A task flow control method and system of a multi-agent system and a storage medium

Abstract

The application relates to the technical field of multi-agent systems, and provides a task flow control method and system of a multi-agent system and a storage medium. The method comprises the following steps: obtaining a target task execution plan; loading a multi-dimensional interaction contract for a current task node, and controlling a matched execution agent to execute the current task node to generate a target product containing a support data set and an output conclusion, wherein the multi-dimensional interaction contract comprises constraint tools, product output specifications and product verification rules; performing multi-dimensional verification on the target product based on the product verification rules in the multi-dimensional interaction contract, wherein the multi-dimensional verification at least comprises data support degree verification; extracting an uncertainty index of the target product based on a multi-dimensional verification result, and performing dynamic gating flow control on the target product in combination with a risk evaluation level of a downstream associated node; and the application blocks the flow of low-quality products through structured contracts and uncertainty gating, thereby improving the reliability and engineering controllability of a multi-agent task chain.

Description

Technical Field

[0001] This application relates to the field of multi-agent system technology, and more specifically to a task flow control method, system, and storage medium for a multi-agent system. Background Technology

[0002] With the development of artificial intelligence technology, multi-agent systems are widely used in scenarios such as complex data processing, scientific research, and industrial decision-making. In these systems, multiple agents with different functions work collaboratively through task chains, with the output of upstream agents serving as input data for downstream related nodes, forming a cascaded data flow relationship.

[0003] Existing multi-agent collaboration schemes typically rely on natural language interaction or simple format validation to constrain data transfer between agents. However, this approach suffers from the following technical problems:

[0004] First, there is a lack of machine-verifiable input-output boundary constraints. Existing solutions lack hard limits on the scope of tools that agents can invoke, the structural specifications of output data, and the traceability of data sources. This can lead to agents invoking unauthorized tools or generating structurally incompatible products.

[0005] This can cause downstream related nodes to fail to resolve or the task chain to break, resulting in the ineffective consumption of system computing resources.

[0006] Second, there is a lack of a mechanism to verify the substantive credibility of the output content. Existing solutions typically only check the format integrity of the output data, while ignoring the logical support relationship between the output conclusion and its supporting data. When an upstream agent generates a conclusion lacking sufficient evidence, this conclusion may be treated as fact by downstream related nodes, causing errors to cascade and amplify in the task chain, ultimately triggering costly and ineffective operations and resulting in serious waste of system resources.

[0007] Third, there is a lack of dynamic flow control mechanisms based on uncertainty indicators and downstream task risks. Existing solutions mostly use fixed thresholds or manually preset flow rules, which cannot adaptively adjust according to the linkage between product quality and the error sensitivity of downstream related nodes. This results in either task blockage caused by overly strict fixed rules or low-quality data flowing into high-value consumption links due to overly lenient release policies.

[0008] Therefore, how to construct machine-verifiable structured constraints and dynamic gating flow control mechanisms in multi-agent systems to block the propagation of low-quality and highly uncertain data in the task chain and reduce invalid computation and system resource consumption is a technical problem that urgently needs to be solved. Summary of the Invention

[0009] To address the aforementioned problems in existing technologies, this application provides a task flow control method, system, and storage medium for a multi-agent system. It constrains the execution boundaries and product specifications of agents through structured multi-dimensional interactive contracts, and performs dynamic gating flow control based on the linkage between the uncertainty index of product verification results and the risk assessment level of downstream related nodes. This effectively blocks the propagation of low-quality and highly uncertain data in the task chain, improves the data parsing success rate and overall reliability of long-link task execution, and reduces invalid computation and system resource consumption caused by error propagation.

[0010] To achieve the above objectives, this application adopts the following technical solution: a task flow control method for a multi-agent system, comprising: obtaining an execution plan for a target task, the execution plan including multiple task nodes; loading an associated multidimensional interaction contract for the current task node, and controlling a matched execution agent to execute the current task node under the constraints of the multidimensional interaction contract to generate a target product, the target product including a supporting data set and output conclusions, the multidimensional interaction contract being used to limit tool call constraints, product output specifications, and product verification rules; performing multidimensional verification on the target product based on the product verification rules in the multidimensional interaction contract, the multidimensional verification including at least data support degree verification; extracting uncertainty indicators of the target product based on the results of the multidimensional verification, and combining them with the risk assessment levels of downstream associated nodes among the multiple task nodes, performing dynamic gating flow control on the target product.

[0011] The above solution, by configuring structured multi-dimensional interaction contracts for task nodes, solidifies the scope of tool calls, output formats, and verification rules of intelligent agents into machine-verifiable hard constraints, thereby standardizing the data transmission boundaries in multi-agent collaboration from the source. Simultaneously, by performing multi-dimensional verification, including data support, on the target product, it verifies the substantive credibility of the product content rather than merely checking format compliance. Furthermore, by linking the uncertainty indicators in the verification results with the risk assessment levels of downstream related nodes, it achieves dynamic matching between product quality and downstream risk, avoiding the rigidity caused by fixed thresholds, and ensuring the security of high-risk processes while retaining the flexibility of low-risk processes.

[0012] Preferably, the target product is subjected to multi-dimensional verification based on the product verification rules in the multi-dimensional interaction contract. This multi-dimensional verification includes at least data support verification, comprising: obtaining the supporting data set contained in the target product and the output conclusion of the target product; evaluating the quantity, data type, and reliability level of the supporting data in the supporting data set; and determining the data support verification result of the target product based on the logical support relationship between the supporting data set and the output conclusion. This preferred solution transforms data support verification from a subjective judgment to a deterministic calculation process, effectively preventing the downstream propagation of products lacking sufficient evidence.

[0013] Preferably, based on the results of the multi-dimensional verification, uncertainty indicators of the target product are extracted, and combined with the risk assessment levels of downstream related nodes, dynamic gating flow control is performed on the target product. This includes: extracting confidence indicators and data conflict indicators of the target product from the results of the multi-dimensional verification as the uncertainty indicators; determining gating trigger conditions based on the confidence indicators, the data conflict indicators, and the risk assessment levels of the downstream related nodes; and performing corresponding gating actions on the target product in response to the satisfaction of the gating trigger conditions. The gating actions include at least one of data interception, rework and recertification, upgrade review, or route release. This preferred solution achieves differentiated gating action mapping by coupling uncertainty indicators with downstream risk levels to form a multi-dimensional decision matrix.

[0014] Preferably, the multidimensional interaction contract includes tool invocation constraints and product output specifications. The tool invocation constraints are used to limit the whitelist boundaries of tools that the executing agent can invoke and the types of actions that are prohibited from being executed. The product output specifications are used to limit the input and output data structure, required fields, and data source anchors of the target product. This preferred solution ensures the security sandbox constraints of the executing agent at the system architecture level, while ensuring the structural standardization and source traceability of the product data.

[0015] Preferably, in response to the data support verification result indicating that the data support of the target product does not meet a preset support threshold, the flow of the target product to the downstream associated nodes is blocked. The downstream associated nodes include high-cost execution nodes whose resource consumption exceeds a preset consumption threshold. Simultaneously with blocking the flow, a supplementary data acquisition task is automatically generated and triggered. This task acquires supplementary support data that meets the preset support threshold and integrates it into the target product to re-execute the multi-dimensional verification. This preferred solution accurately blocks the flow of low-evidence products to high-cost nodes and automatically triggers a supplementary evidence loop, improving the robustness and continuity of the system.

[0016] Preferably, determining the gating trigger condition based on the confidence index, the data conflict index, and the risk assessment level of the downstream associated nodes includes: when the risk assessment level of the downstream associated node exceeds a preset risk threshold, a high-risk gating condition is triggered, and the gating action corresponding to the high-risk gating condition is mandatory escalation review, regardless of whether the confidence index meets the standard; when the data conflict index indicates that the supporting data set simultaneously contains a first type of data supporting the output conclusion and a second type of data refuting the output conclusion, a conflict gating condition is triggered, and the gating action corresponding to the conflict gating condition includes generating a data conflict report and intercepting the flow of the target product. This preferred scheme establishes two types of security fallback mechanisms: high-risk gating forces manual review of key business processes, and conflict gating makes implicit evidence contradictions explicit.

[0017] Preferably, the multi-dimensional verification of the target product based on the product verification rules in the multi-dimensional interaction contract further includes structural verification and permission verification of the product output specifications. The method also includes: in response to the structural verification indicating that the target product's input / output data structure is missing required fields or that the data units do not meet the parsing requirements of the downstream associated nodes, returning the target product to the executing agent with a structural verification report; and in response to the permission verification indicating that the executing agent has invoked an unauthorized tool or data source outside the whitelist boundary, marking the target product as untrusted and triggering an unauthorized audit process for the executing agent. This preferred solution constructs a dual defense of structural verification and permission verification, respectively achieving a self-healing closed loop for format errors and a complete security response to unauthorized behavior.

[0018] Preferably, in response to the target product completing the dynamic gating flow control, a signature product package is generated based on the data hash value of the target product, the verification report of the multi-dimensional verification, and the data source reference; a globally unique identifier is assigned to the signature product package, and the globally unique identifier is written into downstream task requests and system logs to support product status tracking across task nodes; in response to detecting a change in the input data or associated model version of the target product, the affected downstream associated nodes are located based on the globally unique identifier, and local state updates are performed on the affected downstream associated nodes. This preferred solution encapsulates the product as an indivisible trust unit and implements node-level incremental updates based on the globally unique identifier, improving the system's adaptability to dynamic changes.

[0019] This application also provides a task flow control system for a multi-agent system, comprising: a plan acquisition unit for acquiring an execution plan for a target task, the execution plan including multiple task nodes; a contract loading and execution unit for loading an associated multi-dimensional interaction contract for the current task node, and controlling a matched execution agent to execute the current task node under the constraints of the multi-dimensional interaction contract to generate a target product, the target product including a supporting data set and an output conclusion, the multi-dimensional interaction contract being used to limit tool call constraints, product output specifications, and product verification rules; a product verification unit for performing multi-dimensional verification on the target product based on the product verification rules in the multi-dimensional interaction contract, the multi-dimensional verification including at least data support verification; and a gated flow unit for extracting uncertainty indicators of the target product based on the results of the multi-dimensional verification, and combining the risk assessment levels of downstream associated nodes among the multiple task nodes to perform dynamic gated flow control on the target product. The above-described task flow control system for a multi-agent system is used to execute a task flow control method for a multi-agent system.

[0020] The aforementioned system, by solidifying the task flow control method into the collaborative operation of functional units, achieves precise control over the quality of data flow during multi-agent collaboration, ensuring the compliance of execution boundaries and the real-time nature of flow decisions at the system architecture level.

[0021] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the task flow control method for a multi-agent system as described in any of the foregoing method embodiments.

[0022] The beneficial effects of this application include:

[0023] (1) By configuring a structured multidimensional interaction contract for task nodes, the scope of tool calls, output format and verification rules of the agent are solidified into machine-verifiable hard constraints, which standardizes the data transmission boundary in multi-agent collaboration from the source, eliminates the risk of unauthorized operation caused by model illusion or prompt word injection, and improves the predictability and traceability of system task execution.

[0024] (2) By performing multi-dimensional verification, including data support, on the target product, the logical support relationship between the supporting data set and the output conclusion is verified, and the quality control is elevated from formal compliance to substantive credibility. This effectively blocks the spread of products lacking sufficient evidence in the task chain, reduces the probability of downstream related nodes performing invalid operations due to consuming incorrect data, and reduces the waste of system computing resources.

[0025] (3) By linking the uncertainty index in the product verification result with the risk assessment level of the downstream associated node, a dynamic gating flow control mechanism is constructed, which realizes the adaptive matching of product quality and downstream risk, avoids the rigidity problem caused by fixed threshold, and maintains the flexibility of low-risk exploration while ensuring the safety of high-risk links, thus achieving a dynamic balance between system resource efficiency and safety.

[0026] (4) By constructing an automated self-healing closed loop of verification, blocking, supplementary verification, and re-verification, as well as a security fallback mechanism for high-risk mandatory review and explicit conflict, the system can automatically handle various abnormal scenarios without human intervention, which significantly improves the robustness, continuity and engineering controllability of the multi-agent system when dealing with complex long-link tasks. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the multi-agent system task flow control system architecture according to an embodiment of this application;

[0028] Figure 2 This is a schematic diagram illustrating the relationship between the multidimensional interactive contract and the product data structure in an embodiment of this application;

[0029] Figure 3 This is a schematic flowchart of the task flow control method according to an embodiment of this application;

[0030] Figure 4 This is a schematic diagram of the exception handling and gating branch logic in an embodiment of this application;

[0031] Figure 5 This is a schematic diagram of a multi-agent collaborative application scenario according to an embodiment of this application.

[0032] Among them, 10-Plan Acquisition Unit, 20-Contract Loading and Execution Unit, 30-Product Verification Unit, 40-Gated Flow Unit, 60-Execution Agent, and 70-Downstream Related Node. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0035] Example 1: As Figure 3As shown, this embodiment provides a task flow control method for a multi-agent system. This method achieves precise control over the data flow quality during multi-agent collaboration through structured contract constraints and a dynamic risk linkage mechanism. The method mainly includes the following steps:

[0036] Step S100: Obtain the execution plan of the target task, which contains multiple task nodes.

[0037] Specifically, the target task refers to a data processing or scientific research request that requires collaboration among multiple roles, such as new material screening or fault attribution analysis. The execution plan is a machine-readable description of the target task after its structured decomposition; its data structure is typically represented as a directed acyclic graph (DAG) or a state machine model. Each task node represents an atomic processing unit, and data dependencies, control flow order, and preconditions are defined between nodes through edges. By parsing the execution plan, the system can clearly identify the position of the current node to be executed in the entire task chain, its upstream input sources, and downstream consumer objects, thus providing a global context for subsequent contract loading and flow control.

[0038] Step S200: Load the associated multidimensional interaction contract for the current task node, and control the matched execution agent to execute the current task node under the constraints of the multidimensional interaction contract to generate the target product. The target product includes the supporting data set and the output conclusion. The multidimensional interaction contract is used to limit the tool call constraints, product output specifications and product verification rules.

[0039] Specifically, the multidimensional interaction contract is the core technical feature that distinguishes this application from traditional natural language prompts. It is a machine-verifiable and enforceable structured protocol. This contract is typically stored in the form of JSON Schema, OpenAPI specifications, or custom DSLs, and includes three dimensions of hard constraints: tool call constraints, limiting the APIs that the agent can call and prohibiting dangerous operations; output specifications, defining the field names, data types, required fields, and data source anchor formats of the output data; and output verification rules, pre-setting evaluation indicators or logical assertions to verify the quality of the output. The system injects the multidimensional interaction contract into the runtime environment of the executing agent, requiring it to adhere to the preset structure and permission boundaries when generating content. The target product consists of two parts: a "supporting data set" and an "output conclusion." The supporting data set serves as the evidence carrier for the output conclusion, including original observations, simulation results, literature excerpts, or computational intermediates. This two-element binding structure of "conclusion + evidence" fundamentally prevents the agent from generating unfounded, illusory conclusions.

[0040] Step S300: Perform multi-dimensional verification on the target product based on the product verification rules in the multi-dimensional interactive contract. The multi-dimensional verification includes at least the data support verification.

[0041] Specifically, the execution logic of multi-dimensional verification is entirely driven by the rules defined in the multi-dimensional interaction contract. In addition to routine format integrity verification, this step particularly emphasizes the necessity of data support verification, aiming to verify whether the supporting data set is sufficient to derive the output conclusion. Verification dimensions include the quantity of evidence, type matching degree, source reliability level, and semantic implication relationship between evidence and conclusion. Only when the data support verification passes can it be proven that the target product possesses the minimum credibility threshold to enter the next processing stage, elevating quality control from "formal compliance" to "substantial credibility."

[0042] Step S400: Based on the results of multi-dimensional verification, extract the uncertainty index of the target product, and combine it with the risk assessment level of downstream related nodes in multiple task nodes to perform dynamic gating flow control on the target product.

[0043] Specifically, dynamic gating workflow control solves the problem that fixed thresholds in traditional workflows cannot adapt to complex scenarios. The system extracts uncertainty indicators (such as confidence scores, data conflict ratios, or evidence coverage) from multi-dimensional verification results and links them with the risk assessment levels of downstream related nodes. The risk assessment level of a downstream related node reflects the node's sensitivity to input data errors and the cost of error correction. For example, high-cost wet testing nodes or decision-making nodes involving safety and compliance usually have high risk levels. When the uncertainty indicator is high and the downstream risk level is also high, the system triggers strict gating actions; while when the uncertainty indicator is slightly lower but the downstream is only a low-risk exploratory node, the system may allow downgrading. This two-factor linkage mechanism of "product quality - downstream risk" achieves a dynamic balance between resource efficiency and security. Through the coordinated operation of steps S100 to S400, this embodiment constructs a technical framework of "contractual constraints + evidence verification + risk linkage".

[0044] Example 2: In this example, the multi-dimensional verification and dynamic gating flow control in Example 1 are further refined, focusing on the quantitative evaluation mechanism of data support and the triggering logic of uncertainty gating. Preferably, multi-dimensional verification is performed on the target product based on the product verification rules in the multi-dimensional interaction contract. Multi-dimensional verification includes at least data support verification, specifically: obtaining the supporting data set contained in the target product and the output conclusion of the target product; evaluating the quantity, data type, and reliability level of the supporting data in the supporting data set; and determining the data support verification result of the target product based on the logical support relationship between the supporting data set and the output conclusion.

[0045] Specifically, the system performs multi-dimensional assessments of the supporting data's fundamental attributes. The quantity assessment of supporting data uses a weighted statistical method based on a pre-defined evidence quota model; for example, peer-reviewed journal articles are counted as 1.0 standard evidence unit, and preprints as 0.3 standard evidence units. Data type assessment verifies the completeness of the evidence modality, while reliability assessment dynamically scores the data source's authority, timeliness, and citation frequency.

[0046] This embodiment introduces a quantitative determination of logical support relationships, verifying whether the supporting data can lead to the output conclusion through semantic implication analysis or a causal strength scoring model. Data support score. It can be calculated using the following formula:

[0047]

[0048] in, This represents the total number of supporting evidence items in the dataset; Indicates the first The type weight coefficient of each piece of evidence reflects the probative value benchmark of that type of evidence in the current task scenario; Indicates the first The reliability rating of the evidence typically ranges from 0 to 1. Indicates the first Item of evidence With output conclusion The logical support strength function between them is calculated using a natural language inference model or a knowledge graph path search algorithm. When the data support threshold is lower than the preset threshold in the multidimensional interaction contract, the system determines that the data support verification fails.

[0049] Based on this, uncertainty indicators of the target product are extracted from the results of multi-dimensional verification, and combined with the risk assessment levels of downstream related nodes in multiple task nodes, dynamic gating flow control is performed on the target product. Specifically, this includes: extracting confidence indicators and data conflict indicators of the target product from the results of multi-dimensional verification as uncertainty indicators; determining gating trigger conditions based on confidence indicators, data conflict indicators, and risk assessment levels of downstream related nodes; and performing corresponding gating actions on the target product in response to the satisfaction of the gating trigger conditions. The gating actions include at least one of data interception, rework and recertification, upgrade review, or route release.

[0050] Specifically, the system extracts two types of core uncertainty indicators from the verification results: the confidence index reflects the overall degree of certainty about the quality of the product, which is usually determined by... After normalization, the data conflict index is obtained; it quantifies the consistency within the dataset, and the conflict index increases significantly when positive and negative evidence coexist. The system couples these indicators with the risk assessment levels of downstream related nodes. For example, experimental nodes involving expensive reagent consumption have a high-risk level, while ordinary text summary nodes have a low-risk level.

[0051] like Figure 4 As shown, the determination of gating trigger conditions adopts a multi-dimensional decision matrix logic. When the confidence index is lower than the first threshold and the downstream risk assessment level is higher than the second threshold, the high-risk, low-confidence gating condition is triggered; when the data conflict index is higher than the third threshold, the conflict gating condition is triggered. In response to different gating trigger conditions, the system executes differentiated gating actions: data interception temporarily stores the product in an isolation area; rework and supplementary evidence automatically generates a new task node for supplementary evidence; upgraded review routes the product to a reviewer with higher authority; and route release allows the data to continue flowing under the condition of satisfying security constraints.

[0052] Example 3: In this example, the internal structure of the multidimensional interaction contract and its feedback loop mechanism under abnormal scenarios are further described in detail. Preferably, the multidimensional interaction contract includes tool call constraints and product output specifications.

[0053] Tool invocation constraints are used to define the whitelist boundaries of tools that the executing agent can invoke and the types of actions that are prohibited. The whitelist boundaries are typically represented by an enumerated list of allowed API endpoints, function signatures, or database query commands. The system uses a proxy layer or interceptor to perform real-time matching of external requests from the executing agent; requests not on the whitelist are directly rejected. The prohibited action types define a list of negative behaviors, such as prohibiting the deletion of raw data or access to unauthorized network domains. These are prevented before the command is issued using regular expression matching or a semantic analysis engine.

[0054] The output specification defines the input and output data structures, required fields, and data source anchors for the target product. Input and output data structures are typically described using standard schema definition languages ​​such as JSON Schema and Protocol Buffers, clearly specifying the name, data type, value range, and nesting relationships of each field. Required fields define essential core elements of the target product; their absence results in invalid output. Data source anchors mandate that key data items include traceable source identifiers, such as DOI (Document Title), database record ID, or simulation task UUID.

[0055] like Figure 2As shown, a structured constraint-output mapping relationship is formed between the multidimensional interaction contract and the target product. The tool invocation constraints in the multidimensional interaction contract limit the set of tools accessible to the executing agent and the list of prohibited behaviors from the input side. The product output specification defines the data structure template that the target product must follow from the output side. The product verification rules serve as a verification bridge connecting the contract and the product, performing compliance checks on the product after its generation. The target product consists of two parts: a supporting data set and an output conclusion. Each piece of data in the supporting data set establishes a traceable reference relationship with an external data source through a data source anchor. The output conclusion ensures its structural integrity through mandatory field constraints. This three-layer architecture of "contract constraint input—specification definition output—verification bridge verification" makes the multidimensional interaction contract not only a static configuration agreement but also a dynamic governance framework that runs through the entire product lifecycle.

[0056] Based on the above structured definition, this embodiment also constructs an automated anomaly feedback loop. Preferably, multi-dimensional verification is performed on the target product based on the product verification rules in the multi-dimensional interaction contract, including structural verification and permission verification of the product output specifications. The verification process not only focuses on the substantive credibility of the content but also simultaneously considers the compliance of the form and the legality of the behavior, forming the second line of defense in the system's defense architecture.

[0057] For handling structure validation failures, in response to the structure validation indication that the target product's input / output data structure is missing required fields or that the data units do not meet the parsing requirements of downstream related nodes, the target product is returned to the executing agent along with a structure validation report. When the system detects that the target product fails JSON Schema validation or unit conversion checks, it automatically generates a machine-readable structure validation report, including the error type, the path of the erroneous field, the expected format, and repair suggestions. The system then sends the product along with the validation report back to the executing agent, triggering an automatic retry or correction process, achieving a self-healing closed loop for format errors and reducing task chain breaks caused by data format incompatibility.

[0058] Regarding the handling of permission verification failures, in response to the permission verification indication that the executing agent has invoked an unauthorized tool or data source outside the whitelist boundary, the target artifact is marked as untrusted, and an unauthorized audit process for the executing agent is triggered. When the system detects that the executing agent has invoked an unauthorized tool or accessed a restricted data domain, the artifact is immediately marked as "untrusted" and prevented from entering the flow queue. The system automatically triggers the unauthorized audit process, recording a complete audit log including the agent ID, task node ID, name of the violating tool, request parameters, and timestamp. In cases of serious violations, the system can also suspend the executing agent's subsequent task scheduling permissions and send a security alert. The above-mentioned exception handling for structure verification and permission verification is fully automatic and machine-driven, requiring no manual intervention to complete the judgment and response.

[0059] Example 4: In this example, the differentiated handling strategy of the system when facing abnormal verification results is further elaborated. Unlike the general dynamic gating described in Example 2, this example focuses on specific control logic under two extreme scenarios: one is an automatic verification and self-healing mechanism for repairable defects, and the other is a forced circuit breaker mechanism for unacceptable risks. These two mechanisms together constitute a deep defense system for the system to cope with complex uncertainties.

[0060] Preferably, in response to the data support verification result indicating that the data support of the target product does not meet the preset support threshold, the flow of the target product to downstream related nodes is blocked. When the data support score is lower than the minimum admission standard defined in the multidimensional interaction contract, the system determines that the current product is in a "insufficient evidence" state and accurately identifies the downstream related nodes that the product was originally planned to flow to. In this application, downstream related nodes include high-cost execution nodes whose resource consumption exceeds a preset consumption threshold. High-cost execution nodes refer to processing units that will generate significant irreversible resource consumption once started, such as wet laboratory nodes that consume expensive chemical reagents or long-cycle simulation nodes that occupy large-scale supercomputing clusters. The system dynamically identifies such nodes by reading the metadata tags of each node in the execution plan.

[0061] While blocking the flow, a supplementary data acquisition task is automatically generated and triggered. This task acquires supplementary supporting data that meets a preset support threshold and integrates it into the target product for re-performing multi-dimensional verification. When insufficient support is detected, the system automatically parses the verification report, locates the specific missing evidence dimension, and instantiates a supplementary data acquisition task accordingly. This task automatically selects a suitable external data source or computing tool based on the tool whitelist in the multi-dimensional interaction contract. The acquired supplementary supporting data is incrementally appended to the supporting data set of the target product, preserving the integrity of the original data for auditing purposes. After integration, the system automatically resubmits the product to step S300 for multi-dimensional verification, forming an automated closed loop of "verification-blocking-supplementary evidence-re-verification". Only when the re-verified data support meets the preset support threshold is the flow blockade on downstream high-cost nodes lifted.

[0062] As another key branch of anomaly handling, this embodiment also defines a mandatory circuit breaker mechanism based on specific risk patterns. Preferably, the gating trigger conditions are determined based on confidence indices, data conflict indices, and the risk assessment levels of downstream related nodes, including the following two special cases:

[0063] The first scenario is a high-risk fallback logic. When the risk assessment level of a downstream related node exceeds a preset risk threshold, a high-risk gating condition is triggered. The corresponding gating action is a mandatory escalation review, regardless of whether the confidence index meets the standard. This logic establishes the system principle of "safety priority over efficiency." In critical business processes, even if the confidence index of the upstream product shows "highly reliable," if the downstream node is marked as extremely high risk, the system will still forcibly block automatic flow and route the product to a designated human expert or a higher-level review agent for verification. "Ignoring the confidence index" means that this gating condition has the highest priority, preventing catastrophic consequences due to model overconfidence or blind spots in assessment indicators. The preset risk threshold can be dynamically adjusted according to the project stage.

[0064] The second scenario involves explicit contradiction logic. When data conflict indicators suggest the simultaneous existence of both first-type data supporting the output conclusion and second-type data refuting it within the supporting dataset, a conflict gating condition is triggered. The corresponding gating actions include generating a data conflict report and intercepting the flow of the target product. When the system discovers, through semantic analysis or source comparison, that both positive and negative evidence exist within the same product, simple confidence scores are often averaged out, masking potential cognitive risks. The system automatically generates a structured data conflict report, listing the specific content, source, and reliability level of both conflicting parties, and summarizing the possible causes of the conflict. Subsequently, the system immediately intercepts the subsequent flow of the product, forcing manual intervention for adjudication or triggering a more in-depth comparative verification task.

[0065] like Figure 4 As shown, the two circuit breaker mechanisms mentioned above, together with the basic gating logic in Example 2, constitute a hierarchical decision-making system. The system prioritizes judging high-risk gating conditions or conflicting gating conditions, then proceeds to the conventional confidence-risk linkage matrix judgment, and finally checks the data support conditions to supplement the judgment. The system makes a flow decision in the order of "safety first, contradiction priority, and self-healing fallback".

[0066] Example 5: This example further elaborates on the data credibility solidification and end-to-end state tracking mechanism of the multi-agent system during task flow. Preferably, in response to the target product completing dynamic gating flow control, a signature product package is generated based on the target product's data hash value, multi-dimensional verification reports, and data source citations. Specifically, after the target product passes various verifications and gating decisions, the system encapsulates it into a signature product package with self-description and self-proof capabilities. The data hash value is a digest obtained by performing a one-way hash operation on the core content of the target product, used to verify data integrity; the verification report records the product's scoring and judgment criteria in terms of data support, structural compliance, and authorization legality; the data source citation includes the original URI, database record ID, or document DOI of all supporting data. By binding these three elements and applying a digital signature or Message Authentication Code (MAC), the signature product package becomes an indivisible trust unit.

[0067] Based on this, a globally unique identifier is assigned to the signed product packet, and this identifier is written to downstream task requests and system logs to support product status tracking across task nodes. The globally unique identifier (GUID) is a key index connecting discrete task nodes. The system uses a UUID generation algorithm conforming to RFC 4122 or a namespace-based deterministic hash algorithm to assign an identifier to the signed product packet. This identifier is stored as the primary key of the signed product packet in object storage or a database and is forcibly injected into all relevant downstream task request headers, message queue metadata, and system audit logs. This allows the system to quickly associate the entire lifecycle event chain of the product from generation, verification, and flow to consumption through lightweight identifier matching.

[0068] More importantly, this embodiment introduces an identifier-based incremental update mechanism to address the frequent data iterations and model upgrades that occur during scientific research. Preferably, in response to the detection of changes in the input data or associated model version of the target product, the affected downstream associated nodes are located based on globally unique identifiers, and local state updates are performed on these affected downstream associated nodes. This application solves this problem by maintaining a dynamic dependency graph with globally unique identifiers as nodes. When the system detects a change in a source data or model version, it first locates the signature product package identifier corresponding to the changed object, and then performs a breadth-first search or topological sorting traversal in the dependency graph to accurately identify all downstream associated nodes that depend on this identifier. For affected nodes, the system automatically marks their status as "pending update" or "invalid," and only triggers the re-execution of these nodes and their necessary predecessor nodes; unaffected branches retain their original cached results. This local state update mechanism refines the change response granularity from "task level" to "node level," improving the system's adaptability to dynamic changes while ensuring data consistency.

[0069] Example 6: As Figure 1 As shown, this embodiment provides a task flow control system for a multi-agent system. This system solidifies the processing logic from the aforementioned method embodiments into specific functional units and achieves precise control of the multi-agent collaboration process through the coordinated configuration of hardware and software. The system mainly includes a plan acquisition unit 10, a contract loading and execution unit 20, an output verification unit 30, and a gated flow unit 40.

[0070] The plan acquisition unit 10 is used to acquire the execution plan of the target task, which contains multiple task nodes. The plan acquisition unit 10 can consist of a processor and memory, and communicates with the contract loading and execution unit 20 via a high-speed data bus or distributed message queue, distributing the parsed task node metadata in the form of structured messages. This unit also supports interface with external task orchestration platforms via RESTful API or gRPC interface, converting YAML and JSON format workflow descriptions into directed acyclic graph data structures.

[0071] The contract loading and execution unit 20 loads the associated multidimensional interaction contract for the current task node and controls the matched execution agent to execute the current task node under the constraints of the multidimensional interaction contract to generate the target product. The target product includes a supporting data set and output conclusions. The multidimensional interaction contract is used to limit tool call constraints, product output specifications, and product verification rules. The contract loading and execution unit 20 integrates a contract repository submodule and an agent scheduling submodule. The contract repository submodule maintains a persistent multidimensional interaction contract database, supporting quick retrieval and instantiation of contract configuration objects based on task node type. The agent scheduling submodule starts the matched execution agent 60 through a container orchestration interface and injects the multidimensional interaction contract into its runtime sandbox. This unit also acts as a proxy gateway, intercepting and whitelisting external tool call requests from the execution agent 60 in real time. After the target product is generated, this unit temporarily stores it in a shared storage area and sends a verification trigger event to the product verification unit 30.

[0072] Product verification unit 30 is used to perform multi-dimensional verification on the target product based on the product verification rules in the multi-dimensional interaction contract. The multi-dimensional verification includes at least data support verification. Product verification unit 30 can be deployed as an independent microservice cluster, internally containing a schema verification engine, a semantic implication analysis engine, and an evidence quota statistics engine. These engines read the target product and the multi-dimensional interaction contract in parallel, performing structural compliance checks, data support metric assessments, and permission legality audits, respectively. The verification results are serialized into standardized verification report objects. For data support verification, this unit has a built-in pre-trained inference model or knowledge graph query interface to automatically calculate the logical support strength between the supporting data set and the output conclusion.

[0073] The gating flow unit 40 is used to extract uncertainty indicators of the target product based on the results of multi-dimensional verification, and combine them with the risk assessment levels of downstream related nodes in multiple task nodes to perform dynamic gating flow control on the target product. The gating flow unit 40 is the decision-making center of the system, including a high-performance processor and non-volatile storage media for storing the gating decision matrix and historical flow logs. This unit asynchronously obtains verification reports and uncertainty indicators by subscribing to the verification completion event of the product verification unit 30, and obtains the risk assessment levels of downstream related nodes 70 in real time. The internal state machine engine calculates gating action instructions according to a preset linkage strategy and sends corresponding control signals to downstream related nodes 70 or external review terminals via the message bus.

[0074] It should be understood that the aforementioned units can be integrated within the same physical server or distributed across cloud clusters, edge computing nodes, or local workstations. The executing agent 60 and downstream associated nodes 70, as external interaction objects of the system, can be integrated with this system via standard network protocols, and their specific implementation is not limited by this application.

[0075] Example 7: This example provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the task flow control method of the multi-agent system described in any of the foregoing examples.

[0076] Computer-readable storage media can be non-volatile memory (such as ROM, EEPROM, flash memory, SSD, optical disc, magnetic tape, or disk), volatile memory (such as RAM, DRAM, or SRAM), or logical storage units in network storage devices, cloud object storage buckets, or distributed file systems.

[0077] The computer program stored on the medium contains a sequence of machine-executable instructions that logically correspond to the method steps described in Embodiments 1 to 6. When a processor (including a CPU, GPU, NPU, or ASIC) reads and executes these instructions from the storage medium, it drives the computer system to perform precise control over the flow of multi-agent tasks.

[0078] It should be understood that the computer-readable storage medium provided in this embodiment is the physical carrier form of the aforementioned method embodiments, ensuring that the technical effects of this application can be reproduced in different hardware platforms or operating system environments. Furthermore, since the program stored on this medium fully implements all the aforementioned method steps, it also possesses beneficial effects such as blocking the propagation of low-quality data, reducing the consumption of ineffective computing resources, and improving the reliability of multi-agent cooperation.

[0079] Example 8: As Figure 5As shown, this embodiment provides a specific application example of a task flow control method for a multi-agent system in an AI4S research expert team collaboration scenario. The target task is decomposed into a serial task chain of literature review, theoretical calculation, and wet experimental verification, involving three types of execution roles: literature expert agent, computational expert agent, and experimental expert agent.

[0080] The system acquires an execution plan containing multiple task nodes and loads associated multidimensional interaction contracts for the literature survey nodes. The contracts configure strict output specifications and tool call constraints for the literature expert agent. For example, when the literature expert agent outputs a conclusion about the stability of a novel ligand, the contract requires that the target product must contain at least three independent literature source anchors and corresponding experimental condition parameter tables, and that the output conclusion fields must conform to a predefined JSON Schema format.

[0081] When the target product generated by the literature expert agent is transferred to the theoretical calculation node, the system performs multi-dimensional verification, including structural verification of the product output specifications. If the ligand molecule structure file output by the literature expert agent is missing a key charge state field or uses non-standard abbreviations for energy units, the product verification unit will detect the structural verification failure and automatically return the target product to the literature expert agent along with a structural verification report.

[0082] After the computational expert agent generates a target product containing stability predictions, the system performs a data support verification. Suppose the output conclusion claims "this ligand has excellent stability at high temperatures," but the supporting dataset only contains room-temperature literature citations and lacks high-temperature molecular dynamics simulation data. The system determines the data support score is below a preset support threshold. The system identifies the downstream wet experimental verification node as a high-cost execution node with resource consumption exceeding a preset consumption threshold. It immediately blocks the flow of the target product to the wet experimental node and automatically triggers a supplementary data acquisition task, instructing the computational expert agent to call high-temperature simulation tools to obtain the missing thermal stability data. After fusion, the multi-dimensional verification is re-executed.

[0083] Before the target product enters the wet testing verification stage, the system executes dynamic gating flow control based on risk assessment. Even if the stability conclusion has passed data support verification and has a high confidence level, if the material involves high-risk categories such as highly toxic chemicals or high-pressure reactions, the risk assessment level of its downstream related nodes will exceed the preset risk threshold. The system determines that a high-risk gating condition has been triggered, disregards whether the confidence index meets the standard, and forcibly executes an escalation review action, routing the target product and verification report to a senior safety expert for manual review.

[0084] As can be seen from the AI4S research expert team application scenario in this embodiment, this application solves practical technical problems such as uncontrollable data quality, cascading error propagation, and ineffective consumption of high-value resources in multi-agent collaboration through specific technical means such as multi-dimensional interactive contracts, data support verification, and dynamic gating flow control.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A task flow control method for a multi-agent system, characterized in that, include: Obtain the execution plan for the target task, wherein the execution plan contains multiple task nodes; Load the associated multidimensional interaction contract for the current task node, and control the matching execution agent to execute the current task node under the constraints of the multidimensional interaction contract to generate the target product. The target product includes a supporting data set and output conclusions. The multidimensional interaction contract is used to limit tool call constraints, product output specifications, and product verification rules. Based on the product verification rules in the multidimensional interaction contract, a multidimensional verification is performed on the target product, and the multidimensional verification includes at least a data support degree verification. Based on the logical support relationship between the supporting data set and the output conclusion, the data support verification result of the target product is determined. Based on the results of the multi-dimensional verification, the uncertainty index of the target product is extracted, and combined with the risk assessment level of the associated nodes located downstream of the current task node in the multiple task nodes, dynamic gating flow control is performed on the target product.

2. The task flow control method for a multi-agent system according to claim 1, characterized in that, The multi-dimensional verification of the target product is performed based on the product verification rules in the multi-dimensional interaction contract. This multi-dimensional verification includes at least a data support verification, comprising: Obtain the supporting data set contained in the target product and the output conclusion of the target product; Evaluate the quantity, data type, and reliability level of the supporting data in the supporting data set.

3. The task flow control method for a multi-agent system according to claim 1, characterized in that, The uncertainty index of the target product is extracted based on the results of the multi-dimensional verification, and dynamic gating flow control is performed on the target product in combination with the risk assessment level of downstream related nodes, including: The confidence index and data conflict index of the target product are extracted from the results of the multi-dimensional verification and used as the uncertainty index. Based on the confidence index, the data conflict index, and the risk assessment level of the downstream associated nodes, the gating triggering conditions are determined. In response to the fulfillment of the gating triggering condition, a corresponding gating action is performed on the target product, the gating action including at least one of data interception, rework and recertification, upgrade review, or route release.

4. The task flow control method for a multi-agent system according to claim 1, characterized in that, The multidimensional interaction contract includes tool call constraints and product output specifications; The tool invocation constraints are used to limit the whitelist boundaries of the tools that the executing agent can invoke and the types of actions that are prohibited from being executed; The product output specification is used to define the input and output data structure, required fields, and data source anchor points of the target product.

5. The task flow control method for a multi-agent system according to claim 2, characterized in that, The method further includes: In response to the data support verification result indicating that the data support of the target product does not meet the preset support threshold, the flow of the target product to the downstream associated node is blocked. The downstream associated nodes include high-cost execution nodes whose resource consumption exceeds a preset consumption threshold; While blocking the flow, a supplementary data acquisition task is automatically generated and triggered. The supplementary data acquisition task is used to acquire supplementary support data that meets the preset support threshold, and to integrate the supplementary support data into the target product to re-execute the multi-dimensional verification.

6. The task flow control method for a multi-agent system according to claim 3, characterized in that, The process of determining the gating trigger conditions based on the confidence index, the data conflict index, and the risk assessment level of the downstream associated nodes includes: When the risk assessment level of the downstream associated node exceeds the preset risk threshold, a high-risk gating condition is triggered. The gating action corresponding to the high-risk gating condition is a mandatory upgrade review, regardless of whether the confidence index meets the standard. When the data conflict indicator indicates that the supporting data set contains both a first type of data that supports the output conclusion and a second type of data that refutes the output conclusion, a conflict gating condition is determined to be triggered. The gating action corresponding to the conflict gating condition includes generating a data conflict report and intercepting the flow of the target product.

7. The task flow control method for a multi-agent system according to claim 4, characterized in that, The multi-dimensional verification of the target product based on the product verification rules in the multi-dimensional interactive contract also includes structural verification and permission verification of the product output specification. The method further includes: In response to the structure verification indicating that the input and output data structure of the target product is missing a required field or the data unit does not meet the parsing requirements of the downstream associated node, the target product is returned to the executing agent along with a structure verification report. In response to the permission verification indicating that the executing agent has invoked an unauthorized tool or unauthorized data source outside the whitelist boundary, the target product is marked as untrusted, and an unauthorized audit process for the executing agent is triggered.

8. The task flow control method for a multi-agent system according to claim 1, characterized in that, The method further includes: In response to the completion of the dynamic gating flow control of the target product, a signature product package is generated based on the data hash value of the target product, the verification report of the multi-dimensional verification, and the data source reference. Assign a globally unique identifier to the signature product packet and write the globally unique identifier into downstream task requests and system logs to support product status tracking across task nodes; In response to the detection of a change in the input data or associated model version of the target product, the affected downstream associated nodes are located based on the globally unique identifier, and a local state update is performed on the affected downstream associated nodes.

9. A task flow control system for a multi-agent system, characterized in that, include: The plan acquisition unit is used to acquire the execution plan of the target task, wherein the execution plan contains multiple task nodes; The contract loading and execution unit is used to load the associated multidimensional interaction contract for the current task node, and control the matching execution agent to execute the current task node under the constraints of the multidimensional interaction contract to generate the target product. The target product includes a supporting data set and output conclusions. The multidimensional interaction contract is used to limit tool call constraints, product output specifications and product verification rules. The product verification unit is used to perform multi-dimensional verification on the target product based on the product verification rules in the multi-dimensional interaction contract. The multi-dimensional verification includes at least data support verification. The gated flow unit is used to extract the uncertainty index of the target product based on the results of the multi-dimensional verification, and combine it with the risk assessment level of the downstream related nodes to perform dynamic gated flow control on the target product. The task flow control system of the multi-agent system is used to execute the task flow control method of the multi-agent system as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the task flow control method for a multi-agent system as described in any one of claims 1 to 8.

Images