New product insight analysis method and system based on multi-intelligent agent collaborative arrangement

Abstract

The application provides a new product insight analysis method and system based on multi-intelligent agent collaborative arrangement, and relates to the technical field of data analysis. The method comprises the following steps: analyzing user intention and initializing analysis parameters to construct a unified analysis context; deploying a plurality of independent parallel analysis agents in the first stage, each agent completes analysis based on its own data source and writes structured conclusions back to the unified analysis context, and the parallel agents do not call each other; deploying an evidence chain diagnosis agent in the bridging stage, automatically selecting one of the three execution modes of completeness, limitation or baseline according to the source verification coverage, establishing a field-level data snapshot with a state label and calculating a concept readiness score; deploying a read-only agent in the downstream stage, which strictly reads only the existing conclusions in the unified analysis context and inherits the fields with permission status in the order of dependence, calculates a multi-dimensional weighted opportunity score for the candidate new product direction, and controls the downstream permission and blocking in combination with the concept readiness score; and maintaining a total control state machine, which reschedules the corresponding agent when the state of the key field in the context changes or the blocking item is triggered. The application effectively improves the consistency, traceability and conclusion reliability of new product insight analysis.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a new product insight analysis method and system based on multi-intelligent agent collaborative orchestration. Background Technology

[0002] Current new product insight analysis typically relies on a single analytical tool or a single large language model, processing issues such as market size, consumer demand, and competitor landscape separately, and then manually piecing together the conclusions of each part into an analytical report. For example, companies often use spreadsheets to compile sales data, use questionnaires to obtain consumer preferences, and then combine competitor data for qualitative judgments, with analysts ultimately summarizing and writing the report. This approach is mostly based on discrete and fragmented analytical steps, lacking a unified scheduling and context management mechanism, making it difficult for the conclusions of each step to remain consistent within the same context.

[0003] Some improvement solutions attempt to introduce a single large language model to process the entire process end-to-end, or to delegate the processing logic of each analysis stage to the large language model for sequential execution in the form of configuration files. However, due to the lack of explicit constraints on the dependencies between analysis stages and the lack of structured management of intermediate conclusions between stages, the model is prone to deviating from the conclusions produced in previous stages in subsequent stages, regenerating or inferring content that contradicts previous conclusions. This leads to inconsistencies in the final report and problems with the inability to trace the source of conclusions. At the same time, existing methods lack hierarchical and verification mechanisms for the data sources on which the analysis is based, often directly including unverified marketing statements, second-hand information, or model memory as facts in the conclusions, making it difficult to guarantee the credibility of the analysis results.

[0004] Furthermore, existing methods lack explicit gate control over the sufficiency of upstream evidence and the verification of sources when generating new product concepts, scoring opportunities, or outputting external communication content. This can easily lead to the direct output of product-level conclusions or external claims without sufficient evidence, resulting in compliance and decision-making risks. Existing methods also typically provide only one-time analysis results. When upstream data or verification status changes, they cannot automatically identify the affected links and re-execute them accordingly, causing downstream conclusions to become disconnected from the latest data status. This static, fragmented analytical framework, lacking source governance and stage dependency constraints, limits the performance of new product insight analysis in terms of consistency, traceability, and the credibility of conclusions. Summary of the Invention

[0005] This invention provides a new product insight analysis method and system based on multi-intelligent agent collaborative orchestration, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a new product insight analysis method based on multi-intelligent agent collaborative orchestration, comprising: parsing user input, extracting analysis parameters including target category, core object identifier, object identifier to be analyzed, analysis target, analysis market, and data cutoff date; writing the analysis parameters into a unified analysis context, and initializing a general control state machine for recording the execution status of each agent and downstream permissions; deploying multiple independent parallel analysis agents in a first stage, each parallel analysis agent being bound to its own data source and analysis responsibility, performing analysis in parallel without calling or referencing each other's intermediate results, and writing its own analysis conclusions back to the unified analysis context in the form of structured fields; deploying an evidence chain diagnostic agent in a bridging stage, the evidence chain diagnostic agent scanning local data resources and performing source verification on the object to be analyzed, automatically selecting an execution mode from full diagnostic mode, limited diagnostic mode, and baseline diagnostic mode based on verification coverage, and establishing a value, source, and other parameters for each key field. The system takes field-level data snapshots of source hierarchy and status labels, calculates concept readiness scores based on multiple diagnostic sub-items, and writes the field-level data snapshots, execution modes, and concept readiness scores back to the unified analysis context. In the downstream stage, multiple read-only downstream agents are deployed. Each read-only downstream agent reads only the existing structured conclusions in the unified analysis context according to a preset dependency order, and inherits fields marked as directly usable or conditionally usable in the field-level data snapshot according to the field's permission status. Regeneration or inference of conclusions not existing in the context is prohibited. A multi-dimensional weighted opportunity score is calculated for candidate new product directions, and the multi-dimensional weighted opportunity score is placed alongside the concept readiness score. Downstream permission and blocking controls are executed on subsequent agents based on the comparison result of the concept readiness score and a preset threshold. The system maintains completed agents, agents to be executed, field permission status, and blocking items in the overall control state machine. When the status of key fields in the unified analysis context changes or a blocking item appears, the corresponding agent is rescheduled.

[0007] In one optional embodiment, in the first phase, multiple independent parallel analysis agents are deployed, each of which writes its analysis conclusions back to the unified analysis context in the form of structured fields, including:

[0008] The multiple independent parallel analysis agents are divided into a macro engine agent group and a micro engine agent group. The macro engine agent group includes an omnichannel market performance analysis agent, a consumer profile and scenario insight analysis agent, and a competitor landscape and white space analysis agent. The micro engine agent group includes a consumer original voice insight analysis agent and a social media content trend analysis agent.

[0009] Before execution, each of the parallel analysis agents loads its own role settings and data governance rules. The data governance rules include arbitrating information conflicts that are comparable in time, for the same analysis problem, the same market, and the same object version, according to the priority of customer-provided data, industry authoritative source data, general authoritative source data, and other source data.

[0010] Each of the parallel analysis agents generates analysis results according to its own analysis steps, and organizes the analysis results into a preset structured field template. The structured field template includes a conclusion field, a quantitative indicator field, a data source field, and a data credibility level field.

[0011] Each parallel analysis agent does not call or reference the other's intermediate results within the same stage. They only write the structured fields back to the unified analysis context after completing their own analysis, so that downstream agents in subsequent stages can read them uniformly.

[0012] In an optional embodiment, the evidence chain diagnostic agent automatically selects the execution mode and establishes field-level data snapshots based on verification coverage, including:

[0013] The evidence chain diagnostic agent recursively scans local data resources and reads the source list, performs official source verification for each object to be analyzed, and records the official source verification status, official source link, verification date, target market and object version.

[0014] When the target object to be analyzed and at least one comparison object are both verified by official sources and have usable materials, select the full diagnostic mode; when only some objects or some fields are verified by official sources, select the limited diagnostic mode and create a supplementary list for the objects and fields that have not passed verification; when no target object that can be verified by official sources is found, select the baseline diagnostic mode and only output the baseline of evidence, experience and visual asset requirements at the category level.

[0015] For each key field, establish a field-level data snapshot containing the value, source, source level, acquisition time, market and version, and status label. The status label is taken from one of the following: verified, marketing statement only, incomplete conditions, cannot be directly compared, to be tested, and prohibited from use.

[0016] In the selected execution mode, network retrieval is prohibited when local data is sufficient; when local data is partially missing, external data collection is performed only for the missing fields; and external data collection is performed according to the source priority when local data is completely missing. Objects and fields that have not passed official source verification are prohibited from entering object-level benchmarking, comparison conclusions, recommendations, or external content.

[0017] In one optional embodiment, the read-only downstream agent inherits fields according to their permission status and performs downstream permission and blocking control, including:

[0018] The overall control state machine maintains a permission state for each field in the field-level data snapshot. The permission state is taken from one of direct use, conditional use, and prohibited use. The field marked as conditional use also carries its restriction description and source.

[0019] When reading the unified analysis context, each read-only downstream agent inherits only fields with a permission status of direct use or conditional use. For inherited conditional use fields, the restriction description is inherited synchronously. Fields with a permission status of prohibited use are rejected for inheritance and marked as not covered in the output.

[0020] When the conclusion of a certain preceding agent is not written back to the unified analysis context, the read-only downstream agent that depends on the conclusion will mark the corresponding field as not covered and prompt the corresponding preceding agent to be executed first, prohibiting manual rewriting.

[0021] The central control state machine performs permission or blocking actions on downstream agents based on the field permission status and blocking items. When a blocking item is marked as needing to be repaired, the blocking is performed on subsequent agents that depend on that field until the blocking item is removed.

[0022] In one optional embodiment, calculating a multidimensional weighted chance score for candidate new product directions includes:

[0023] Read the candidate new product directions output by the preceding agent from the unified analysis context, as well as the quantitative indicators required to calculate the multidimensional weighted opportunity score;

[0024] For each candidate new product direction, a relevance score, engagement score, coverage score, and innovation score are calculated. The relevance score is determined based on the conclusions of the consumer profile and scenario insight analysis agency and the consumer voice insight analysis agency. The engagement score is determined based on the conclusions of the social media content trend analysis agency. The coverage score is determined based on the conclusions of the omnichannel market performance analysis agency. The innovation score is determined based on the conclusions of the competitor landscape and white space analysis agency.

[0025] Preset weights are assigned to the relevance score, participation score, coverage score and innovation score respectively. The scores are weighted and summed with their corresponding weights to obtain the multidimensional weighted opportunity score for the candidate new product direction. A comprehensive rating is assigned to the candidate new product direction based on the comparison between the multidimensional weighted opportunity score and the preset rating range.

[0026] Using coverage score as the first coordinate axis, innovation score as the second coordinate axis, the sum of relevance score and participation score as the marker size, and comprehensive rating as the marker color, a new product opportunity distribution map is generated, and the priority of candidate new product directions is output from high to low according to the multi-dimensional weighted opportunity score.

[0027] In one optional embodiment, the step of calculating a concept readiness score based on multiple diagnostic sub-items and performing downstream permission and blocking control based on a comparison result of the concept readiness score and a preset threshold includes:

[0028] Five diagnostic sub-items are set up: claim clarity, evidence completeness, competitor differentiation, experience fulfillment, and visual presentation readiness, and preset weights are assigned to each diagnostic sub-item.

[0029] For the current object to be analyzed, the scores of each diagnostic item are evaluated one by one. The scores of each diagnostic item are weighted and summed with their corresponding weights to obtain a concept readiness score ranging from zero to one hundred. The concept readiness score is calculated independently and does not change the multidimensional weighted chance score.

[0030] Set a first threshold and a second threshold below the first threshold. When the concept readiness score is not less than the first threshold, allow subsequent concept board generation agents and consumer verification agents to proceed. When the concept readiness score is between the second threshold and the first threshold, require the completion of critical repair tasks before proceeding. When the concept readiness score is less than the second threshold, block the final concept board generation agent and return to concept optimization.

[0031] When the evidence chain diagnostic agent in the bridging phase is skipped or not executed, the concept readiness score is marked as unevaluated, and estimation or backfilling is prohibited.

[0032] In an optional embodiment, the downstream stage further includes deploying a demand space mapping agent to perform exponential clustering, including:

[0033] The demand space mapping agent reads the conclusions written back by the consumer original voice insight analysis agent, the competitor landscape and white space analysis agent, and the social media content trend analysis agent. By conducting structured quantitative question and answer with a preset number of simulated consumers, it collects quantitative data including purchase trigger scenarios, purchase obstacles, demographic information, product accessibility, and core attitudes.

[0034] For each feature, the index value of the feature is calculated by multiplying the ratio of the feature’s proportion in the target segment population to the feature’s proportion in the total sample by the base. When the index value is higher than a preset upper threshold, it is marked as a highly significant feature; when the index value is lower than a preset lower threshold, it is marked as a low significant feature.

[0035] Using the target audience as the first clustering axis, usage scenarios and trigger points as the second clustering axis, and core demands and obstacles as the third clustering axis, the collected quantitative data is cross-clustered to generate multiple independent demand spaces, and the sub-groups identified in each demand space are given names.

[0036] For each demand space, the purchase trigger scenarios and purchase obstacles are sorted from high to low according to their corresponding index values. High-index trigger scenarios and high-index obstacles that are not yet covered by existing comparison objects are marked as white spaces. For each demand space, a segmented population profile is output, including market consumption scale, relevance and growth momentum, demographics, accessibility, core attitudes, trigger points, obstacles, and competitor positioning.

[0037] In one alternative embodiment, the deployment includes:

[0038] Configure corresponding skill description files for the parallel analysis agent, evidence chain diagnosis agent and read-only downstream agent respectively. Each skill description file records the corresponding agent's role setting, data governance rules, execution steps, and structured field template for writing back the conclusion, as well as routing rules containing the identifier of the preceding dependent agent and the triggering conditions.

[0039] The skill scheduling engine parses user input to determine the set of proxies to be executed, and constructs a dependency graph between proxies based on the pre-dependent proxies identifiers recorded in each skill description file. The execution order and parallel grouping of each proxy are determined according to the dependency graph.

[0040] When scheduling the execution of a certain agent, the skill description file corresponding to the agent and the existing structured conclusions in the unified analysis context are provided to the large language model. The large language model then performs the corresponding analysis or orchestration duties based on the skill description file and writes the output back to the unified analysis context in structured text form according to the structured field template.

[0041] The skill scheduling engine controls whether to schedule the skill description file corresponding to the downstream agent that depends on the corresponding field based on the field permission status and blocking items recorded in the master control state machine, thereby implementing the downstream permission and blocking control in a programmatic manner.

[0042] A second aspect of this invention provides a new product insight analysis system based on multi-intelligent agent collaborative orchestration, comprising: a parameter initialization unit, configured to parse user input and extract analysis parameters, write the analysis parameters into a unified analysis context, and initialize a central control state machine; a parallel analysis unit, configured to deploy multiple independent parallel analysis agents in a first phase, wherein each parallel analysis agent performs analysis in parallel without calling each other and writes structured conclusions back to the unified analysis context; an evidence chain diagnosis unit, configured to deploy evidence chain diagnosis agents, automatically select execution modes based on source verification coverage, establish field-level data snapshots with status labels, and calculate concept readiness scores; a downstream orchestration unit, configured to deploy multiple read-only downstream agents, read only existing conclusions in the unified analysis context in dependency order and inherit fields according to permission status, calculate multidimensional weighted opportunity scores, and execute downstream permission and blocking control based on concept readiness scores; and a status monitoring unit, configured to maintain the central control state machine, and trigger rescheduling of corresponding agents when the status of key fields in the unified analysis context changes or blocking items appear.

[0043] A third aspect of the present invention provides an electronic device, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the aforementioned method.

[0044] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0045] In this embodiment of the invention, multiple intelligent analysis agents are collaboratively orchestrated through a unified analysis context and a central control state machine. The new product insight analysis is divided into three stages: parallel analysis, evidence chain bridging, and read-only downstream processing. Explicit stage dependencies constrain the read and write behavior of each agent, significantly improving the consistency of conclusions across the entire analysis chain. In the first stage, multiple parallel analysis agents operate independently and do not call each other, each writing structured conclusions back to the unified analysis context. This ensures the independence of each analysis perspective and provides a unified, comparable, and traceable source of conclusions for subsequent stages, avoiding conclusion splicing errors caused by fragmented analysis processes. In the bridging stage, the evidence chain diagnostic agent automatically selects three execution modes—complete, restricted, or baseline—based on source verification coverage. It also creates a field-level data snapshot with status tags for each key field. Combined with priority arbitration of customer-provided data, industry authoritative sources, general authoritative sources, and other sources, as well as an official source verification gate, unverified data cannot enter object-level conclusions, effectively... This enhances the credibility and compliance of the analysis results. The read-only agents in the downstream stage strictly read only existing conclusions from the context and inherit fields according to their permission status, prohibiting regeneration or inference. This fundamentally avoids inconsistencies caused by large language models rewriting themselves without prior conclusions. The multi-dimensional weighted opportunity scoring quantifies and ranks candidate new product directions based on four dimensions: relevance, engagement, coverage, and innovation. Each dimension is derived from the conclusions of designated upstream agents, ensuring clear data support and interpretability for the scoring results. Concept readiness scoring is calculated independently of opportunity scoring and uses dual thresholds to control the generation and validation of downstream concept boards and consumer validation, ensuring that product-level conclusions or external content are not output when evidence is insufficient. The central control state machine continuously monitors the state changes and blocking items of key fields in the context. Once upstream data or verification status is updated, affected agents can be rescheduled in a targeted manner, preventing downstream conclusions from becoming disconnected from the latest data status and significantly improving the timeliness and accuracy of new product insight analysis. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the new product insight and analysis method based on multi-intelligent agent collaborative orchestration, as described in an embodiment of the present invention.

[0047] Figure 2 This is a logical framework diagram of the three-stage proxy orchestration and stage dependency relationship in Embodiment 3 of the present invention;

[0048] Figure 3 This is a flowchart illustrating the automatic selection of the evidence chain diagnosis proxy execution mode in an embodiment of the present invention.

[0049] Figure 4 This is a schematic diagram illustrating the field permission state propagation of the unified analysis context and the overall control state machine in this embodiment of the invention;

[0050] Figure 5 This is a schematic diagram of the multidimensional weighted opportunity score and new product opportunity distribution diagram in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] The technical solution of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. It should be noted that the embodiments of the present invention are illustrated using new product insight analysis in a certain consumer goods industry as an example, but the technical solution of the present invention is not limited to a specific industry and is equally applicable to insight analysis scenarios for other new consumer goods products.

[0053] Figure 1 This is a flowchart illustrating a new product insight analysis method based on multi-intelligent agent collaborative orchestration, as described in an embodiment of the present invention. The method includes: parsing user input and initializing analysis parameters to construct a unified analysis context; deploying multiple independent parallel analysis agents in the first stage, with each agent writing structured conclusions back to the unified analysis context; deploying an evidence chain diagnostic agent in the bridging stage, automatically selecting the execution mode and establishing field-level data snapshots and concept readiness scores; deploying read-only downstream agents in the downstream stage, inheriting fields according to permission status and calculating multi-dimensional weighted opportunity scores, and executing downstream permission and blocking controls based on the concept readiness scores; maintaining the overall control state machine and rescheduling corresponding agents when key field states change or blocking items appear.

[0054] In one specific implementation, the agents and orchestration logic described in this invention are implemented through a workflow, where each node of the workflow corresponds to one agent. The system includes a skill scheduling engine, skill description files corresponding one-to-one with each agent, and a large language model. Each skill description file is a structured document that records the corresponding agent's role settings, data governance rules, execution steps, and structured field templates for writing back the conclusion, as well as routing rules containing identifiers of prerequisite dependent agents and triggering conditions. The skill scheduling engine parses user input to determine the set of agents to be executed, and constructs a dependency graph between agents based on the identifiers of prerequisite dependent agents recorded in each skill description file. It then determines the execution order and parallel grouping of each agent according to the dependency graph. Multiple agents without dependencies are grouped into the same parallel group for parallel execution in the first stage, while agents with dependencies are executed sequentially in subsequent stages according to their topological order.

[0055] When scheduling an agent for execution, the skill scheduling engine provides the skill description file corresponding to the agent and the existing structured conclusions in the unified analysis context to the large language model. The large language model then completes the corresponding analysis or orchestration responsibilities according to the roles, data governance rules, and execution steps specified in the skill description file, and writes the output back to the unified analysis context in structured text format according to the structured field templates specified in the skill description file. The unified analysis context and the central control state machine are stored in a persistent, readable, and writable structured text format. Each agent's writing back and reading of the context uses this structured text as the carrier, rather than relying on the implicit memory of the large language model. Based on the field permission status and blocking items recorded in the central control state machine, the skill scheduling engine controls whether to provide the skill description file corresponding to the downstream agent that depends on the corresponding field to the large language model for execution. When the permission status of a field is prohibited or there is a blocking item marked as requiring repair, the skill scheduling engine does not schedule the downstream agent that depends on that field, thus implementing the aforementioned downstream permission and blocking control programmatically. By adopting the above implementation method, the domain logic of each link in the new product insight analysis is decoupled from the scheduling engine in the form of skill description files. This not only facilitates the independent maintenance and replacement of the analysis logic of a single link, but also ensures the consistency, traceability and reproducibility of the overall analysis link through the programmatic control of the context state and permission state by the scheduling engine.

[0056] In one specific implementation, such as Figure 2As shown, the system sets up an orchestration agent as the sole central control entry point, responsible for parameter initialization, agent routing, precondition checks, context state writing back, and downstream call permissions. Other business agents only perform their respective duties. The orchestration agent first parses the user input, extracts the analysis parameters, and writes them into the unified analysis context. These analysis parameters include fields such as target category, core object identifier, the identifier of the target object to be analyzed, the identifier of the comparison object, analysis objective, key issues, analysis market, data cutoff date, and analysis depth. If a parameter is not explicitly given in the user input, it is written as a default value or marked as unspecified, without affecting the startup of subsequent agents.

[0057] While writing analysis parameters, the orchestration agent initializes the central control state machine. This central control state machine maintains state fields such as the list of completed agents, the list of agents to be executed, the chain of evidence diagnostic status, the diagnostic execution mode, object source verification records, data patterns, object comparability, opportunity scoring permissions, concept readiness scores, and downstream permissions and blocking items, using structured fields rather than natural language memory. The central control state machine runs throughout the entire analysis chain. After execution, each agent must write its conclusions back to the unified analysis context using structured fields and update the corresponding state in the central control state machine, thereby ensuring that the orchestration agent can make scheduling decisions based on explicit states rather than implicit memory.

[0058] In one optional embodiment, in the first stage, multiple independent parallel analysis agents are deployed. Each parallel analysis agent writes its analysis conclusions back to the unified analysis context in the form of structured fields. This includes: dividing the multiple independent parallel analysis agents into a macro-engine agent group and a micro-engine agent group; loading each parallel analysis agent's role settings and data governance rules before execution; generating analysis results according to each parallel analysis agent's analysis steps and organizing them into preset structured field templates; and ensuring that each parallel analysis agent does not call or reference each other's intermediate results within the same stage, only writing the structured fields back to the unified analysis context after completing its own analysis.

[0059] In one specific implementation, the multiple parallel analysis agents in the first phase include: an omnichannel market performance analysis agent, used to analyze the market size, growth trend, channel structure, and price range evolution of the target category based on omnichannel sales data, and to identify high-growth sub-segments; a consumer profile and scenario insight analysis agent, used to analyze consumer concerns and construct target audience profiles based on search behavior and hot keyword data; a competitor landscape and untapped market analysis agent, used to track the dynamics of comparison targets, dissect their positioning structure, and identify untapped market opportunities; a consumer voice insight analysis agent, used to extract high-frequency pain points and unmet needs from consumer review data; and a social media content trend analysis agent, used to analyze content traffic trends and content opportunities on social media platforms. These agents belong to the macro-engine agent group and the micro-engine agent group, each bound to different data sources and analysis responsibilities.

[0060] Before execution, each parallel analysis agent loads its respective role settings and data governance rules. These rules stipulate that for information conflicts involving the same analytical problem, the same market, the same object version, and comparable timeframes, arbitration is conducted in the following order of priority: client-provided data, industry-authoritative source data, general authoritative source data, and other source data. In cases of source conflicts, all conflicting values ​​are retained, and the version, market, and time are recorded; silent selection of favorable values ​​is prohibited. After each agent generates its analysis results according to its respective analysis steps, the results are organized into a pre-defined structured field template. This template includes at least a conclusion field, a quantitative indicator field, a data source field, and a data credibility level field, such as structured conclusions like market size, fastest-growing channels, mainstream price ranges, high-growth sub-segments, competitive niches, high-frequency pain points, and content trends.

[0061] It is worth emphasizing that the parallel analysis agents in the first phase are independent of each other and operate independently. Within the same phase, they do not call or reference each other's intermediate results, and only write the structured fields back to the unified analysis context after their respective analyses are completed. This parallel and independent design allows the agents to execute concurrently to shorten the overall analysis time, while ensuring the independence of each analysis perspective and preventing intermediate biases from one agent from spreading within a phase. At the same time, it provides downstream agents in subsequent phases with a unified, comparable, and traceable source of conclusions.

[0062] In one optional embodiment, the evidence chain diagnostic agent automatically selects the execution mode and establishes field-level data snapshots based on verification coverage, including: recursively scanning local data resources and reading the source list, performing official source verification for each object to be analyzed; automatically selecting between full diagnostic mode, limited diagnostic mode, and baseline diagnostic mode based on verification coverage; establishing a field-level data snapshot for each key field, including value, source, source level, acquisition time, market and version, and status label; and constraining the network retrieval scope according to local data coverage under the selected execution mode.

[0063] In one specific implementation, such as Figure 3 As shown, the Evidence Chain Diagnostic Agent, as a dedicated bridging module connecting market and consumer insights with product concept and communication design, is fixed after the multiple basic analyses in the first stage and before demand space mapping and new product direction refinement. It is automatically registered as a pending task by the orchestration agent upon receiving a new product insight request and will not be closed or skipped due to the user not specifying it, not confirming the scope a second time, or not initially providing the object to be analyzed. After the Evidence Chain Diagnostic Agent is started, it first recursively scans the local data resource directory and reads the source list, identifying official page snapshots, materials, reviews, and prior conclusions of both the user's own object and the comparison object.

[0064] The Evidence Chain Diagnostic Agent performs official source verification for each specific object to be analyzed, recording the official source verification status, official source link, verification date, target market, and object version. Based on verification coverage, the Evidence Chain Diagnostic Agent automatically selects the execution mode: when both the client object and at least one comparison object have passed official source verification and have available materials or evidence, the complete diagnostic mode is selected to complete multi-dimensional benchmarking, evidence coverage matrix, concept readiness scoring, and action list; when only some objects or some fields can complete official source verification, the restricted diagnostic mode is selected to perform diagnostics on verified objects and fields, while the rest are marked as pending verification or data is temporarily unavailable; when no client object that can complete official source verification is found, the baseline diagnostic mode is selected, outputting only the baseline of claims, evidence, experience, and visual asset requirements for that category, as well as a list of objects, materials, and test fields to be supplemented.

[0065] The evidence chain diagnostic agent establishes a field-level data snapshot for each key field. Each key field records at least its value, source, source level, acquisition time, market and version, and status label. The status label is taken from one of the following: verified, marketing statement only, incomplete conditions, cannot be directly compared, pending testing, or prohibited from use. In the selected execution mode, network retrieval is prohibited when local data is sufficient. When local data is partially missing, external data collection is only performed for the missing fields. When local data is completely missing, external data collection is performed according to the aforementioned source priority. Any new data must have its source, time, market and version, and testing conditions written back. Official source verification constitutes a mandatory authenticity gate for object-level conclusions. Objects and fields that fail official source verification cannot be included in object-level benchmarking, comparison conclusions, recommendations, or external content; they can only be retained as objects pending verification or anonymous aggregated data.

[0066] In one optional embodiment, the read-only downstream agent inherits fields based on their permission status and performs downstream permission and blocking control, including: maintaining a permission status for each field in the field-level data snapshot, which is taken from one of direct use, conditional use, and prohibited use; each read-only downstream agent only inherits fields with a permission status of direct use or conditional use, and synchronously inherits restriction descriptions for conditional use fields; when a certain preceding agent's conclusion is not written back to the context, marking the corresponding field as not covered and prompting the preceding agent to be executed first; the central control state machine performs allow or block actions on the downstream agents based on the field permission status and blocking items.

[0067] In one specific implementation, such as Figure 4 As shown, the central control state machine maintains a permission state for each field in the field-level data snapshot. The permission state is taken from one of three: directly usable, conditionally usable, or prohibited. Fields marked as conditionally usable also carry their restriction descriptions and sources. Downstream agents from the second stage onwards, including demand space mapping agents, new product direction extraction agents, new product track strategy analysis agents, new product opportunity quantitative scoring agents, and the third stage's new product concept aggregation agents, market size calculation agents, concept board generation agents, content solution generation agents, and consumer validation agents, are all read-only downstream agents, with their sole data source being the unified analysis context.

[0068] When reading the unified analysis context, each read-only downstream agent strictly reads only the existing structured conclusions. It is prohibited to regenerate, infer, or supplement any conclusions not present in the context based on its own knowledge. It only inherits fields with a permission status of direct use or conditional use. For inherited conditional use fields, it simultaneously inherits their restriction descriptions and sources. Fields with a permission status of prohibited use are rejected for inheritance and marked as not covered in the output. When a conclusion from a preceding agent is not written back to the unified analysis context, the read-only downstream agent relying on that conclusion marks the corresponding field as not covered and prompts the corresponding preceding agent to be executed first, prohibiting it from manually adding the missing data. For example, the new product direction extraction agent must inherit fields such as recommendation proposition, benefits, supporting reasons, credibility, experience and usage, and visual tasks; if missing, it is marked as not covered and not manually added. The external content solution generation agent must not use fields marked as pending testing, not directly comparable, or prohibited.

[0069] The central control state machine grants or blocks downstream agents based on field permission status and blocking items. When a blocking item is marked as requiring repair, the blocking depends on subsequent agents for that field until the blocking item is lifted. Through the hierarchical propagation and blocking control of field-level permission status, this embodiment ensures at the mechanism level that each field of the downstream conclusion is traceable to its source and verification status, and avoids generating object-level facts, comparative conclusions, or external claims when evidence is insufficient or the object has not been verified.

[0070] In one optional embodiment, calculating a multidimensional weighted opportunity score for candidate new product directions includes: reading candidate new product directions and required quantitative indicators from a unified analysis context; calculating a relevance score, participation score, coverage score, and innovation score for each candidate direction, with each score determined based on the conclusions of a designated upstream agent; configuring preset weights for each score and summing them to obtain a multidimensional weighted opportunity score, and assigning a comprehensive rating accordingly; generating a new product opportunity distribution map and outputting priority rankings from high to low scores.

[0071] In one specific implementation, the new product opportunity quantification scoring agent reads candidate new product directions, typically two to five, from the unified analysis context, output by the preceding agent, while simultaneously backtracking to calculate the required key quantitative indicators. For each candidate new product direction, scores are calculated across four dimensions: relevance score, measuring the degree to which the direction aligns with consumers' core pain points and high-frequency concerns, determined based on the conclusions of the consumer profile and scenario insight analysis agent and the consumer voice insight analysis agent; engagement score, measuring the potential for the direction to generate discussion and dissemination on social media, determined based on the conclusions of the social media content trend analysis agent; coverage score, measuring the market size and growth potential corresponding to the direction, determined based on the conclusions of the omnichannel market performance analysis agent; and innovation score, measuring the degree to which the direction avoids competitive red oceans and occupies untapped markets, determined based on the conclusions of the competitor landscape and untapped market analysis agent.

[0072] Preset weights are assigned to relevance score, engagement score, coverage score, and innovation score. In one example, the relevance weight is 30%, engagement weight is 20%, coverage weight is 30%, and innovation weight is 20%. The weighted sum of the four dimensions' scores and their corresponding weights yields a multi-dimensional weighted opportunity score ranging from 0 to 100. Based on the comparison between this score and a preset rating range, a comprehensive rating is assigned to the candidate new product direction. For example, a score of 80 or higher is the highest rating, 70 to 79 is the second highest rating, 60 to 69 is a middle rating, and less than 60 is a recommended elimination rating. Further, a new product opportunity distribution map is generated using the coverage score as the first axis, the innovation score as the second axis, the sum of the relevance score and engagement score as the marker size, and the comprehensive rating as the marker color. Figure 5 As shown, the closer the marker is to the high value region of the first and second coordinate axes and the larger the marker size, the more worthy it is to prioritize investment in the candidate new product direction; finally, the priority ranking of the candidate new product directions is output from high to low according to the multidimensional weighted opportunity score.

[0073] In one optional embodiment, calculating a concept readiness score based on multiple diagnostic sub-items and executing downstream permission and blocking controls accordingly includes: setting five diagnostic sub-items—claim clarity, evidence completeness, competitor differentiation, experience fulfillment, and visual expression readiness—and configuring preset weights; evaluating each item of the current object and weighting and summing them to obtain a concept readiness score; setting a first threshold and a second threshold and executing three levels of permission or blocking on downstream agents accordingly; and marking the concept readiness score as unevaluated when the evidence chain diagnosis is skipped.

[0074] In one specific implementation, the concept readiness score is calculated independently by the evidence chain diagnostic agent without altering the multidimensional weighted opportunity score. Five diagnostic sub-items are set: claim clarity, evidence completeness, competitor differentiation, experience fulfillment, and visual expression readiness. In one example, these are assigned weights of 20%, 25%, 20%, 15%, and 20%, respectively. The score for each diagnostic sub-item is evaluated for the current object to be analyzed, and the scores of each sub-item are weighted and summed with their corresponding weights to obtain a concept readiness score ranging from zero to one hundred.

[0075] A first threshold and a second threshold below the first threshold are set. In one example, the first threshold is set to 80 and the second threshold to 65. When the concept readiness score is not less than the first threshold, the subsequent concept board generation agent and consumer verification agent are allowed to proceed. When the concept readiness score is between the second and first thresholds, the critical tasks marked as requiring repair must be completed before proceeding. When the concept readiness score is less than the second threshold, the final concept board generation agent is blocked and the process returns to concept optimization. When the evidence chain diagnostic agent in the bridging phase is skipped or not executed, the concept readiness score is marked as unevaluated, and estimation or backfilling is prohibited. By making the concept readiness score and the multidimensional weighted opportunity score parallel and independent of each other, this embodiment retains the quantitative ranking of business value by the opportunity score and independently controls the sufficiency of evidence through the readiness gate, avoiding the direct generation of external concept boards when the evidence is insufficient.

[0076] In one optional embodiment, the downstream stage further includes deploying a demand space mapping agent to perform index clustering, including: reading conclusions from original consumer insights, competitor landscape and white space, and social media content trends, and collecting quantitative data through structured quantitative question-and-answer sessions with a preset number of simulated consumers; calculating the index value of each feature according to the proportion ratio and marking its significance according to the threshold; performing cross-clustering with the target audience, usage scenarios and trigger points, and core demands and obstacles as three clustering axes to generate multiple demand spaces and naming sub-groups; sorting the trigger points and obstacles in each demand space according to the index value and marking the white space, and outputting a segmented audience profile.

[0077] In one specific implementation, the demand space mapping agent reads the conclusions written back by the consumer original voice insight analysis agent, the competitor landscape and white space analysis agent, and the social media content trend analysis agent. By conducting structured quantitative question and answer on a preset number of simulated consumers, it collects quantitative data including purchase trigger scenarios, purchase obstacles, demographic information, product accessibility, and core attitudes. The purchase trigger scenarios and purchase obstacles each list a sufficient number of candidate options for multiple selection.

[0078] For each feature, an index value is calculated by multiplying the ratio of the feature's proportion in the target sub-group to its proportion in the total sample by the base value. That is, the index value equals the proportion in the sub-group divided by the proportion in the total sample, then multiplied by the base value. When the index value is higher than a preset upper threshold, the feature is considered significantly higher than the overall feature in this sub-group and is marked as a highly significant feature; when the index value is lower than a preset lower threshold, the feature is considered significantly lower than the overall feature in this sub-group and is marked as a low-significant feature; values ​​between the upper and lower thresholds are considered within the normal range. Using the target sub-group as the first clustering axis, usage scenarios and trigger points as the second clustering axis, and core demands and obstacles as the third clustering axis, cross-clustering is performed on the collected quantitative data to generate multiple independent demand spaces. Within each demand space, several sub-groups with differentiated characteristics are identified and named.

[0079] For each demand space, its purchase trigger scenarios and purchase obstacles are sorted from high to low according to their corresponding index values. High-index trigger scenarios represent core opportunities for new product concepts, while high-index obstacles represent key hurdles that the new product concept needs to overcome. High-index trigger scenarios and high-index obstacles not yet covered by existing comparisons are marked as "white space." For each demand space, a segmented demographic profile is generated, including market consumption scale, relevance and growth momentum, demographics, accessibility, core attitudes, trigger points, obstacles, and competitor positioning. This profile allows new product teams to refine their strategies and anchor each candidate direction to the named demand space.

[0080] In one specific implementation, the agents in the second and third stages execute in an explicit dependency order. That is, after the completion of multiple basic analyses in the first stage, the following steps are executed sequentially: evidence chain diagnosis, demand space mapping, new product direction refinement, new product track strategy assessment, new product opportunity quantification scoring, market size calculation, and new product concept aggregation. Subsequently, users actively initiate steps such as concept board generation, consumer verification, content solution generation, and launch strategy. Among these, demand space mapping is a necessary prerequisite for new product direction refinement, and market size calculation is a necessary prerequisite for the new product concept aggregation report. Each downstream agent is blocked and prompted to complete the corresponding prerequisite analysis first if the conclusions of its prerequisite agents are not written back into the context.

[0081] In addition, after outputting analysis conclusions, the orchestration agent continuously monitors the state changes of key fields and blocking items in the unified analysis context. For each key field, it compares its current state with the state at the time of the last scheduling. When there is a field value update, a change in source verification status, or a new blocking item marked as requiring repair, the orchestration agent identifies the downstream agents affected by the field based on the agent dependencies recorded in the master control state machine and triggers the rescheduling of the corresponding agents. This ensures that the downstream conclusions are always consistent with the latest data and verification status without requiring a full recalculation of the entire analysis chain.

[0082] A second aspect of this invention provides a new product insight analysis system based on multi-intelligent agent collaborative orchestration, comprising: a parameter initialization unit, configured to parse user input and extract analysis parameters, write the analysis parameters into a unified analysis context, and initialize a central control state machine; a parallel analysis unit, configured to deploy multiple independent parallel analysis agents in a first phase, wherein each parallel analysis agent performs analysis in parallel without calling each other and writes structured conclusions back to the unified analysis context; an evidence chain diagnosis unit, configured to deploy evidence chain diagnosis agents, automatically select execution modes based on source verification coverage, establish field-level data snapshots with status labels, and calculate concept readiness scores; a downstream orchestration unit, configured to deploy multiple read-only downstream agents, read only existing conclusions in the unified analysis context in dependency order and inherit fields according to permission status, calculate multidimensional weighted opportunity scores, and execute downstream permission and blocking control based on concept readiness scores; and a status monitoring unit, configured to maintain the central control state machine, and trigger rescheduling of corresponding agents when the status of key fields in the unified analysis context changes or blocking items appear.

[0083] A third aspect of the present invention provides an electronic device, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the aforementioned method.

[0084] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0085] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A new product insight analysis method based on multi-intelligent agent collaborative orchestration, characterized in that, include: Parse user input, extract analysis parameters including target category, core object identifier, object identifier to be analyzed, analysis target, analysis market and data cutoff date, write the analysis parameters into a unified analysis context, and initialize the overall control state machine used to record the execution status of each agent and downstream permissions; In the first phase, multiple independent parallel analysis agents are deployed. Each parallel analysis agent is bound to its own data source and analysis responsibility. They perform analysis in parallel without calling or referencing each other's intermediate results, and write their analysis conclusions back to the unified analysis context in the form of structured fields. During the bridging phase, an evidence chain diagnostic agent is deployed. The evidence chain diagnostic agent scans local data resources and performs source verification on the object to be analyzed. Based on the verification coverage, it automatically selects an execution mode from the full diagnostic mode, the limited diagnostic mode, and the baseline diagnostic mode. For each key field, a field-level data snapshot containing the value, source, source level, and status label is created. Based on multiple diagnostic sub-items, a concept readiness score is calculated. The field-level data snapshot, execution mode, and concept readiness score are written back to the unified analysis context. In the downstream stage, multiple read-only downstream agents are deployed. Each read-only downstream agent reads only the structured conclusions already existing in the unified analysis context according to a preset dependency order, and inherits the fields marked as directly usable or conditionally usable in the field-level data snapshot according to the permission status of the fields. It is prohibited to regenerate or infer conclusions that do not exist in the context. A multidimensional weighted opportunity score is calculated for the candidate new product direction, and the multidimensional weighted opportunity score is listed alongside the concept readiness score. Downstream permission and blocking control is performed on subsequent agents based on the comparison result of the concept readiness score and a preset threshold. The system maintains completed agents, agents to be executed, field permission status, and blocking items in the overall control state machine. When the status of a key field in the unified analysis context changes or a blocking item appears, the system triggers the rescheduling of the corresponding agent.

2. The method according to claim 1, characterized in that, In the first phase, multiple independent parallel analysis agents are deployed. Each parallel analysis agent writes its analysis conclusions back to the unified analysis context in the form of structured fields, including: The multiple independent parallel analysis agents are divided into a macro engine agent group and a micro engine agent group. The macro engine agent group includes an omnichannel market performance analysis agent, a consumer profile and scenario insight analysis agent, and a competitor landscape and white space analysis agent. The micro engine agent group includes a consumer original voice insight analysis agent and a social media content trend analysis agent. Before execution, each of the parallel analysis agents loads its own role settings and data governance rules. The data governance rules include arbitrating information conflicts that are comparable in time, for the same analysis problem, the same market, and the same object version, according to the priority of customer-provided data, industry authoritative source data, general authoritative source data, and other source data. Each of the parallel analysis agents generates analysis results according to its own analysis steps, and organizes the analysis results into a preset structured field template. The structured field template includes a conclusion field, a quantitative indicator field, a data source field, and a data credibility level field. Each parallel analysis agent does not call or reference the other's intermediate results within the same stage. They only write the structured fields back to the unified analysis context after completing their own analysis, so that downstream agents in subsequent stages can read them uniformly.

3. The method according to claim 1, characterized in that, The evidence chain diagnostic agent automatically selects the execution mode and establishes field-level data snapshots based on verification coverage, including: The evidence chain diagnostic agent recursively scans local data resources and reads the source list, performs official source verification for each object to be analyzed, and records the official source verification status, official source link, verification date, target market and object version. When the target object to be analyzed and at least one comparison object are both verified by official sources and have usable materials, select the full diagnostic mode; when only some objects or some fields are verified by official sources, select the limited diagnostic mode and create a supplementary list for the objects and fields that have not passed verification; when no target object that can be verified by official sources is found, select the baseline diagnostic mode and only output the baseline of evidence, experience and visual asset requirements at the category level. For each key field, establish a field-level data snapshot containing the value, source, source level, acquisition time, market and version, and status label. The status label is taken from one of the following: verified, marketing statement only, incomplete conditions, cannot be directly compared, to be tested, and prohibited from use. In the selected execution mode, network retrieval is prohibited when local data is sufficient; when local data is partially missing, external data collection is performed only for the missing fields; and external data collection is performed according to the source priority when local data is completely missing. Objects and fields that have not passed official source verification are prohibited from entering object-level benchmarking, comparison conclusions, recommendations, or external content.

4. The method according to claim 1, characterized in that, The read-only downstream agent inherits fields based on their permission status and performs downstream permission and blocking control, including: The overall control state machine maintains a permission state for each field in the field-level data snapshot. The permission state is taken from one of direct use, conditional use, and prohibited use. The field marked as conditional use also carries its restriction description and source. When reading the unified analysis context, each read-only downstream agent inherits only fields with a permission status of direct use or conditional use. For inherited conditional use fields, the restriction description is inherited synchronously. Fields with a permission status of prohibited use are rejected for inheritance and marked as not covered in the output. When the conclusion of a certain preceding agent is not written back to the unified analysis context, the read-only downstream agent that depends on the conclusion will mark the corresponding field as not covered and prompt the corresponding preceding agent to be executed first, prohibiting manual rewriting. The central control state machine performs permission or blocking actions on downstream agents based on the field permission status and blocking items. When a blocking item is marked as needing to be repaired, the blocking is performed on subsequent agents that depend on that field until the blocking item is removed.

5. The method according to claim 1, characterized in that, The calculation of a multidimensional weighted opportunity score for candidate new product directions includes: Read the candidate new product directions output by the preceding agent from the unified analysis context, as well as the quantitative indicators required to calculate the multidimensional weighted opportunity score; For each candidate new product direction, a relevance score, engagement score, coverage score, and innovation score are calculated. The relevance score is determined based on the conclusions of the consumer profile and scenario insight analysis agency and the consumer voice insight analysis agency. The engagement score is determined based on the conclusions of the social media content trend analysis agency. The coverage score is determined based on the conclusions of the omnichannel market performance analysis agency. The innovation score is determined based on the conclusions of the competitor landscape and white space analysis agency. Preset weights are assigned to the relevance score, participation score, coverage score and innovation score respectively. The scores are weighted and summed with their corresponding weights to obtain the multidimensional weighted opportunity score for the candidate new product direction. A comprehensive rating is assigned to the candidate new product direction based on the comparison between the multidimensional weighted opportunity score and the preset rating range. Using coverage score as the first coordinate axis, innovation score as the second coordinate axis, the sum of relevance score and participation score as the marker size, and comprehensive rating as the marker color, a new product opportunity distribution map is generated, and the priority of candidate new product directions is output from high to low according to the multi-dimensional weighted opportunity score.

6. The method according to claim 1, characterized in that, The process of calculating a concept readiness score based on multiple diagnostic sub-items and performing downstream permission and blocking control based on the comparison result of the concept readiness score and a preset threshold includes: Five diagnostic sub-items are set up: claim clarity, evidence completeness, competitor differentiation, experience fulfillment, and visual presentation readiness, and preset weights are assigned to each diagnostic sub-item. For the current object to be analyzed, the scores of each diagnostic item are evaluated one by one. The scores of each diagnostic item are weighted and summed with their corresponding weights to obtain a concept readiness score ranging from zero to one hundred. The concept readiness score is calculated independently and does not change the multidimensional weighted chance score. Set a first threshold and a second threshold below the first threshold. When the concept readiness score is not less than the first threshold, allow subsequent concept board generation agents and consumer verification agents to proceed. When the concept readiness score is between the second threshold and the first threshold, require the completion of critical repair tasks before proceeding. When the concept readiness score is less than the second threshold, block the final concept board generation agent and return to concept optimization. When the evidence chain diagnostic agent in the bridging phase is skipped or not executed, the concept readiness score is marked as unevaluated, and estimation or backfilling is prohibited.

7. The method according to claim 1, characterized in that, The downstream stage also includes deploying demand space mapping agent execution index clustering, including: The demand space mapping agent reads the conclusions written back by the consumer original voice insight analysis agent, the competitor landscape and white space analysis agent, and the social media content trend analysis agent. By conducting structured quantitative question and answer with a preset number of simulated consumers, it collects quantitative data including purchase trigger scenarios, purchase obstacles, demographic information, product accessibility, and core attitudes. For each feature, the index value of the feature is calculated by multiplying the ratio of the feature’s proportion in the target segment population to the feature’s proportion in the total sample by the base. When the index value is higher than a preset upper threshold, it is marked as a highly significant feature; when the index value is lower than a preset lower threshold, it is marked as a low significant feature. Using the target audience as the first clustering axis, usage scenarios and trigger points as the second clustering axis, and core demands and obstacles as the third clustering axis, the collected quantitative data is cross-clustered to generate multiple independent demand spaces, and the sub-groups identified in each demand space are given names. For each demand space, the purchase trigger scenarios and purchase obstacles are sorted from high to low according to their corresponding index values. High-index trigger scenarios and high-index obstacles that are not yet covered by existing comparison objects are marked as white spaces. For each demand space, a segmented population profile is output, including market consumption scale, relevance and growth momentum, demographics, accessibility, core attitudes, trigger points, obstacles, and competitor positioning.

8. The method according to claim 1, characterized in that, The deployment includes: Configure corresponding skill description files for the parallel analysis agent, evidence chain diagnosis agent and read-only downstream agent respectively. Each skill description file records the corresponding agent's role setting, data governance rules, execution steps, and structured field template for writing back the conclusion, as well as routing rules containing the identifier of the preceding dependent agent and the triggering conditions. The skill scheduling engine parses user input to determine the set of proxies to be executed, and constructs a dependency graph between proxies based on the pre-dependent proxies identifiers recorded in each skill description file. The execution order and parallel grouping of each proxy are determined according to the dependency graph. When scheduling the execution of a certain agent, the skill description file corresponding to the agent and the existing structured conclusions in the unified analysis context are provided to the large language model. The large language model then performs the corresponding analysis or orchestration duties based on the skill description file and writes the output back to the unified analysis context in structured text form according to the structured field template. The skill scheduling engine controls whether to schedule the skill description file corresponding to the downstream agent that depends on the corresponding field based on the field permission status and blocking items recorded in the master control state machine, thereby implementing the downstream permission and blocking control in a programmatic manner.

9. A new product insight analysis system based on multi-intelligent agent collaborative orchestration, used to implement the method as described in any one of claims 1 to 8, characterized in that, include: The parameter initialization unit is used to parse user input and extract analysis parameters, write the analysis parameters into a unified analysis context, and initialize the overall control state machine. A parallel analysis unit is used to deploy multiple independent parallel analysis agents in the first phase. Each of the parallel analysis agents performs analysis in parallel without calling each other and writes the structured conclusions back to the unified analysis context. The evidence chain diagnostic unit is used to deploy the evidence chain diagnostic agent, automatically select the execution mode based on the source verification coverage, create field-level data snapshots with status labels, and calculate the concept readiness score. Downstream orchestration unit is used to deploy multiple read-only downstream agents, read only the existing conclusions in the unified analysis context in the order of dependency and inherit fields according to the permission status, calculate multidimensional weighted chance scores and execute downstream permission and blocking control based on concept readiness scores; The status monitoring unit is used to maintain the overall control state machine and triggers the rescheduling of the corresponding agent when the status of key fields in the unified analysis context changes or when a blocking item appears.

10. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 8.

11. A computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 8.

Images