Agent activity history management system

Abstract

This system provides an agent behavior history management system that collects each processing step performed by an AI agent as an action event, correlates these events, allows for the reconstruction of the processing path and decision-making process at a later date, and stores them as audit records in a format that allows for tamper detection. [Solution] In the agent behavior history management system, the history collection unit 20 collects the behavior history of the AI ​​agent 10, the history association unit 30 associates causal relationships, time-series relationships, and reference relationships based on an event classification vocabulary system (three levels: L1 major classification (processing phase), L2 medium classification (processing type), L3 minor classification (specific event)), the history reconstruction unit 40 reconstructs the history from an audit perspective and for single / multiple agents, and the audit record unit 50 generates a record that can be detected for tampering in hash chain format.

Description

Technical Field

[0001] The present invention relates to an agent behavior history management system, and particularly to a system that collects action events, associates them with each other, reconstructs a processing path or a judgment process later, and stores them as audit records in a form that can detect tampering, regarding the process in which a plurality of AI agents share and execute processing such as proposal generation, evaluation, integration, or decision support.

Background Art

[0002] In recent years, systems in which a plurality of AI agents share and execute processing such as proposal generation, evaluation, integration, and decision support have been widely used. In such systems, it is required to clarify the basis of proposals and judgment results generated by AI agents and to be able to track the processing path and judgment process later.

[0003] In conventional systems, it was common to record only the final output results generated by AI agents as logs. However, since intermediate processing, reference rule information, input data, context, etc. up to the final output are not recorded, it was difficult to reconstruct the processing path and judgment process later.

[0004] Also, in an environment where a plurality of AI agents are involved, it is necessary to associate the processing executed by each agent with each other and be able to track a consistent processing path for the same case. In particular, from the viewpoints of auditing and accountability, it is important to clarify which AI agent reached which output or judgment based on which reference rule information or input data.

[0005] In response to such problems, a system is required that collects each processing step executed by an AI agent as an action event, associates them with each other, and enables the reconstruction of the processing path and judgment process later.

Prior Art Documents

[0006] [Patent Document 1] Japanese Patent Application No. 2026-005527 discloses a corporate management system equipped with machine-readable articles of incorporation, an AI execution engine, audit nodes, and an information sharing platform. While it shows the overall architecture of corporate management, it does not disclose an action history management system that collects, correlates, reconstructs processing paths, and stores the action events of multiple AI agents in a tamper-detectable format.

[0007] [Patent Document 2] Japanese Patent Application No. 2026-007023 discloses data items such as AI_registry_id, task_id, timestamp, and evidence_ref, as well as task generation and evidence trail recording. While it provides the basics of AI agent management, it does not describe the collection of behavioral events of multiple AI agents, their interrelationship, the reconstruction of processing paths, or tamper-detectable audit records.

[0008] [Patent Document 3] U.S. Patent No. 11570264 (US11570264B1) "Provenance audit trails for microservices architectures" discloses techniques for tracking data provenance between microservices, but does not disclose the collection, correlation, restructuring of processing paths, or tamper-detectable audit records of behavioral events from multiple AI agents.

[0009] [Patent Document 4] U.S. Patent Application Publication No. 2026 / 0017525 (US20260017525A1) "Validating autonomous artificial intelligence (AI) agents using generative AI" Validates autonomous AI agents using generative AI. While it discloses the verification of agent behavior data, it does not describe the collection of behavioral events of multiple AI agents, their interrelationship, the reconstruction of processing paths, or the storage of tamper-detectable audit records.

[0010] [Patent Document 5] U.S. Patent No. 12244723 (US12244723B2) Codenotary "Cryptographically-verifiable immutable database" is a cryptographically verifiable immutable database technology. It discloses the efficient maintenance of immutable and verifiable records, but does not provide specific implementations for collecting, correlating, reconstructing processing paths for the behavior of multiple AI agents, or creating tamper-detectable audit records. [Non-patent literature]

[0011] [Non-Patent Document 1] R. Souza et al., "PROV-AGENT: Unified Provenance for Tracking AI Agent Interactions in Agentic Workflows", arXiv:2508.02866, 2025 https: / / arxiv.org / abs / 2508.02866 A unified provenance for tracking AI agent interactions in agentic workflows. The paper proposes a provenance model that extends W3C PROV, but does not show an implementation configuration for collecting behavioral events of multiple AI agents, correlating them, reconstructing processing paths, or creating tamper-detectable audit records.

[0012] [Non-Patent Document 2] MDPI Data, "A Dataset Capturing Decision Processes, Tool Interactions and Provenance Links in Autonomous AI Agents", 2026 https: / / www.mdpi.com / 2306-5729 / 11 / 4 / 66 A dataset that captures the decision-making processes and tool interactions of LLM agents. It discusses fine-grained tracking of agent inference and action execution, but does not cover the correlation of behavioral events across multiple AI agents, processing path reconstruction, or tamper-detectable audit records.

[0013] [Non-Patent Document 3] LoginRadius Blog, "Ensuring Log Integrity and Non-Repudiation for AI Agents", 2026 https: / / www.loginradius.com / blog / engineering / ensure-log-integrity-non-repudiation-ai-agents Ensuring log integrity and non-repudiation for AI agents. While it discusses maintaining the reliability of legal evidence through immutable logging infrastructure, it does not show the implementation configuration for collecting, correlating, and reconstructing the processing paths of multiple AI agents' behavioral events, or for creating tamper-detectable audit records. [Overview of the project]

[0014] Conventional systems typically only record the final output generated by the AI ​​agent, lacking sufficient mechanisms to track intermediate processing and decision-making processes. This makes it difficult to verify the processing path and decision-making rationale later, hindering compliance with audit and accountability requirements.

[0015] In addition, in an environment where multiple AI agents are involved, the mechanism for correlating the processes executed by each agent with each other was insufficient, and it was difficult to trace a consistent processing path for the same case.

[0016] The present invention has been made to solve the above problems, and collects each processing step executed by an AI agent as an action event, correlates these with each other, enables the subsequent reconstruction of the processing path and the judgment process, and stores them in an audit record in a tamper-detectable format. The purpose is to provide an agent action history management system.

Means for Solving the Problems

[0017] The agent action history management system according to one aspect of the present invention is an information processing system in which a plurality of AI agents share and execute at least part of proposal generation, evaluation, integration, or decision support, for each AI agent or for each process executed by the AI agent, at least a history collection unit that acquires an action event including a request identifier, an agent identifier, a process type, reference rule information, input or reference data, an output result, an execution time, and related information indicating context; a history association unit that associates a plurality of action events acquired by the history collection unit with each other based on at least one of the same case, the same processing sequence, a reference source event, or a subsequent event; a history reconstruction unit that reconstructs at least part of the processing path, the judgment process, or the output generation process by the plurality of AI agents based on the group of action events associated by the history association unit; an audit record unit that stores the action event, the association result, and the reconstruction result as an audit record in a tamper-detectable format; Equipped with, Based on the audit records stored in the aforementioned audit record unit, it is possible to track which AI agent arrived at what output or decision for a given case, based on what reference rule information or input or reference data. It is characterized by the following:

[0018] The history association unit is characterized by associating the plurality of action events based on at least two of the following: case identifier, request identifier, event reference identifier, processing sequence identifier, and execution time.

[0019] The audit record unit is characterized by generating a hash value for each action event or audit record, and storing the hash value in association with the hash value corresponding to the record preceding the action event or audit record.

[0020] The history reconstruction unit is characterized by detecting at least one of the following based on the associated group of action events: an unprocessed event, a conflicting judgment result, or a branching event that does not satisfy a predetermined condition.

[0021] The history reconstruction unit is characterized by generating a comparison result for each of the plurality of AI agents that includes at least one of the processing path from the start of processing to the completion of processing, response time, output result, or judgment difference.

[0022] The audit records stored in the aforementioned audit record unit are configured to be accessible from at least one of the administrator terminal, audit terminal, or supervisor terminal.

[0023] The aforementioned action event is characterized by including at least one of the consent information, approval policy, version information, reference rule information, or setting basis information referenced in the processing.

[0024] The history collection unit is characterized by collecting action events corresponding to at least one of the following: proposal generation, evaluation execution, integrated result output, decision support result output, hold determination, or re-verification request. [Effects of the Invention]

[0025] According to the present invention, each processing step performed by an AI agent can be collected as an action event, correlated with each other, and the processing path and decision-making process can be reconstructed later. This data can then be stored as an audit record in a tamper-proof format. Furthermore, by classifying and collecting action events based on a fixed vocabulary system, searchability and traceability can be improved. In addition, by storing the audit record using a hash chain, an audit trail can be maintained in a tamper-proof format. [Brief explanation of the drawing]

[0026] [Figure 1] This figure shows the overall configuration of an agent behavior history management system according to an embodiment of the present invention. [Figure 2] This diagram shows a list of the types of processing that the AI ​​agent will perform. [Figure 3] This figure shows the processing flow according to an embodiment of the present invention. [Figure 4] This diagram shows the hierarchical structure of the event classification vocabulary system. [Figure 5] This is a table showing common attributes of events. [Figure 6] This is a diagram showing the history association. [Figure 7] This is a diagram showing the reconstruction of the history. [Figure 8] This diagram shows the hash chain of audit records. [Figure 9] This diagram shows the references on the audit terminal. [Figure 10] This diagram shows how supporting information is linked. [Figure 11]This figure shows the hardware configuration of an agent behavior history management system according to an embodiment of the present invention. [Modes for carrying out the invention]

[0027] (Overall structure) Figure 1 shows an embodiment of the present invention. form This diagram shows the overall configuration of the agent behavior history management system. The system comprises multiple AI agents 10, a history collection unit 20, a history association unit 30, a history reconstruction unit 40, an audit recording unit 50, and an audit terminal 60.

[0028] AI agent 10 performs processing by sharing at least a part of proposal generation, evaluation, integration, or decision support. Examples of the types of processing performed by AI agent 10 include document generation processing, data inference processing, analysis execution processing, suitability evaluation processing, quality evaluation processing, risk evaluation processing, result integration processing, data aggregation processing, consensus generation processing, recommendation presentation processing, risk warning processing, decision option presentation processing, hold judgment processing, additional confirmation request processing, manual confirmation branching processing, authentication re-verification request processing, authority re-verification request processing, condition re-verification request processing, etc. (see Figure 2).

[0029] The history collection unit 20 acquires an action event for each AI agent 10 or for each process performed by that AI agent, including a request identifier, agent identifier, processing type, reference rule information, input or reference data, output result, execution time, and related information indicating the context. The action events are classified based on a fixed vocabulary system.

[0030] The history association unit 30 associates multiple action events acquired by the history collection unit 20 with each other based on at least one of the following: the same case, the same processing sequence, the source event, or the subsequent event. The history association unit 30 associates action events based on at least two of the following: case identifier, request identifier, event reference identifier, processing sequence identifier, and execution time.

[0031] The history reconstruction unit 40 reconstructs at least a portion of the processing paths, decision processes, or output generation processes of multiple AI agents based on the group of action events associated by the history association unit 30. Based on the group of action events associated, the history reconstruction unit 40 detects at least one of the following: an unprocessed event, a conflicting decision result, or a branching event that does not satisfy a predetermined condition.

[0032] The audit record unit 50 stores the action event, association results, and reconstruction results as audit records in a format that allows for tamper detection. The audit record unit 50 generates a hash value for each action event or audit record and stores the hash value in association with the hash value corresponding to the record preceding the action event or audit record (hash chaining).

[0033] The audit terminal 60 is a terminal used by the system administrator or auditor, and it is possible to refer to the audit records stored in the audit record unit 50 and audit the processing execution history by the AI ​​agent 10.

[0034] (Processing flow) Figure 3 shows an embodiment of the present invention. form This diagram shows the processing flow related to this implementation. form Then, during the process in which the AI ​​agent 10 executes the processing, the history collection unit 20 acquires the action events, the history association unit 30 associates the action events, the history reconstruction unit 40 reconstructs the processing path, and the audit record unit 50 saves it as an audit record.

[0035] Step S1: AI agent 10 executes the process and generates action events at each step, such as the start, in progress, completion, and error of the process.

[0036] Step S2: The history collection unit 20 acquires action events from the AI ​​agent 10. Action events include a request identifier, agent identifier, processing type, reference rule information, input or reference data, output result, execution time, and related information indicating the context.

[0037] Step S3: The history association unit 30 associates the acquired action events with each other based on at least one of the following: the same case, the same processing sequence, the source event, or the subsequent event.

[0038] Step S4: The history reconstruction unit 40 reconstructs the processing path, decision process, or output generation process based on the associated action event group. Step S5: The audit record unit 50 saves the action events, association results, and reconstruction results as audit records.

[0039] (Event classification vocabulary system) This implementation form Behavioral events are classified based on a fixed vocabulary system. Figure 4 shows the hierarchical structure of the event classification vocabulary system. Event classification consists of three levels: L1 major classification (processing phase), L2 medium classification (processing type), and L3 minor classification (specific event).

[0040] (L1 Major Category: Processing Phase) The L1 major classification divides the overall phases of processing performed by the AI ​​agent. These include the E1 proposal generation phase, E2 evaluation execution phase, E3 integrated result output phase, E4 decision support result output phase, E5 hold judgment phase, E6 re-verification request phase, E7 external system integration phase, and E8 management / audit phase.

[0041] (E1: Proposal generation phase) The E1 proposal generation phase is the phase in which the AI ​​agent generates new proposals, drafts, and data. There are three subcategories: E1-1 document generation, E1-2 data inference, and E1-3 analysis execution. For example, for E1-1 document generation, four subcategories of events are defined: E1-1-01 document generation start, E1-1-02 document generation in progress, E1-1-03 document generation complete, and E1-1-99 document generation error.

[0042] The document generation start event (E1-1-01) collects the request ID, agent ID, processing type, reference rule, and execution time. The document generation in progress event (E1-1-02) collects the progress rate, intermediate data, and execution time. The document generation completion event (E1-1-03) collects the output result, execution time, and subsequent event reference ID. The document generation error event (E1-1-99) collects the error details, execution time, and context.

[0043] (E2: Evaluation Execution Phase) The E2 evaluation execution phase is the phase in which the AI ​​agent evaluates and scores existing proposals and data. There are three subcategories: E2-1 conformity evaluation, E2-2 quality evaluation, and E2-3 risk evaluation. For example, for E2-1 conformity evaluation, four sub-category events are defined: E2-1-01 conformity evaluation start, E2-1-02 conformity evaluation in progress, E2-1-03 conformity evaluation completion, and E2-1-99 conformity evaluation error.

[0044] The conformity assessment start event (E2-1-01) collects the request ID, agent ID, item to be assessed, reference rule version, and execution time. The conformity assessment execution event (E2-1-02) collects the intermediate score, reference items, and execution time. The conformity assessment completion event (E2-1-03) collects the assessment result, score, reason for judgment, and execution time.

[0045] (E3: Integrated Result Output Phase) The E3 integrated results output phase is the phase in which multiple proposals and data are integrated and the results are output. There are three subcategories: E3-1 results integration, E3-2 data aggregation, and E3-3 consensus generation. For example, for E3-1 results integration, four sub-category events are defined: E3-1-01 results integration start, E3-1-02 results integration in progress, E3-1-03 results integration complete, and E3-1-99 results integration error.

[0046] (E4: Decision support result output phase) The E4 decision support result output phase is the phase in which information and recommendations are output to guide the final decision. There are three subcategories: E4-1 recommendation presentation, E4-2 risk warning, and E4-3 decision option presentation. For example, for E4-1 recommendation presentation, four sub-category events are defined: E4-1-01 recommendation presentation start, E4-1-02 recommendation presentation in progress, E4-1-03 recommendation presentation completion, and E4-1-99 recommendation presentation error.

[0047] (E5: Pending Decision Phase) The E5 hold determination phase is the phase in which processing is determined and executed for hold or additional confirmation requests. There are three subcategories: E5-1 hold determination, E5-2 additional confirmation request, and E5-3 manual confirmation branch. For example, for E5-1 hold determination, four sub-category events are defined: E5-1-01 hold determination start, E5-1-02 hold determination execution in progress, E5-1-03 hold determination completion, and E5-1-99 hold determination error.

[0048] (E6: Re-verification request phase) The E6 re-verification request phase is the phase in which a re-verification of authentication, authorization, and conditions is requested. There are three subcategories: E6-1 authentication re-verification request, E6-2 authorization re-verification request, and E6-3 condition re-verification request. For example, for the E6-1 authentication re-verification request, four sub-category events are defined: E6-1-01 authentication re-verification request started, E6-1-02 authentication re-verification request in progress, E6-1-03 authentication re-verification request completed, and E6-1-99 authentication re-verification request error.

[0049] (Common attributes for events) Figure 5 is a table showing common event attributes. All action events have the following common attributes: event_id (event-specific identifier), request_id (request identifier), agent_id (agent identifier), event_classification_L1 (L1 major category ID), event_classification_L2 (L2 subcategory ID), event_classification_L3 (L3 minor category ID), event_name (minor category event name), processing_type (processing type), input_data (input or reference data), output_result (output result), reference_rule_info (reference rule information), execution_timestamp (execution time), parent_event_id (context / parent event ID), child_event_ids (context / child event ID group), processing_series_id (processing series identifier), justification_info (justification information), hash_value (hash value), previous_hash_value (previous hash value).

[0050] These common attributes enable the association of behavioral events, the reconstruction of processing paths, and the detection of tampering.

[0051] (Hash chain of audit records) Figure 8 shows a hash chain of audit records. The audit record unit 50 generates a hash value for each action event or audit record, and stores the hash value in association with the hash value corresponding to the record preceding the action event or audit record.

[0052] Specifically, a hash value H(N) is generated for each action event N. The hash value H(N) is generated using the content of action event N and the hash value H(N-1) of the preceding action event (N-1) as input. This allows for a chain of action events in chronological order, and if any event is tampered with, the hash values ​​of subsequent events will be mismatched, making it possible to detect tampering.

[0053] For generating hash values, it is preferable to use a cryptographic hash function such as SHA-256.

[0054] (Hardware configuration) Figure 11 shows the hardware configuration of an agent behavior history management system according to an embodiment of the present invention. This system includes a CPU 101, memory 102, storage 103, network interface 104, and input / output interface 105.

[0055] The CPU 101 executes the programs deployed in the memory 102 to realize the functions of the history collection unit 20, history association unit 30, history reconstruction unit 40, and audit record unit 50. The memory 102 is a volatile memory such as RAM, and temporarily stores programs and data. The storage 103 is a non-volatile storage device such as an HDD or SSD, and permanently stores programs, data, audit records, etc.

[0056] The network interface 104 is an interface for communication with the AI ​​agent 10, the audit terminal 60, etc. The input / output interface 105 is an interface for connecting to input / output devices such as a keyboard, mouse, and display. [Examples]

[0057] (History association) Figure 6 shows the history association. The history association unit 30 associates multiple action events based on at least two of the following: case identifier (request_id), event reference identifiers (parent_event_id, child_event_ids), processing series identifier (processing_series_id), and execution time (execution_timestamp).

[0058] For example, if, for case A, AI agent 10-1 performs document generation processing and AI agent 10-2 evaluates the results, the document generation completion event (E1-1-03) and the suitability evaluation start event (E2-1-01) are associated by the same request_id. Additionally, the parent_event_id of the conformance assessment start event is set to the event_id of the document generation completion event.

[0059] By linking multiple behavioral events in this way, it becomes possible to track the processing path and decision-making process afterward. [Examples]

[0060] (History reconstruction) Figure 7 shows the history reconstruction process. The history reconstruction unit 40 reconstructs the processing path, decision process, or output generation process based on the associated action event group.

[0061] For example, for case A, the history reconstruction unit 40 reconstructs the following processing path: document generation start (E1-1-01) → document generation in progress (E1-1-02) → document generation completed (E1-1-03) → conformity evaluation start (E2-1-01) → conformity evaluation in progress (E2-1-02) → conformity evaluation completed (E2-1-03).

[0062] Furthermore, the history reconstruction unit 40 detects at least one of the following based on the associated action event group: an unprocessed event, a conflicting judgment result, or a branching event that does not meet predetermined conditions. For example, if a document generation start event exists but a document generation completion event does not, it is detected as an unprocessed event. Also, if the evaluation results of the conformity evaluation completion event and the quality evaluation completion event are contradictory, it is detected as a conflicting judgment result.

[0063] Furthermore, the history reconstruction unit 40 generates a comparison result for each of the multiple AI agents, which includes at least one of the processing path from the start to the completion of processing, the response time, the output result, or the difference in judgment. This makes it possible to compare and analyze the processing efficiency and judgment accuracy among the AI ​​agents. [Examples]

[0064] (Reference on audit terminal) Figure 9 shows the access on the audit terminal. The audit records stored in the audit record unit 50 are configured to be accessible from at least one of the administrator terminal, audit terminal 60, or supervisor terminal.

[0065] The audit terminal 60 can track which AI agent arrived at what output or decision for a given case, based on what reference rule information or input or reference data. It can also efficiently search for specific types of events based on an event classification vocabulary system.

[0066] For example, a search such as "Show the results of the E2-1 conformity assessment for project A" can extract all events related to the start of the conformity assessment (E2-1-01), the execution of the conformity assessment (E2-1-02), and the completion of the conformity assessment (E2-1-03). [Examples]

[0067] (Linking to supporting information) Figure 10 shows the linking of supporting information. In this embodiment, an action event includes at least one of the consent information, approval policy, version information, reference rule information, or setting supporting information referenced in the process.

[0068] For example, in the conformity assessment completion event (E2-1-03), in addition to the assessment result, score, and reason for the judgment, the version information of the referenced approval policy, the version information of the referenced rule, and the basis for setting the assessment criteria are recorded as justification_info. This makes it possible to clarify the basis for the assessment result at a later date.

[0069] Furthermore, for processes that require user consent information, it becomes possible to prove that the process was executed based on consent by associating the consent acquisition event with the process execution event. [Industrial applicability]

[0070] The agent behavior history management system of the present invention is applicable to all information processing systems in which multiple AI agents share and execute complex processing tasks. For example, it can be applied to corporate business systems, financial institution risk management systems, medical institution diagnostic support systems, government agency review support systems, and so on. [Explanation of Symbols]

[0071] 10 AI Agents 10-1, 10-2, 10-3 AI Agents (referring to one of several 10 AI agents) 20 History Collection Department 30 History linking section 40 History Reconstruction Unit 50 Audit Records Department 60 Audit terminals 101 CPU 102 memory 103 Storage 104 Network Interfaces 105 Input / Output Interfaces

Claims

1. An information processing system in which multiple AI agents share the task of generating, evaluating, integrating, or supporting decision-making, and perform the processing accordingly. A history collection unit acquires an action event for each AI agent or for each process performed by said AI agent, which includes at least a request identifier, agent identifier, processing type, reference rule information, input or reference data, output result, execution time, and related information indicating the context. A history association unit associates multiple action events acquired by the history collection unit with each other based on at least one of the following: the same case, the same processing sequence, the source event, or the subsequent event. A history reconstruction unit reconstructs at least a part of the processing path, decision process, or output generation process of the multiple AI agents based on the group of behavioral events associated by the history association unit, An audit record unit that stores the aforementioned action events, association results, and reconstruction results as audit records in a format that allows for tamper detection, Equipped with, Based on the audit records stored in the aforementioned audit record unit, it is possible to track which AI agent arrived at what output or decision for a given case, based on what reference rule information or input or reference data. An agent behavior history management system characterized by the following:

2. In the agent behavior history management system described in claim 1, The history association unit associates the multiple action events based on at least two of the following: case identifier, request identifier, event reference identifier, processing sequence identifier, and execution time. An agent behavior history management system characterized by the following:

3. In the agent behavior history management system according to claim 1 or 2, The audit record unit generates a hash value for each action event or audit record, and stores the hash value in association with the hash value corresponding to the record preceding the action event or audit record. An agent behavior history management system characterized by the following:

4. In the agent behavior history management system according to any one of claims 1 to 3, The history reconstruction unit detects at least one of the following based on the associated group of action events: an unprocessed event, a conflicting judgment result, or a branching event that does not satisfy a predetermined condition. An agent behavior history management system characterized by the following:

5. In the agent behavior history management system according to any one of claims 1 to 4, The history reconstruction unit generates a comparison result for each of the plurality of AI agents, which includes at least one of the processing path from the start of processing to the completion of processing, the response time, the output result, or the judgment difference. An agent behavior history management system characterized by the following:

6. In the agent behavior history management system according to any one of claims 1 to 5, The audit records stored in the aforementioned audit record unit are configured to be accessible from at least one of the administrator terminal, audit terminal, or supervisor terminal. An agent behavior history management system characterized by the following:

7. In the agent behavior history management system according to any one of claims 1 to 6, The aforementioned action event includes at least one of the consent information, approval policy, version information, reference rule information, or setting basis information referenced in the processing, An agent behavior history management system characterized by the following:

8. In the agent behavior history management system according to any one of claims 1 to 7, The history collection unit collects action events corresponding to at least one of the following: proposal generation, evaluation execution, integrated result output, decision support result output, hold judgment, or re-verification request. An agent behavior history management system characterized by the following:

Images