

[0100] Audit traceability categories: user operation logs, version numbers for each step, and version information.

[0101] These elements collectively support the core capabilities that this invention aims to achieve: step-level pause, partial re-execution, and audit replay. When a user pauses a task midway and modifies the parameters of a certain step, the "product index table" and "version numbers of each step" stored in the execution context enable the system to accurately calculate the set of affected steps. When it is necessary to replay the historical process, the "user operation log" and "global variable dictionary" provide complete traceability evidence. It can be said that the execution context is the key data structure that realizes the design goals of "interruptibility, intervention capability, and partial re-execution capability."

[0102] The collaborative task step diagram provides the skeleton of the task—what steps are involved, the dependencies between steps, and the topological order. The "list of completed steps" and "list of steps to be executed" in the execution context are initialized based on the topological structure of the step diagram.

[0103] The step description provides metadata for each step—step identifier, step type, input / output contract, agent model interface, parameter set, prompt word reference, etc.

[0104] The step diagram determines the "how steps are arranged" part of the execution context, while the step description determines the "what each step is and what it produces" part. When constructing the execution context, the system initializes the step list and cursor position with the step diagram as the skeleton, and then fills the empty spaces for the initial version number and product index with the step description, ultimately forming a complete state container that can be shared and updated by all steps during runtime.

[0105] By defining standardized step descriptions for each step and establishing a shared execution context, it is possible to precisely manage the inputs, outputs, parameters, versions, and states of individual steps, providing a technical foundation for pausing, resuming, and re-executing.

[0106] Step S5: Set up a separate step status visualization control unit for each execution step to provide a step execution status display and manual intervention interface.

[0107] The step status visualization control unit is used to display in real time the step name, step status, step task parameters, running time, input summary, output summary, current progress and execution status, intermediate results and status control buttons for a single execution step.

[0108] The step status includes: pending execution, executing, paused, resumed, completed, failed, invalidated, and pending rerun. "Invalidated" means that the historical result of this step is no longer a valid result due to changes in upstream input; "pending rerun" means that the system has calculated that this step belongs to the affected step set, but has not yet started re-execution.

[0109] The status control buttons are used to provide an interface for manual intervention in the execution of steps, such as triggering pause, resume execution, stop or restart the execution of steps.

[0110] In this invention, steps S2 to S5 are the foundation for realizing the interruptible step orchestration function. They realize the division of collaborative tasks into multiple steps with clear dependencies and their step status visualization and control units, laying the technical foundation for realizing the orchestration mechanism of pausing, resuming, modifying and re-executing a certain step.

[0111] Step S6: Based on the execution context and the collaborative task step diagram, execute each execution step.

[0112] Specifically, each execution step is executed sequentially or in parallel according to the topological order, scheduling order, or parallel strategy of the collaborative task step diagram; when each execution step is executed, it calls the corresponding intelligent agent model to perform the task through the intelligent agent model interface corresponding to its step description.

[0113] For example, the data collection step calls the data collection agent model, the extraction step calls the information extraction agent model, the retrieval step calls the retrieval or question-answering agent model, the synthesis step calls the synthesis generation agent model, and the reporting step calls the report template to fill the agent model.

[0114] For example, when an execution step begins execution, it reads the output of the upstream execution step from the "global variable dictionary" and "product index table" of the execution context as its own input;

[0115] During execution, it updates the "current step cursor" and writes its own state to the "task-level state";

[0116] After execution, it registers its products in the "Product Index Table" and moves itself from the "List of Steps to be Executed" to the "List of Completed Steps".

[0117] If a pause or anomaly occurs, the "User Operation Log" records all intervention actions, providing a basis for subsequent partial re-execution and audit playback.

[0118] In this invention, step S5 focuses on "who calls whom and when," while the execution context focuses on "the memory shared by all steps during the call process." Each call to the agent model is accompanied by reading and writing to the execution context. In this way, this invention sets "multi-agent collaboration" as a method for step orchestration and global state control.

[0119] Step S7: After each execution step is completed, an intermediate result is obtained, and an intermediate product package is generated based on the intermediate result and the execution context.

[0120] After each execution step is completed, the intelligent agent model invoked in that step feeds back the intermediate results, which are stored in the data storage table. Simultaneously, an intermediate product package is generated based on the intermediate results and the current execution context. For intermediate results with large amounts of text, the main text can be stored in object storage or document storage, while only the reference addresses are recorded in the intermediate product package.

[0121] The intermediate product package is a standardized intermediate result encapsulation data package used to support result reuse between steps, user viewing of intermediate processes, playback of historical results, and team sharing.

[0122] The intermediate product package includes: task identifier, step identifier, input information reference list, output result reference list, parameter snapshot, prompt word version, agent version, running time, status code and timestamp, replay pointer, exception information, log, upstream step version number and current step version number.

[0123] The input information reference list is used to indicate which upstream step products were actually read in the current step;

[0124] The output result reference list is used to indicate where the output result of the current step is stored and which downstream steps or downstream fields can continue to reference it.

[0125] Both the input information reference list and the output result reference list are structured collections of reference records, rather than lists that directly store the full text content. Each reference record includes at least: source step identifier, source intermediate product package identifier, source version number, field name or field path, storage location pointer, data digest value and / or checksum; wherein, the storage location pointer is one or more of the following: object storage address, database primary key, document fragment identifier, paragraph offset, or fragment number.

[0126] Step S8: When manual intervention occurs, determine the set of affected steps based on the dependencies in the collaborative task step diagram, the output contract of the currently intervened execution step, and the reference relationship between the downstream execution steps and the upstream output results.

[0127] During the task execution phase, if a user needs to adjust the step task parameters or other information of a certain execution step, such as adjusting the prompts, filtering conditions, search conditions, agent model, input range, or supplementing or deleting evidence fragments, manual intervention will occur if the user restarts the execution step after updating the step task parameters, or if the user pauses or stops the execution step.

[0128] When manual intervention occurs, the current execution context, step cursor, parameter snapshot, and input / output references are first frozen to form a pause snapshot. Then, each manual intervention is saved as a new step version, and the user, operation time, and differences before and after the modification are recorded. Finally, the execution steps that are affected by manual intervention are calculated, and these affected steps are counted as the affected step set.

[0129] Freezing input / output references refers to freezing the structured reference records established in the current step, including the source step identifier, intermediate product package identifier, version number, field path, storage location pointer, data digest value, and downstream consumption relationship, in order to form a pause snapshot corresponding to this manual intervention; the frozen objects are reference metadata and location information, rather than copying the entire text of the referenced text or the original content of the image or video to the pause snapshot.

[0130] The affected step set is a set of steps determined after human intervention, based on the step dependencies in the collaborative task step diagram, the changes in the output contract of the intervened step, and the reference relationships between downstream steps and the output results of the intervened step; the affected step set includes at least the currently intervened step and downstream steps that directly or indirectly depend on its output results.

[0131] Preferably, the affected step set includes at least the current step itself and all downstream steps that directly or indirectly depend on the output of the current step; steps that do not depend on the output of the current step are not included in the affected step set.

[0132] This invention enables step-level pauses, manual intervention, and partial re-execution by setting a set of affected steps. When a step fails or the user wants to make adjustments, there is no need to rerun the entire process, which significantly reduces the consumption of computing resources and improves the flexibility of interaction.

[0133] Step S9: Retain the intermediate results and intermediate product packages corresponding to the execution steps that have not entered the affected step set, and re-execute only the execution steps in the affected step set; generate updated intermediate results for each re-executed execution step, and generate corresponding intermediate product packages based on the updated intermediate results and the execution context.

[0134] Switch the step status of each execution step in the affected step set to invalid or pending rerun state; retain the intermediate results and intermediate product packages of the unaffected execution steps; only re-execute the steps in the affected step set, thereby avoiding the entire collaborative task from starting from scratch.

[0135] For example, the step status visualization control unit of a certain execution step displays its step status as "to be rerun", indicating that the system has calculated that the step belongs to the affected step set, but has not yet started to re-execute.

[0136] Furthermore, the step states switch according to state transition rules: the pending state can switch to the executing state; the executing state can switch to the completed, failed, or paused state; the paused state can switch to the executing, stopped, resumed, or pending rerun state; the completed state can switch to the invalidated state when the upstream result changes; the invalidated state switches to the pending rerun state when rescheduled; and the pending rerun state switches back to the executing state. This state machine can fully describe the dynamic process of collaborative tasks.

[0137] Preferably, each execution step is assigned an input signature (Hashin) and a parameter signature (Hashparam) attribute. When a user modifies a step, the signature value of that step changes; the signatures before and after the modification are compared, and only the corresponding step and its dependent steps are added to the affected step set if the signature has changed. This approach can further reduce unnecessary reruns.

[0138] In this invention, steps S8 and S9 implement the partial re-execution function of intelligence analysis tasks, that is, while retaining the results of unaffected steps, only the current step and downstream steps that depend on its output are re-executed.

[0139] Step S10: Output the final task result based on the intermediate product package and generate an audit playback object.

[0140] Once all necessary steps have been executed, the final task result is generated based on the final intermediate product package.

[0141] The final task outcome is an intelligence briefing, thematic analysis, Q&A results, map analysis results, or a comprehensive research report.

[0142] The audit replay object is a data object used to restore the historical execution process, supporting result review, sharing, template reuse, and permission auditing, so that collaborative tasks have complete traceability.

[0143] The audit playback objects include: step diagram version, execution context snapshot, intermediate product package index for each step, parameter snapshot, prompt word version, agent model version, and operation log, etc.

[0144] Step S11 involves encapsulating the successfully executed collaborative task step diagram, prompts, parameters, and template into a collaborative template. Subsequent users can reuse similar tasks by adjusting only a few parameters, and the system retains its step-level interruptibility, modifiability, and re-execution capability during reuse.

[0145] Method Example 2

[0146] For example, the intelligence analysis task described by natural language is: "Generate an intelligence briefing on the business activities of key institutions of Project XX over the past week." This collaborative task request is parsed to extract the task objectives as Project XX, key institutions, and business activities, with a time frame of the past week, and the output format as an intelligence briefing.

[0147] The collaborative task step diagram is then generated, which includes the following steps in sequence: collection step A1, extraction step A2, retrieval step A3, synthesis step A4, and reporting step A5.

[0148] The process involves several steps: Step A1 involves calling a data collection agent model to extract new data related to the XX project from relevant websites over the past week, outputting a candidate document set D1; Step A2 involves calling an information extraction agent model to perform structured processing on the text in D1, completing the processing of entities, relationships, and events, and outputting an extraction result set E2; Step A3 involves calling a retrieval or question-answering agent model based on E2 to obtain a set of evidence fragments R3; Step A4 involves calling a synthesis generation agent model to form a draft conclusion C4 based on "XX project, key institutions, and business dynamics"; and Step A5 involves calling a report template filling agent model to fill in a Word template and generate an intelligence briefing file P5.

[0149] Each step generates a corresponding intermediate product package and displays the step time, input / output summary, and version information.

[0150] If an analyst finds that the conclusions in the summary output of integration step A4 are too generalized and fail to highlight the time boundary of "the past week," they can click "Pause Current Step," switching the status of A4 from executing to paused and freezing the execution context snapshot CTX-4. The analyst can then modify the prompt text of A4, replacing the original prompt text with "Only output key events in chronological order based on new developments verified in the past week, without outputting speculative conclusions," thus triggering manual intervention.

[0151] After manual intervention, the differences in prompts before and after the modification are recorded, and the version number of step A4 is upgraded from v1 to v2. The set of affected steps is determined based on the dependencies in the collaborative task step graph.

[0152] Since A5 directly depends on the output of A4, while A1 to A3 do not depend on the result of A4, the affected step set is {A4, A5}. The execution status of A4 and A5 is marked as pending rerun, while the valid intermediate product packages of A1 to A3 are retained. Subsequently, only A4 and A5 are re-executed to generate new synthesis conclusions and new briefing documents, without re-executing the acquisition, extraction, and retrieval processes, thus saving execution time and computational resources.

[0153] Method Example 3

[0154] For example, based on method embodiment 2, a trend analysis step A4a and an institutional statistics step A4b are added.

[0155] The trend analysis step calculates a daily popularity curve based on the retrieved evidence fragment set; the institutional statistics step aggregates institutional entities in the documents to generate a ranking of key institutions. The synthesis step then summarizes the trend analysis and institutional statistics results into a structured conclusion for analysts.

[0156] If a user finds that the recall scope of the retrieval step is too broad during operation and wishes to limit the source range to a few specified sites, they can pause retrieval step A3 and modify its filtering parameters. The system recalculates the affected step set as {A3, A4a, A4b, A5}, while the results of the collection and extraction steps are retained. Therefore, this invention is not simply about "stopping the entire task and starting over," but rather about selectively rerunning steps.

[0157] System Implementation Examples

[0158] A second aspect of the present invention provides an intelligence analysis multi-agent collaborative system with interruptible step orchestration, as shown in the appendix. Figure 2 As shown, it specifically includes:

[0159] The task receiving and parsing module is used to receive collaborative task requests and parse them to obtain the initial task object.

[0160] The step diagram generation module is used to generate a collaborative task step diagram based on the initial task object and preset capability mapping rules.

[0161] The step standardization description module is used to convert each execution step in the collaborative task step diagram into a standardized step description.

[0162] The execution context management module is used to construct an execution context based on the collaborative task step diagram and the step description.

[0163] The state machine management module is used to set up a separate step status visualization and control unit for each execution step, providing a step execution status display and manual intervention interface, and maintaining step status management.

[0164] The step scheduling module is used to call each agent model to execute each execution step based on the execution context and the collaborative task step diagram; and to re-execute the execution steps in the affected step set.

[0165] The intermediate artifact package management module is used to obtain intermediate results after each execution step is completed, generate intermediate artifact packages based on the intermediate results and the execution context, and manage the intermediate artifact package version.

[0166] The affected step set determination module is used to determine the affected step set when human intervention occurs.

[0167] The final result module is used to output the final task result based on the intermediate product package.

[0168] The audit replay module is used to generate audit replay objects and reproduce historical execution processes.

[0169] Furthermore, in terms of data storage, this system also includes data storage tables, including a collaborative task table, an execution step table, a step version table, an intermediate product package table, an execution context table, an operation log table, and a replay object table. The collaborative task table stores task-level information; the execution step table stores the collaborative task step graph structure; the step version table stores multiple modified versions of the same step; the intermediate product package table stores output results and references; the execution context table stores the task running status; the operation log table stores action logs such as pause, resume, modification, and rerun; and the replay object table stores aggregated indexes that can be used to reproduce historical processes. These data storage tables can be stored in text files or a database.

[0170] Furthermore, this system also includes intelligent agent models, categorized by function as follows: data collection intelligent agent models, information extraction intelligent agent models, retrieval or question answering intelligent agent models, comprehensive generation intelligent agent models, report template filling intelligent agent models, export intelligent agent models, translation intelligent agent models, graph analysis intelligent agent models, trend analysis intelligent agent models, or verification intelligent agent models. These intelligent agent models belong to the larger AI model set. The execution steps involve calling the interface of the intelligent agent model, sending input information to the intelligent agent model, and obtaining the result information fed back from the intelligent agent model.

[0171] The second aspect of this invention has the same technical concept as the first aspect, therefore the remaining technical features will not be repeated here. Those skilled in the art can adapt the specific steps disclosed in the first aspect of this invention to one or more modules disclosed in the second aspect, and can also adaptably increase or decrease the number of system modules, all of which fall within the protection scope of this invention.

[0172] A third aspect of the present invention provides a processor-readable storage medium storing a computer program, wherein the processor, when executing the computer program, implements the interruptible step orchestration method for multi-agent collaborative intelligence analysis described in the first aspect of the present invention. The third aspect of the present invention shares the same technical concept as the first aspect, therefore, the remaining technical features will not be repeated here.

[0173] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A multi-agent collaborative method for intelligence analysis with interruptible step scheduling, characterized in that, include: Step S1: Receive and parse the collaborative task request to obtain the initial task object; Step S2: Generate a collaborative task step diagram based on the initial task object and preset capability mapping rules; Step S3: Each execution step in the collaborative task step diagram is transformed into a standardized step description. The step description provides metadata for each step, including: step identifier, step type, input contract, output contract, agent model interface, parameter set, prompt word reference, error handling strategy, failure strategy, maximum number of retries, manual intervention permission, local re-execution strategy, and global re-execution strategy. The output contract is used to specify the output data structure and external provision method generated by the current execution step, including: output field name, field type, field meaning, output format, version number generation rule, data storage location or location pointer, summary information, interface fields for downstream steps to read, and exception return format. Step S4: Based on the collaborative task step diagram and the step description, construct an execution context. The execution context is a set of state information shared by all execution steps during the operation of the collaborative task, and maintains version continuity when paused, resumed, or re-executed. Step S5: Set up a separate step status visualization control unit for each execution step to provide a step execution status display and manual intervention interface; Step S6: Based on the execution context and the collaborative task step diagram, execute each execution step; Step S7: After each execution step is completed, an intermediate result is obtained. Based on the intermediate result and the execution context, an intermediate artifact package is generated. The intermediate artifact package includes: task identifier, step identifier, input information reference list, output result reference list, parameter snapshot, prompt word version, agent version, execution time, status code and timestamp, replay pointer, exception information, log, upstream step version number, and current step version number. The intermediate artifact package is used to support result reuse between steps, user viewing of intermediate processes, historical result replay, and team sharing. Step S8: When manual intervention occurs, determine the set of affected steps based on the dependencies in the collaborative task step diagram, the changes in the output contract of the currently intervened execution step, and the reference relationship between the downstream execution steps and the upstream output results. Step S9: Retain the intermediate results and intermediate product packages corresponding to the execution steps that did not enter the affected step set, and re-execute only the execution steps in the affected step set; generate updated intermediate results for each re-executed execution step, and generate corresponding intermediate product packages based on the updated intermediate results and the execution context; Step S10: Output the final task result based on the intermediate product package and generate an audit playback object.

2. The multi-agent collaborative intelligence analysis method with interruptible step orchestration according to claim 1, characterized in that, Also includes: Step S11: The collaborative task step diagram, prompts, parameters, and templates that have been successfully executed for intelligence analysis tasks are packaged into a collaborative template for reuse.

3. The multi-agent collaborative method for intelligence analysis with interruptible step orchestration according to claim 1, characterized in that: The initial objects of the task include: task topic, task objective, time range, source range, output format, priority, confidentiality level, constraints and / or user identity identifier.

4. The multi-agent collaborative method for intelligence analysis with interruptible step orchestration as described in claim 1, characterized in that: The execution steps include: collection step, extraction step, retrieval step, synthesis step, and reporting step.

5. The multi-agent collaborative method for intelligence analysis with interruptible step orchestration according to claim 4, characterized in that: The execution steps also include: an export step, a translation step, a mapping step, a trend analysis step, and / or a verification step.

6. The multi-agent collaborative method for intelligence analysis with interruptible step orchestration according to claim 1, characterized in that: The step status visualization control unit is used to display in real time the step name, step status, step task parameters, running time, input summary, output summary, current progress and execution status, intermediate results and status control buttons for a single execution step; The step status includes: pending execution, in execution, paused, completed, failed, invalidated, and pending rerun. The status control button is used to trigger pause, resume execution, stop, or restart the execution of steps.

7. The multi-agent collaborative method for intelligence analysis with interruptible step orchestration according to claim 1, characterized in that: The affected step set is a set of steps determined after human intervention, based on the step dependencies in the collaborative task step diagram, the changes in the output contract of the intervened step, and the reference relationships between downstream steps and the output results of the intervened step; the affected step set includes at least the currently intervened step and downstream steps that directly or indirectly depend on its output results.

8. A multi-agent collaborative intelligence analysis system using the interruptible step arrangement of the method according to any one of claims 1 to 7, characterized in that, include: The task receiving and parsing module is used to receive collaborative task requests and parse them to obtain the initial task object. The step diagram generation module is used to generate a collaborative task step diagram based on the initial task object and preset capability mapping rules; The standardized step description module is used to convert each execution step in the collaborative task step diagram into a standardized step description. The execution context management module is used to construct an execution context based on the collaborative task step diagram and the step description; The state machine management module is used to set up a separate step status visualization and control unit for each execution step, providing a step execution status display and manual intervention interface, and maintaining step status management; The step scheduling module is used to call each agent model to execute each execution step based on the execution context and the collaborative task step diagram; and to re-execute the execution steps in the affected step set. The intermediate artifact package management module is used to obtain intermediate results after each execution step is completed, generate intermediate artifact packages based on the intermediate results and the execution context, and manage the intermediate artifact package version. The affected step set determination module is used to determine the affected step set when human intervention occurs; The final result module is used to output the final task result based on the intermediate product package.

9. A multi-agent collaborative intelligence analysis system with interruptible step orchestration according to claim 8, characterized in that, Also includes: The audit replay module is used to generate audit replay objects and reproduce historical execution processes; The intelligent agent model is used to execute steps by calling the interface of the intelligent agent model, sending input information to the intelligent agent model, and obtaining result information fed back from the intelligent agent model; The intelligent agent model includes: a data acquisition intelligent agent model, an information extraction intelligent agent model, a retrieval or question answering intelligent agent model, a comprehensive generation intelligent agent model, a report template filling intelligent agent model, an export intelligent agent model, a translation intelligent agent model, a graph analysis intelligent agent model, a trend analysis intelligent agent model, and / or a verification intelligent agent model.

10. A processor-readable storage medium, characterized in that, The processor-readable storage medium stores a computer program, and when the processor executes the computer program, it implements an interruptible step-arranged multi-agent collaborative intelligence analysis method as described in any one of claims 1 to 7.

Images