Information processing system

JP7912365B1Active Publication Date: 2026-08-28METELIX CO LTD
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
JP2026022227
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2026-02-13
Publication Date
2026-08-28
Estimated Expiration
2046-02-13

AI Technical Summary

Benefits of technology

【0019】 本発明によれば、上位エージェントが、指示情報の特徴に応じて補助情報を動的に生成し、下位エージェントに送信するため、下位エージェントに付与すべきコンテキスト情報を過不足なく与えやすくなり、下位エージェントのタスク処理の精度及び効率を向上させることができる。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007912365000001_ABST
    Figure 0007912365000001_ABST
Patent Text Reader

Abstract

To provide a system that allows higher-level agents to appropriately delegate tasks to lower-level agents. [Solution] The upper-level agent includes a memory information storage unit and an external information acquisition unit via an external connection infrastructure, and an instruction information generation unit generates instruction information for lower-level agents. The auxiliary information generation unit dynamically generates auxiliary information from memory information or external information based on the characteristics of the instruction information, and excludes information above a predetermined confidentiality level based on confidentiality level attributes and authorization information. Furthermore, it selects a subset of information based on the determination of the judgment criterion category, determines whether or not to include action guideline information, and injects it into the auxiliary information. The transmission unit transmits the instruction information and auxiliary information so that they are included in the input context of the lower-level agent, and the upper-level agent verifies the execution result against the action guideline information, and regenerates and retransmits the instruction information or auxiliary information as necessary.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to an information processing system in which an upper agent generates instruction information for a lower agent, and dynamically generates and transmits auxiliary information based on characteristics of the instruction information.

Background Art

[0002] In recent years, information processing technology that efficiently processes complex tasks has been used, in which a plurality of agents using language models or the like are combined, and an upper agent decomposes tasks and allocates them to lower agents.

[0003] In this type of technology, when an upper agent gives an instruction to a lower agent, the processing accuracy of the lower agent is improved by also providing context information such as background information necessary for task execution, preconditions, and materials to be referred to.

[0004] Furthermore, technology for connecting to external tools or external data sources, acquiring external information, and using it for tasks has also become widespread, enabling upper agents to combine external information and internal stored information to delegate work to lower agents.

[0005] On the other hand, if an upper agent transmits excessive information to a lower agent, useful information may be buried or information processing costs may increase due to the input constraints of the lower agent. In addition, if the transmitted information includes confidential information, the risk of confidential information leakage increases via a lower agent with low authority or a lower agent operating in an external environment.

Prior Art Documents

Patent Documents

[0006]

Patent Document 1

Patent Document 2

[0007] Patent Document 1 describes a task support system including multiple AI agents in which a supervisor agent sends instructions to sub-agents and supports tasks using connectors for communication with external systems and various data sources.

[0008] Furthermore, Patent Document 2 describes a technique in which multiple hierarchical agents collect information, generate selected information or summary information using the learning results, and transmit it to a higher level.

[0009] However, in situations where a higher-level agent issues instructions to a lower-level agent, there is no concrete mechanism in place to dynamically control "what information, to what extent, and at what level of granularity" should be provided to the lower-level agent in response to the instructions, while simultaneously preventing the leakage of confidential information and obtaining results that are consistent with the organization's decision-making criteria and guidelines.

[0010] In particular, as the channels for acquiring external information diversify and integration with external tools becomes easier, the information handled by higher-level agents becomes more diverse, and the level of confidentiality tends to be mixed. As a result, if the control over the information sent to lower-level agents is insufficient, confidential information leaks, inconsistencies in judgment criteria, outputs that violate organizational guidelines, or inconsistencies in quality are likely to occur.

[0011] The present invention aims to provide an information processing system that dynamically generates information to be transmitted from a higher-level agent to a lower-level agent according to the characteristics of the instruction information, appropriately injects it into the input context of the lower-level agent to support task processing while ensuring confidentiality control and consistency of judgment criteria, and further verifies the execution results and makes them regenerative. [Means for solving the problem]

[0012] An information processing system according to one aspect of the present invention comprises a higher-level agent and at least one lower-level agent, wherein the higher-level agent includes a storage information storage unit for storing stored information, an external information acquisition unit capable of acquiring external information via an external connection infrastructure, an instruction information generation unit for generating instruction information for the lower-level agent, an auxiliary information generation unit for dynamically generating auxiliary information based on the characteristics of the instruction information from at least one of the storage information storage unit or the external information acquisition unit, and a transmission unit for transmitting the generated instruction information and the generated auxiliary information to the lower-level agent.

[0013] In one embodiment of the present invention, the auxiliary information generation unit includes a filtering unit that excludes information of a predetermined confidentiality level or higher from the information to be included in the auxiliary information to be transmitted to a subordinate agent, based on confidentiality level attributes assigned to stored information or external information, and the predetermined confidentiality level is determined based on authorization information associated with the subordinate agent.

[0014] In one embodiment of the present invention, the higher-level agent includes a task classification unit that analyzes the characteristics of instruction information and determines the category of judgment criteria required by the instruction information, and an auxiliary information generation unit selects a subset of information from a group of information sources based on the determination result and generates auxiliary information.

[0015] In one embodiment of the present invention, the memory information storage unit includes a behavioral guideline storage area for storing organizational behavioral guideline information, and the auxiliary information generation unit determines whether or not to include the behavioral guideline information in the auxiliary information according to the nature of the instruction information or the determination result of the task classification unit.

[0016] In one embodiment of the present invention, the upper agent comprises a result receiving unit that receives execution results from the lower agent, and a verification unit that verifies the execution results in light of action guideline information included in the auxiliary information, the instruction information generation unit or auxiliary information generation unit regenerates instruction information or auxiliary information based on the verification results, and the transmission unit transmits the regenerated instruction information or auxiliary information to the lower agent.

[0017] In one embodiment of the present invention, the upper agent comprises a result receiving unit that receives execution results from the lower agent, and a verification unit that verifies the execution results in light of action guideline information included in the auxiliary information, the instruction information generation unit or auxiliary information generation unit regenerates instruction information or auxiliary information based on the verification results, and the transmission unit transmits the regenerated instruction information or auxiliary information to the lower agent.

[0018] Furthermore, in one embodiment of the present invention, the transmission unit transmits the generated auxiliary information so that it is included in the input context when a subordinate agent performs information processing based on the instruction information. [Effects of the Invention]

[0019] According to the present invention, the higher-level agent dynamically generates auxiliary information according to the characteristics of the instruction information and transmits it to the lower-level agent. This makes it easier to provide the lower-level agent with the correct amount of contextual information, thereby improving the accuracy and efficiency of the lower-level agent's task processing.

[0020] Furthermore, according to one aspect of the present invention, the auxiliary information generation unit excludes information above a predetermined confidentiality level from the auxiliary information based on confidentiality level attributes and authorization information, thereby reducing the risk of confidential information leakage associated with the transmission of information to lower-level agents.

[0021] Furthermore, according to one aspect of the present invention, the task classification unit analyzes the characteristics of the instruction information to determine the category of the judgment criteria, and the source and scope of the acquisition of auxiliary information are determined based on the determination result. This makes it easier to provide judgment criteria and reference information that match the nature of the instruction information to the subordinate agent, thereby improving the consistency of the judgment.

[0022] Furthermore, according to one aspect of the present invention, since it is determined whether or not to include behavioral guideline information in the auxiliary information and then injects it, it becomes easier to align the output of the subordinate agent with the organization's behavioral guidelines, thereby suppressing deviations from policies in external relations, interpersonal relations, or management decisions.

[0023] Further, according to one aspect of the present invention, a loop can be formed in which the execution result is verified against action guideline information, and instruction information or auxiliary information is regenerated and retransmitted as necessary, making it easy to guarantee the quality of the results of lower-level agents.

[0024] Furthermore, according to one aspect of the present invention, since the transmitting unit transmits the auxiliary information so that it is included in the input context of the lower-level agent, the auxiliary information can be reliably and easily reflected in the processing of the lower-level agent. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] [Figure 1] It is a configuration diagram showing a configuration example of the information processing system of the present invention. [Figure 2] It is a functional block diagram showing a functional configuration example of an upper-level agent. [Figure 3] It is a flowchart showing an example of a processing procedure in which an upper-level agent dynamically generates auxiliary information and transmits it to a lower-level agent. [Figure 4] It is a sequence diagram showing an example of message exchange among an upper-level agent, an external connection infrastructure, and a lower-level agent, and an example of a result verification loop. [Figure 5] It is an explanatory diagram showing a structural example of attribute information imparted to stored information and external information, and a conceptual example of filtering based on a confidentiality level attribute and authority information. MODE FOR CARRYING OUT THE INVENTION

[0026] <Details of Embodiment> Hereinafter, embodiments of the present invention will be described with reference to the drawings. The following description is an example for facilitating understanding of the present invention, and the present invention is not limited to the following embodiments.

[0027] In this specification, "agent" refers to a concept including a software module that receives input information, performs information processing such as inference, planning, and generation, and outputs to other agents or external systems. Agents may be implemented using language models, rule-based inference, machine learning models, or a combination thereof.

[0028] In this specification, "upper agent" refers to an agent that has the function of generating instruction information for at least one lower agent, generating auxiliary information in accordance with said instruction information, and transmitting these to the lower agent.

[0029] In this specification, a "sub-agent" is an agent that performs task processing based on instruction information and auxiliary information transmitted from a higher-level agent and can return the execution results to the higher-level agent. There may be one or more sub-agents.

[0030] In this specification, "characteristics of the instruction information" refers to a set of attributes that can be extracted or estimated from the instruction information, including the content, structure, length, scope, risk, necessary judgment criteria, stakeholders, deadline, or confidentiality of the instruction information, and may be expressed as feature quantities.

[0031] As shown in Figure 1, the information processing system 100 comprises a higher-level agent 110 and at least one lower-level agent 120. The information processing system 100 can be connected to an external connection infrastructure 200 via a network 300.

[0032] The upper agent 110 includes a memory information storage unit 130 for storing memory information, an external information acquisition unit 140 capable of acquiring external information via an external connection base 200, an instruction information generation unit 150 for generating instruction information for the lower agent 120, an auxiliary information generation unit 160, and a transmission unit 170.

[0033] The external connectivity infrastructure 200 is a platform that provides connections to external tools or external data sources 210, and may consist of, for example, external APIs, connectors, or external tool calling protocols. The external connectivity infrastructure 200 may provide connection ports to multiple external tools or external data sources 210.

[0034] The memory information storage unit 130 stores the memory information held by the higher-level agent 110. The memory information may include past task history, project information, stakeholder information, organizational rules, domain knowledge, or behavioral guideline information, as described later.

[0035] The external information acquisition unit 140 acquires external information via the external connection infrastructure 200. The external information may include responses obtained from external tools, search results from external data sources, or logs from external systems.

[0036] The instruction information generation unit 150 generates instruction information that specifies the content of the task that the upper agent 110 wants the lower agent 120 to perform. The instruction information may be in the form of natural language, structured data, or a combination thereof.

[0037] The auxiliary information generation unit 160 dynamically generates auxiliary information based on the characteristics of the instruction information from at least one of the memory information storage unit 130 or the external information acquisition unit 140. The auxiliary information may include preconditions, judgment criteria, related materials, or constraints that the lower agent 120 should refer to in order to process the instruction information.

[0038] The transmission unit 170 transmits the generated instruction information and the generated auxiliary information to the subordinate agent 120. Transmission can be achieved in the form of message transmission, API call, or prompt supply.

[0039] The lower agent 120 takes instruction information and auxiliary information as input to perform task processing and can send the execution results to the upper agent 110. The execution results may be text, structured data, results of operations on external tools, or a combination thereof.

[0040] Each function of the information processing system 100 may be implemented on a single computer or distributed across multiple computers. The upper agent 110 and the lower agent 120 may operate in the same environment or in different environments.

[0041] As shown in Figure 2, the auxiliary information generation unit 160 can determine the information source to be referenced, the reference range, and the representation format of the auxiliary information based on the characteristics of the instruction information. The auxiliary information generation unit 160 may also include functions such as a filtering unit 161, a task classification unit 162, an action guideline content determination unit 163, and a compression unit 240, as needed.

[0042] The filtering unit 161 excludes information above a predetermined confidentiality level from candidate information to be included in the auxiliary information. The task classification unit 162 analyzes the characteristics of the instruction information to determine the category of the judgment criteria. The action guideline inclusion determination unit 163 determines whether or not to include the action guideline information in the auxiliary information.

[0043] The upper agent 110 may include a result receiving unit 180 and a verification unit 190. The result receiving unit 180 receives the execution results from the lower agent 120, and the verification unit 190 verifies the execution results in light of action guideline information, etc.

[0044] The upper agent 110 may include a history storage unit 230. The history storage unit 230 stores previously generated auxiliary information and execution results in association, and the auxiliary information generation unit 160 can update the generation conditions based on the history.

[0045] The upper agent 110 may include a capability determination unit 250 or a model selection unit 260. The capability determination unit 250 determines the capabilities of the lower agent 120, and the auxiliary information generation unit 160 can adjust the level of detail of the auxiliary information according to the determination result. The model selection unit 260 can select a language model to be used by the lower agent 120 based on the characteristics of the instruction information.

[0046] The upper agent 110 may include an execution status receiving unit 270, an additional auxiliary information generation unit 271, and an additional auxiliary information transmission unit 272. This allows it to generate and transmit additional auxiliary information while the lower agent 120 is running.

[0047] Next, referring to Figure 3, an example of a processing procedure in which the upper agent 110 dynamically generates auxiliary information and sends it to the lower agent 120 will be described.

[0048] The higher-level agent 110 generates instruction information for the lower-level agent 120 in response to the occurrence or delegation request of a task (step S101).

[0049] Next, the higher-level agent 110 extracts or estimates the characteristics of the generated instruction information (step S102). The characteristics of the instruction information may include the task's purpose, target, deadline, necessary decision criteria, stakeholders, confidentiality, etc.

[0050] Next, the higher-level agent 110 determines the category of the judgment criteria using the task classification unit 162 as needed (step S103).

[0051] Next, the auxiliary information generation unit 160 uses at least one of the memory information storage unit 130 and the external information acquisition unit 140 as information sources to collect candidate information to be used for generating auxiliary information based on the characteristics of the instruction information or the determined category (step S104).

[0052] Next, the auxiliary information generation unit 160, if necessary, uses the filtering unit 161 to exclude information above a predetermined confidentiality level from the candidate information (step S105).

[0053] Next, the auxiliary information generation unit 160, using the action guideline inclusion determination unit 163, determines whether or not to include the action guideline information in the auxiliary information, and if so, adds at least a portion of the action guideline information to the candidate information (step S106).

[0054] Next, the auxiliary information generation unit 160 selects a subset of information from the candidate information and generates auxiliary information while summarizing or compressing it as necessary (step S107).

[0055] Next, the transmission unit 170 transmits the instruction information and auxiliary information to the subordinate agent 120 (step S108). The transmission unit 170 may transmit the auxiliary information so that it is included in the input context of the subordinate agent 120.

[0056] The lower agent 120 executes task processing based on the instruction information and auxiliary information, and returns the execution result to the upper agent 110 (step S109).

[0057] The higher-level agent 110 receives the execution results via the result receiving unit 180 and verifies the execution results via the verification unit 190 (step S110). Depending on the verification results, the higher-level agent 110 may regenerate and retransmit instruction information or auxiliary information (step S111).

[0058] <Confidentiality level filtering> The following describes confidentiality level filtering. Confidentiality level filtering is a function that controls the information included in the auxiliary information that the upper agent 110 sends to the lower agent 120, thereby preventing the leakage of confidential information.

[0059] As shown in Figure 5, the stored information in the stored information storage unit 130 may be managed as a stored information record that includes the information itself and attribute information. The attribute information may include confidentiality level attribute, creator, creation date and time, target project, target customer, information type, or usage conditions.

[0060] Similarly, external information acquired by the external information acquisition unit 140 may be managed as an external information record that includes the information itself and attribute information. The attribute information of the external information may include the source of acquisition, acquisition date and time, usage permission information, or confidentiality level attribute.

[0061] Confidentiality level attributes can be expressed, for example, as a multi-level value. Confidentiality level attributes can be set based on factors such as the impact of information leakage, the scope of information disclosure, or contractual handling. Confidentiality level attributes can be assigned by administrators, automatically assigned during information generation, or inherited from metadata associated with the information source.

[0062] The subordinate agent 120 may be associated with authorization information. This authorization information may indicate the upper limit of the confidentiality level that the subordinate agent 120 can access or handle, the scope of the target project, or the scope of the target customers. The authorization information may be set according to the role, responsibilities, or operating environment of the subordinate agent 120.

[0063] The filtering unit 161 determines a predetermined confidentiality level based on the authorization information. For example, the filtering unit 161 may adopt the upper limit of the confidentiality level included in the authorization information as the predetermined confidentiality level. Alternatively, the predetermined confidentiality level may be dynamically adjusted by taking into account the estimated confidentiality value included in the characteristics of the instruction information, or the determination result by the task classification unit 162.

[0064] The filtering unit 161 compares the confidentiality level attribute assigned to the candidate information with a predetermined confidentiality level, and excludes candidate information whose confidentiality level attribute is equal to or greater than the predetermined confidentiality level from the auxiliary information. Exclusion includes removing the candidate information from the input for generating the auxiliary information, or not including it in the auxiliary information.

[0065] The exclusion by the filtering unit 161 is not limited to the exclusion of the entire information, but may also be limited to the exclusion of a portion of the information. For example, if candidate information is divided into multiple segments and each segment is assigned a confidentiality level attribute, the filtering unit 161 may exclude only segments that are of a predetermined confidentiality level or higher.

[0066] The filtering unit 161 may perform masking or substitution as an alternative to exclusion. For example, proper nouns, amounts, shareholder structure, or contract terms may be masked and replaced with anonymized expressions that retain their meaning. The degree of masking may be adjusted according to the difference between a predetermined confidentiality level and the confidentiality level attribute of the candidate information, or according to the type of authorization information.

[0067] The filtering unit 161 can also exclude external information by referring to its confidentiality level attribute. If the external information does not have a confidentiality level attribute, a default value may be set according to the source, or the higher-level agent 110 may analyze the content of the external information to estimate its confidentiality and assign an attribute.

[0068] The filtering unit 161 may change the exclusion conditions depending on the type of information. For example, candidate information containing personal information may be excluded or strongly masked regardless of the value of the confidentiality level attribute.

[0069] The filtering unit 161 may tighten the handling of confidential information according to the determination result of the task classification unit 162. For example, if a category involving interpersonal judgment, including external transmission, is determined, the filtering unit 161 may set a lower confidentiality level and exclude a wider range of information.

[0070] The filtering unit 161 may record the identifier of the excluded candidate information, the reason for exclusion, the predetermined confidentiality level applied, and the access information referenced as a log, in order to make the filtering results auditable. The log may be stored in the history storage unit 230 or in a separate audit log area.

[0071] The above level of confidentiality filtering prevents information held by the upper agent 110, whether stored or external, from being included in auxiliary information that exceeds the authority of the lower agent 120, thereby reducing the risk of leakage in actual operation.

[0072] <Category Determination Criteria> The task classification unit 162 will now be described. The task classification unit 162 analyzes the characteristics of the instruction information and determines the category of the judgment criteria required by the instruction information.

[0073] The categories of judgment criteria include, for example, at least one of business judgments, technical judgments, and interpersonal judgments. Business judgments may include judgments on business strategy, finance, contracts, risks, or prioritization. Technical judgments may include judgments on design, implementation, evaluation, or technical feasibility. Interpersonal judgments may include judgments on external communication, relationship management, or the appropriateness of expression.

[0074] The task classification unit 162 may generate feature quantities from keywords, objects, stakeholders, deadlines, or output formats included in the instruction information. These feature quantities can be expressed as natural language features, structured features, or embedded vectors.

[0075] The task classification unit 162 may be implemented using a rule-based classifier, a machine learning model, or a combination thereof. For example, it may use rules that determine categories according to a specific vocabulary or template, or it may use a classification model that takes indicative information as input and outputs categories.

[0076] The task classification unit 162 may output a single category or multiple categories simultaneously. If multiple categories are output, priority or confidence levels may be assigned to them.

[0077] The auxiliary information generation unit 160 determines the source and scope of acquisition of auxiliary information from the information source group based on the determination result of the task classification unit 162. The information source group may include stored information stored in the storage information storage unit 130 and external information that can be acquired by the external information acquisition unit 140.

[0078] For example, when a business decision is made, the auxiliary information generation unit 160 may prioritize acquiring stored information records related to contracts, regulations, financial indicators, risk assessments, etc., and the external information acquisition unit 140 may acquire related information from an external data source.

[0079] For example, if a technical judgment is made, the auxiliary information generation unit 160 may prioritize acquiring stored information records related to design documents, specifications, implementation history, or technical knowledge, and the external information acquisition unit 140 may acquire related information from external technical documents or tools.

[0080] For example, if an interpersonal judgment is determined, the auxiliary information generation unit 160 may prioritize acquiring memory information records related to communication policies, expression rules, key points of past interactions, or behavioral guideline information, and may include behavioral guideline information in the auxiliary information as needed.

[0081] The auxiliary information generation unit 160 determines the acquisition destination and acquisition scope, and then selects a subset of information from the group of information sources. Subset selection may include ranking based on relevance, extraction based on filter conditions, or selection to ensure diversity.

[0082] The auxiliary information generation unit 160 may use the characteristics of the instruction information as additional conditions in subset selection. For example, it may generate short auxiliary information when the deadline is short, and adjust to include more supporting information when the risk is high.

[0083] The auxiliary information generation unit 160 generates auxiliary information based on the selected information after a subset selection. The auxiliary information may be a collection of excerpts from the selected information, a summary text integrating the selected information, or in the form of a checklist or judgment criteria.

[0084] Based on the above, by determining the necessary judgment criterion categories from the characteristics of the instruction information, and controlling the source, scope, and subset selection of information sources according to the determination result, it becomes easier to adapt the auxiliary information provided to the subordinate agent 120 to the task.

[0085] <Injection of behavioral guidelines> The handling of behavioral guideline information will be explained. The memory information storage unit 130 may include a behavioral guideline storage area 131 for storing the organization's behavioral guideline information.

[0086] The code of conduct information may include at least one of the mission, vision, values, and code of conduct. The code of conduct information may be stored as text, clauses of code of conduct, or structured data.

[0087] The behavioral guideline inclusion determination unit 163 determines whether or not to include behavioral guideline information as auxiliary information, depending on the nature of the instruction information or the determination result by the task classification unit 162. For example, in tasks where the appropriateness of external document creation, interpersonal coordination, or decision-making is important, a determination to include behavioral guideline information may be made.

[0088] The behavioral guideline inclusion determination unit 163 may decide which range of behavioral guideline information to include. For example, instead of including the entire code of conduct, it may include only the parts relevant to the task in question.

[0089] The behavioral guideline content determination unit 163 may estimate the purpose of referencing the behavioral guideline information based on the characteristics of the instruction information. The purpose of referencing may include adjusting the tone of expression, clarifying ethical constraints, clarifying customer value priority, or clarifying risk mitigation policies.

[0090] When the auxiliary information generation unit 160 includes behavioral guideline information as auxiliary information, it may include the behavioral guideline information in a format that explicitly states it as a "judgment criterion." For example, it may include it as a constraint such as "the output must satisfy value A."

[0091] The auxiliary information generation unit 160 may include the action guideline information as an "evaluation axis" when including it in the auxiliary information. For example, it may request a comparative output of multiple options from the subordinate agent 120 and instruct it to self-evaluate each option in light of the action guideline information.

[0092] When the auxiliary information generation unit 160 includes behavioral guideline information as auxiliary information, it may include the behavioral guideline information in a summarized or compressed form. Summarization can be achieved by shortening only the relevant normative sentences, depending on the characteristics of the instructional information.

[0093] The behavioral guideline information, like other stored information in the stored information storage unit 130, may be assigned a confidentiality level attribute. If part of the behavioral guideline information is confidential, the filtering unit 161 may mask or exclude it.

[0094] The decision of whether or not to include behavioral guideline information as supplementary information may take into account the authorization information or operating environment of the subordinate agent 120. For example, a subordinate agent 120 operating in an external environment may only include publicly available behavioral guideline information.

[0095] When the auxiliary information generation unit 160 includes action guideline information, it may also indicate which of the judgment criterion categories determined by the task classification unit 162 the action guideline information relates to. This makes it easier for the subordinate agent 120 to understand which parts of the auxiliary information are important for decision-making.

[0096] By injecting these behavioral guidelines, the output of the subordinate agent 120 becomes more easily aligned with organizational policies, improving consistency in external relations and decision-making tasks.

[0097] <Result Verification Loop (Quality Assurance)> The result verification loop is described below. The upper agent 110 may include a result receiving unit 180 that receives execution results from the lower agent 120, and a verification unit 190 that verifies the received execution results in light of the action guideline information.

[0098] The result receiving unit 180 may receive the execution result when the lower agent 120 completes the task processing, or it may receive intermediate results in stages.

[0099] The verification unit 190 verifies the execution results against the action guideline information included in the auxiliary information. Verification against the information includes a process to determine whether the execution results are consistent with the constraints or evaluation axes included in the action guideline information.

[0100] The verification unit 190 may calculate a score indicating the degree of suitability to the behavioral guideline information. The score can be calculated using rule matching, keyword matching, semantic similarity, or a classification model.

[0101] The verification unit 190 may verify not only the conformity to the code of conduct information but also whether or not there are any deviations from confidentiality. For example, if the execution result contains confidential information, it may mask a portion of the output or regenerate it.

[0102] The verification unit 190 may verify the suitability to the judgment criterion category determined by the task classification unit 162. For example, if the output is biased towards technical details despite being a management judgment category, the verification unit 190 may estimate a lack of supplementary information and regenerate the output.

[0103] Based on the verification results, the instruction information generation unit 150 or the auxiliary information generation unit 160 regenerates the instruction information or auxiliary information. Regeneration includes changing the expression of the instruction, adding constraints, adding auxiliary information, adjusting the granularity of the auxiliary information, or reviewing the information source.

[0104] The transmission unit 170 transmits the regenerated instruction information or auxiliary information to the subordinate agent 120. This allows the subordinate agent 120 to perform task processing again based on the updated input context.

[0105] The verification unit 190 similarly verifies the execution results after regeneration and may form multiple loops as necessary. The upper limit of the number of loops or the passing conditions may be set according to the importance or deadline of the task.

[0106] The verification unit 190 may output a trigger if the passing conditions are not met, which would cause the higher-level agent 110 to switch the task to another lower-level agent 120, or to request a human review.

[0107] The verification unit 190 may generate information such as sections deemed non-compliant, applicable action guidelines, and recommended modifications, in order to explain the reasons for the verification results. These can be used to regenerate auxiliary information or to store it in the history storage unit 230.

[0108] The above result verification loop makes it easier to align the output of the lower-level agent 120 with the action guideline information while suppressing variations in quality.

[0109] <Injection of auxiliary information (input context creation)> The injection of auxiliary information by the transmission unit 170 will be described below. When the transmission unit 170 transmits instruction information and auxiliary information to the subordinate agent 120, it may transmit the auxiliary information so that it is included in the input context of the subordinate agent 120.

[0110] Sending information so that it is included in the input context includes sending it so that auxiliary information is stored or attached as part of the input data that the lower agent 120 refers to when performing task processing.

[0111] For example, the transmitting unit 170 may format the instruction information and auxiliary information as a single message and transmit it by incorporating it into the prompt or input message sequence of the subordinate agent 120.

[0112] Alternatively, the transmission unit 170 may transmit instruction information and auxiliary information in separate fields, and the lower-level agent 120 may combine them and treat them as an input context.

[0113] Alternatively, the transmission unit 170 may transmit auxiliary information as data to be registered in the short-term memory area or temporary reference store of the lower agent 120 so that the lower agent 120 can refer to it during processing.

[0114] As a result, auxiliary information is reliably incorporated into the input context of the lower-level agent 120, making it easier to perform task processing using the auxiliary information.

[0115] The auxiliary information generation unit 160 includes a compression unit 240, and may summarize or compress the auxiliary information according to the processing capacity of the lower agent 120. The processing capacity can be determined based on the context window size, the maximum number of input tokens, the amount of memory, or the upper limit of processing time.

[0116] The upper agent 110 is equipped with a capability determination unit 250 and may determine the capability based on the skill information or processing history of the lower agent 120. The auxiliary information generation unit 160 may adjust the level of detail or compression strength of the auxiliary information according to the determined capability.

[0117] The upper agent 110 may include a history storage unit 230 and store previously generated auxiliary information and execution results in association. The auxiliary information generation unit 160 may update the selection criteria for the acquisition destination, acquisition scope, summarization criteria, or exclusion criteria based on the stored history.

[0118] The external information acquisition unit 140 may be capable of acquiring external information from multiple external connection infrastructures 200. The auxiliary information generation unit 160 may generate auxiliary information by integrating the external information acquired from multiple sources.

[0119] The upper agent 110 may include a model selection unit 260 and select a language model to be used by the lower agent 120 based on the characteristics of the instruction information. The transmission unit 170 may transmit model information indicating the selected language model along with the instruction information.

[0120] The auxiliary information generation unit 160 may generate auxiliary information by inputting instruction information and information not excluded by the filtering unit 161 into the large-scale language model 400. In this case, the upper agent 110 may include a pre-processing unit 265 that masks a portion of the input information according to the confidentiality level attribute or a predetermined confidentiality level.

[0121] The upper agent 110 is equipped with an execution status receiving unit 270 and may receive the execution status while the lower agent 120 is running. The additional auxiliary information generation unit 271 may generate additional auxiliary information according to the execution status, and the additional auxiliary information transmission unit 272 may transmit the additional auxiliary information.

[0122] The external connection infrastructure 200 may be configured in accordance with connection specifications that define the connection destination to an external tool or external data source. This allows for standardization of the implementation of external collaboration while differentiating the control algorithm provided by the auxiliary information generation unit 160. <Examples>

[0123] For example, when a higher-level agent 110 generates instruction information to "instruct a lower-level agent 120 to create a draft message for external parties," the task classification unit 162 can determine whether it is a person-to-person interaction. In this case, the auxiliary information generation unit 160 can include key points from past communication history and parts of the action guideline information related to external expression as auxiliary information, and the transmission unit 170 can inject this auxiliary information as input context and transmit it.

[0124] In this case, if the stored information includes customer confidential information, the filtering unit 161 may determine a predetermined confidentiality level based on the access information and generate auxiliary information by excluding or masking information above the predetermined confidentiality level.

[0125] Furthermore, if the lower agent 120 is operating in an external environment, the upper agent 110 may configure the authorization information for the external environment and the filtering unit 161 may control the auxiliary information more conservatively.

[0126] Furthermore, if the draft text returned from the lower agent 120 is inconsistent with the action guideline information, the verification unit 190 can identify the inconsistent parts, the auxiliary information generation unit 160 can revise and regenerate the range of action guideline information injection, and the transmission unit 170 can retransmit it, forming a loop.

[0127] <Variation> In addition to exclusions based on confidentiality level attributes, the filtering unit 161 may configure exclusion conditions by combining multiple conditions such as the source of the information, contract terms, or the presence or absence of personal information.

[0128] The task classification unit 162 may be expanded to include categories such as legal judgments, security judgments, or operational judgments, in addition to management judgments, technical judgments, and interpersonal judgments.

[0129] The behavioral guideline information is not limited to text; it may also be stored as a scored evaluation axis, a set of prohibited actions, or a recommended template. The auxiliary information generation unit 160 can control the output format of the subordinate agent 120 by including these in the auxiliary information.

[0130] In addition to verifying against the action guideline information, the verification unit 190 may comprehensively evaluate the consistency of the output, the presentation of evidence, and whether or not there is a deviation from confidentiality, and update the generation conditions of the auxiliary information generation unit 160 based on the evaluation results.

[0131] The injection by the transmitting unit 170 is not limited to a method that integrates instruction information and auxiliary information, but may also be a method that stores auxiliary information in the internal memory or external reference store of the lower agent 120 and allows it to be referenced. [Explanation of Symbols]

[0132] 100 Information Processing Systems 110 Top Agents 120 Lower-ranking agents 130 Memory Information Storage Unit 131 Action Guidelines Storage Area 140 External Information Acquisition Department 150 Instruction information generation section 160 Auxiliary information generation section 161 Filtering section 162 Task Classification Section 163 Behavioral Guidelines Content Judgment Department 170 Transmitter 180 Result receiving unit 190 Verification Department 200 External connectivity infrastructure 210 External tools or external data sources 230 History Storage Unit 240 Compression section 250 Ability Judgment Department 260 Model Selection Section 265 Preprocessing Unit 270 Execution Status Receiving Unit 271 Additional auxiliary information generation section 272 Additional Auxiliary Information Transmission Unit 300 Networks 400 Large-Scale Language Models

Claims

1. An information processing system comprising a higher-level agent and at least one lower-level agent, The aforementioned higher-level agent, A memory information storage unit for storing memory information, An external information acquisition unit capable of acquiring external information via an external connection infrastructure, An instruction information generation unit that generates instruction information for the subordinate agent, An auxiliary information generation unit selects a subset of information based on the characteristics of the instruction information, using at least one of the stored information in the storage information storage unit and the external information obtainable by the external information acquisition unit as information sources, and dynamically generates auxiliary information that the subordinate agent should refer to in order to process the instruction information based on the selected information. The system includes a transmission unit that transmits the generated instruction information and the generated auxiliary information to the subordinate agent so that they are included in the input context when the subordinate agent performs information processing based on the instruction information. Information processing system.

2. The information processing system according to claim 1, The auxiliary information generation unit includes a filtering unit that excludes information of a predetermined confidentiality level or higher from the information to be included in the auxiliary information to be transmitted to the subordinate agent, based on confidentiality level attributes assigned to at least one of the stored information stored in the stored information storage unit and the external information acquired by the external information acquisition unit. The filtering unit determines the predetermined confidentiality level based on the authorization information associated with the subordinate agent. Information processing system.

3. The information processing system according to claim 1, The aforementioned higher-level agent further comprises a task classification unit that analyzes the characteristics of the instruction information and determines the category of judgment criteria required by the instruction information. The auxiliary information generation unit selects a subset of information from a group of information sources consisting of stored information stored in the stored information storage unit and external information obtainable by the external information acquisition unit, based on the determination result by the task classification unit, and generates the auxiliary information based on the selected information. The categories of the aforementioned judgment criteria include at least one of business judgments, technical judgments, and interpersonal judgments. Information processing system.

4. The information processing system according to claim 3, The aforementioned storage information unit includes a behavioral guideline storage area for storing organizational behavioral guideline information, The auxiliary information generation unit determines whether or not to include the action guideline information in the auxiliary information, depending on the nature of the instruction information or the determination result by the task classification unit. The aforementioned behavioral guidelines information includes at least one of the mission, vision, values, and code of conduct. Information processing system.

5. The information processing system according to claim 4, The aforementioned higher-level agent includes a result receiving unit that receives the execution result of the instruction information from the lower-level agent, The system further includes a verification unit that verifies the received execution result in light of the action guideline information included in the auxiliary information, The instruction information generation unit or the auxiliary information generation unit regenerates the instruction information or the auxiliary information based on the verification results by the verification unit, and the transmission unit transmits the regenerated instruction information or the auxiliary information to the subordinate agent. Information processing system.

6. The information processing system according to claim 1, The auxiliary information generation unit includes a compression unit that summarizes or compresses the auxiliary information according to the processing capacity of the lower-level agent. The processing capacity is determined based on at least one of the context window size, the maximum number of input tokens, the amount of memory, and the upper limit of processing time. Information processing system.

7. The information processing system according to claim 6, The aforementioned higher-level agent further comprises a capability determination unit that determines the capabilities of the lower-level agent based on the skill information or processing history of the lower-level agent, The auxiliary information generation unit adjusts the level of detail of the auxiliary information or the compression strength by the compression unit according to the determined capability. Information processing system.

8. The information processing system according to claim 1, The aforementioned higher-level agent further includes a history storage unit that stores the auxiliary information generated in the past in association with the execution results of the lower-level agent corresponding to said auxiliary information. The auxiliary information generation unit updates at least one of the selection criteria for the acquisition destination, the acquisition range, the summarization criteria, and the exclusion criteria for the filtering unit used to generate the auxiliary information, based on the history stored in the history storage unit. Information processing system.

9. The information processing system according to claim 1, The aforementioned external information acquisition unit is capable of acquiring external information from multiple external connection infrastructures. The auxiliary information generation unit generates the auxiliary information by integrating the external information obtained from multiple sources. Information processing system.

10. The information processing system according to claim 1, The higher-level agent further comprises a model selection unit that selects a language model to be used by the lower-level agent based on the characteristics of the instruction information. The transmitting unit transmits model information indicating the selected language model, along with the instruction information, to the subordinate agent. Information processing system.

11. The information processing system according to claim 2, The auxiliary information generation unit generates the auxiliary information by inputting the instruction information and information not excluded by the filtering unit into the large-scale language model. The higher-level agent further includes a preprocessing unit that, in input to the large-scale language model, masks at least a portion of the input information according to the confidentiality level attribute or a predetermined confidentiality level. Information processing system.

12. The information processing system according to claim 1, The aforementioned higher-level agent includes an execution status receiving unit that receives the execution status from the lower-level agent while the lower-level agent is executing its task, An additional auxiliary information generation unit dynamically generates additional auxiliary information according to the received execution status, The system further comprises an additional auxiliary information transmission unit that transmits the generated additional auxiliary information to the subordinate agent, Information processing system.

13. The information processing system according to claim 1, The aforementioned external connection infrastructure is configured in accordance with connection specifications that define the connection destination to an external tool or external data source. Information processing system.

Citation Information

Patent Citations

  • General knowledge extraction and spatio-temporal standardization expression method and related equipment

    CN119476296A

  • Data braiding system implementation method and device based on dialogue and multi-agent collaboration

    CN119537460A

  • Large-model-driven boiler anti-abrasion and anti-explosion supervision intelligent agent system and method

    CN121302120A

  • Geographic space analysis agent system and method based on large model

    CN121390817A

  • Intelligent question and answer method based on multi-agent collaboration and related device

    CN121457482A