Agent system

The agent system addresses challenges in multi-agent systems by scoring and correcting agent identity using evaluators and PageRank, ensuring reliable and secure collaboration and network expansion.

JP2026182331APending Publication Date: 2026-11-19NOMURA RESEARCH INSTITUTE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
JP2025078479
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2025-05-09
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

In multi-agent systems, maintaining proper task processing while facilitating continuous communication between users and agents is challenging due to environmental threats such as changes in AI learning, tool updates, and tampering, which can lead to catastrophic forgetting and inconsistent agent behavior.

Method used

An agent system where multiple agents cooperate to execute tasks, utilizing evaluators to score the identity of each agent based on similarity of outputs, and correct the identity score using PageRank to ensure reliability, enabling secure and efficient collaboration.

Benefits of technology

The system allows for continuous understanding and scoring of agent identity over time, ensuring reliable selection of appropriate agents for tasks and secure expansion of the agent network.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2026182331000001_ABST
    Figure 2026182331000001_ABST
Patent Text Reader

Abstract

This system enables the scoring and tracking of the identity of each agent in a multi-agent system over time. [Solution] Each node 10 includes an evaluator 12 for each agent 11. The first evaluator 12 for each first agent 11 in each agent 11 aggregates the first scores related to the similarity of the output calculated for each of the other second agents 11 using a predetermined method and transmits them to each second evaluator 12. The second evaluator 12 receives the second score calculated by each second evaluator 12 and transmits it, and calculates an identity score for each agent 11 based on the first and second scores for each agent 11.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical Field]

[0001] The present invention relates to agent system technology, and more particularly to technology that is effective when applied to agent systems in which multiple agents cooperate and collaborate with each other to process tasks. [Background technology]

[0002] With advancements in technologies such as generative AI (Artificial Intelligence), LLM (Large Language Model), and LMM (Large Multimodal Model), research and development of agent systems are progressing in which software (agents) autonomously decide what to do from various pieces of information to achieve a predetermined objective and then process that information. Agent systems include single-agent systems, where a single agent processes a task, and multi-agent systems, where multiple agents cooperate and collaborate with each other to process a task.

[0003] To realize a multi-agent system, Google® has announced the A2A (Agent2Agent) protocol (Non-Patent Literature 1), a standardized open protocol for enabling cooperation between agents regardless of the underlying framework or vendor of each agent. In the A2A protocol, both agents are equipped with an API (Application Programming Interface), enabling bidirectional dialogue between agents. The agents here are not limited to AI, but may be other services, tools, programs, etc.

[0004] Furthermore, there is a technology called MCP (Model Context Protocol) (Non-Patent Literature 2) for realizing agent systems that include generative AI. MCP is a standardized open protocol for generative AI to interact with external tools and data sources, but it is a client / server model, and the initiation of agent-to-agent dialogue is determined by the client's specifications. On the other hand, in the A2A protocol, simultaneous dialogue between agents can occur. A2A is not a technology that competes with MCP, but rather a technology that complements MCP. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] "Announcing the Agent-to-Agent Protocol (A2A): A New Era of Agent Interoperability," [online], Google LLC, [Accessed April 24, 2025], Internet <URL:https: / / cloud.google.com / blog / ja / products / ai-machine-learning / a2a-a-new-era-of-agent-interoperability> [Non-Patent Document 2] “Introduction - Model Context Protocol”, [online], [Accessed April 24, 2025], Internet<URL:https: / / modelcontextprotocol.io / introduction> [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] In multi-agent systems, including generative AI, maintaining proper task processing while facilitating continuous communication between users and agents, and even between agents themselves, requires considering environmental threats such as changes in what the AI ​​has learned over time, tool updates, and tampering (i.e., threats to the agent's identity over time).

[0007] For example, in the generative AI that an agent possesses or uses, learning new information can lead to so-called catastrophic forgetting, potentially causing a significant loss of information about previously learned tasks and resulting in the loss of its original capabilities. Furthermore, updates or replacements of the generative AI or software itself could cause the agent to behave differently than before. Additionally, tampering techniques such as DNS (Domain Name System) cache poisoning could result in a situation where, despite specifying the same URI (Uniform Resource Identifier), the agent is interacting with different generative AIs, tools, or individuals.

[0008] Therefore, an object of the present invention is to provide an agent system that enables the scoring and understanding of the identity of each agent in a multi-agent system over time. The aforementioned and other objects and novel features of the present invention will become apparent from the description herein and the accompanying drawings. [Means for solving the problem]

[0009] A brief overview of some of the representative inventions disclosed in this application is as follows:

[0010] A representative embodiment of the present invention is an agent system in which multiple agents, including those that utilize generative AI, constitute nodes in an agent network, and each of these agents cooperates with each other to execute processing requested by the user.

[0011] Each node includes an evaluator for each agent. The first evaluator for each first agent in each agent aggregates the first scores related to the similarity of the outputs calculated for each of the other second agents using a predetermined first method and transmits them to each second evaluator for each of the second agents. The first evaluator then obtains the second scores transmitted from each second evaluator, which are calculated by each second evaluator using the first method, and calculates an identity score for each agent based on the first and second scores for each agent. [Effects of the Invention]

[0012] The effects obtained by some of the representative inventions disclosed in this application can be briefly explained as follows:

[0013] In other words, according to a typical embodiment of the present invention, it becomes possible to score and understand the identity of each agent in a multi-agent system over time. [Brief explanation of the drawing]

[0014] [Figure 1] This figure outlines an example configuration of an agent system, which is one embodiment of the present invention. [Figure 2] This flowchart outlines an example of a processing flow for scoring the identity of an agent in one embodiment of the present invention. [Figure 3] Figures (a) to (d) are diagrams illustrating an example of converting the similarity score in one embodiment of the present invention into a matrix format. [Figure 4](a) and (b) are diagrams showing an overview of an example of a sparse agent network according to an embodiment of the present invention. [Figure 5] This is a sequence diagram showing an overview of an example of the expansion processing (preliminary preparation) of the agent network according to an embodiment of the present invention. [Figure 6] This is a sequence diagram showing an overview of an example of the expansion processing of the agent network according to an embodiment of the present invention. [Figure 7] This is a sequence diagram showing an overview of another example of the expansion processing of the agent network according to an embodiment of the present invention. [Mode for Carrying Out the Invention]

[0015] Hereinafter, embodiments of the present invention will be described in detail based on the drawings. Note that, in all the drawings for describing the embodiments, the same reference numerals are, in principle, assigned to the same parts, and repeated description thereof will be omitted. On the other hand, regarding portions that have been described with reference numerals in a certain drawing, the same reference numerals may be used to refer to the portions without re-illustrating them in the description of other drawings.

[0016] <Overview> As described above, the agent system that is one embodiment of the present invention is a multi-agent system in which a plurality of agents cooperate and coordinate with each other to process tasks, and enables scoring and grasping the temporal identity of each agent. This makes it possible to select a more appropriate agent for a user's input and process the task.

[0017] In this embodiment of the agent system, cooperation and collaboration between agents (including AI, humans, software, etc.) are described as being implemented using the A2A protocol. However, this is not limited to this, and it goes without saying that other existing technologies and future technologies can be applied as appropriate, as long as each agent can cooperate and collaborate with one another. In particular, if the AI ​​agent uses external software or tools, the cooperation may be implemented in the form of MCP.

[0018] <System Configuration> Figure 1 is a diagram illustrating an example configuration of an agent system, which is one embodiment of the present invention. Agent system 1 is an information processing system that realizes a multi-agent system by comprising one or more server devices, virtual servers built on a cloud computing service, or information processing terminals, and by using a CPU (Central Processing Unit) (not shown) on each device to execute middleware such as an OS (Operating System), DBMS (Database Management System), and Web server programs, as well as software running on them, which are deployed from a storage device such as an HDD (Hard Disk Drive) or SSD (Solid State Drive) into memory.

[0019] This agent system 1 includes, for example, multiple agents implemented as software (in the example in Figure 1, agents 11a, 11b, and 11c, which may be collectively referred to as "agent 11" hereafter), multiple evaluators (in the example in Figure 1, evaluators 12a, 12b, 12c, and 12d, which may be collectively referred to as "evaluator 12" hereafter), an authorization unit 20, and an agent search unit 30, among other components.

[0020] Agent 11 is software that performs processing suited to a specific function or purpose, and may be equipped with or utilize AI such as generative AI, as in Agents 11a and 11c, or it may be a tool or program that does not utilize AI, as in Agent 11b. In addition, some of these Agents 11 may include a host agent that serves as the interface for the agent network 2 for the user (user agent 3), as will be described later. These Agents 11 can communicate with each other and cooperate and operate in cooperation via the A2A protocol over a network not shown, and conceptually constitute the agent network 2 by A2A as nodes (in the example in Figure 1, these are nodes A (10a), B (10b), and C (10c), and each of these nodes may be collectively referred to as "node 10" below).

[0021] Each agent 11 in node 10 is equipped with an evaluator 12 that scores its identity against agents 11 in other nodes 10. In other words, in this embodiment, the identity of a given agent 11 is evaluated from multiple perspectives based on evaluation by others, specifically by the evaluators 12 in each other node 10. Note that each agent 11 in node A (10a), node B (10b), and node C (10c) is equipped with evaluators 12a, 12b, and 12c, respectively. However, it is not necessarily required to configure node 10 with sets of agents 11 and evaluators 12. As will be described later, it is also possible to configure node 10 with only agents 11 without evaluators 12.

[0022] Furthermore, there may be nodes 10 that consist only of evaluators 12d and do not have agents 11, such as node D (10d) located outside agent network 2 (in this case, as will be described later, processing can be carried out without inconsistency by treating the evaluator 12d as equivalent to agent 11). Details of the identity scoring process by evaluator 12 will also be described later.

[0023] In the example shown in Figure 1, agent system 1 is composed of four nodes, A (10a) to D (10d). However, it goes without saying that the number of nodes 10, the presence or absence of agents 11 and evaluators 12, etc., are not limited to the configuration shown in Figure 1. Also, in the example in Figure 1, the arrows between the four nodes 10 indicate a fully connected configuration in which each node 10 interacts with (i.e., evaluates) all the other nodes 10. However, as will be discussed later, a configuration including nodes 10 that interact with only some of the other nodes 10 is also possible.

[0024] The authorization unit 20 has the function of performing authorization processing when adding a new agent 11 to the agent network 2. As will be described later, the addition of a new agent 11 here has the function of performing authorization processing for the authority of the host agent, which is the interface for the agent network 2 for the user (user agent 3), when registering the new agent 11 as a known agent that can cooperate with the agent in the agent list.

[0025] The agent search unit 30 has the function of searching for new agents (agents not included in the agent network 2) that meet the requirements for processing a task. The authorization unit 20 and the agent search unit 30 may also be configured to function as individual agents in the agent system 1.

[0026] The user accesses Agent System 1 via a network (not shown) using an information processing terminal such as a PC (Personal Computer) and requests the processing of a desired task. The software that receives the user's request on the information processing terminal may also be configured to function as an individual agent (User Agent 3) in Agent System 1.

[0027] <Processing flow> As described above, in this embodiment, the time-series identity of a certain agent 11 is evaluated from multiple perspectives based on evaluation by others, specifically by scoring by the evaluators 12 of each other node 10 (agent 11).

[0028] Regarding the evaluation of the identity of things, in the real world, for example, identification information such as IDs, names, and dates of birth is used for identity verification, and in the case of electronic data, identity is sometimes confirmed by electronic signatures. In contrast, in the case of agents consisting of software such as AI, it is sufficient if identification information such as name and version information can be obtained, but not all agents necessarily provide this information, and even AIs with the same name and version may give different answers to the same question depending on the learning content and the temperature (temperature) during inference, so it may not be sufficient.

[0029] Therefore, in this embodiment, the temporal identity of agent 11 is evaluated by the identity of the output in response to the input to agent 11, that is, by how similar the output is when the same input is given to agent 11 as before.

[0030] Figure 2 is a flowchart outlining an example of the process flow for scoring the identity of agent 11 in one embodiment of the present invention. First, for a specific agent 11 to be processed (here, for the sake of explanation, let's call it agent 11a of node A (10a) in the example in Figure 1), a loop process is started that makes all other agents 11 (the agent 11 to be scored or other known agents 11) subject to evaluation by others (S01).

[0031] In the loop processing, a set of input patterns is prepared for the agent 11 to be evaluated (S02). For example, this could be a fixed prompt, an image, or audio (A2A is designed to support various modalities). These may also have perturbations such as noise based on a certain distribution. This input pattern may be set by the user (or user agent 3, hereafter the same) for the evaluator 12a related to the agent 11a to be processed, or the evaluator 12a or the user may select from one or more input patterns pre-registered in the evaluator 12a. Furthermore, the input pattern may be unified among all evaluators 12, or each evaluator 12 may have a different input pattern.

[0032] Next, the prepared input pattern is fed into the agent 11 to be evaluated (S03), and the output result is obtained (S04). Then, it is determined whether or not this is the first time the output value has been obtained from the agent 11 to be evaluated (S05). If it is the first time (Yes in step S05), the output value (or embedded representation) is registered in the evaluator 12a as a registered value to be compared thereafter (S06), and the process moves on to the next agent 11 to be evaluated (S09, S01).

[0033] On the other hand, if the output value from the agent 11 being evaluated is obtained for the second time or later (No in step S05), the evaluator 12a calculates the similarity between the output value (or embedded representation) and the registered value registered in step S06 and uses this as the score (S07). The method for calculating the similarity is not particularly limited, and any appropriate method such as cosine similarity can be used. In addition, if the output value is text, an index such as the BLEU (BiLingual Evaluation Understudy) score may be used, and if it is an image, an index such as the Frechet Inception Distance (FID) may be used, or an index evaluated based on the difference or distribution of the accuracy or loss value of a predetermined task may be used.

[0034] Furthermore, in step S02, multiple input patterns may be prepared, and a score may be calculated based on multiple output values ​​obtained from these input patterns. Additionally, if possible, the registered value registered in step S06 may be modified or updated based on the output values ​​obtained from the second time onward. For example, statistical values ​​such as a moving average of the output values ​​may be calculated based on the output values ​​obtained from the second time onward, and these may be used as the new registered value.

[0035] After calculating the similarity score, the score (or the set of scores calculated based on the output values ​​obtained from each of the multiple input patterns) is converted into a matrix format (adjacency matrix, or an adjacency matrix that allows self-loops, etc.) and recorded (S08).

[0036] Figure 3 is a diagram illustrating an example of converting the similarity score in one embodiment of the present invention into a matrix format. Figure 3(a) shows a matrix in tabular form that holds the similarity scores calculated by the evaluators 12 of "Node A" to "Node D" for the agents 11 of the other nodes 10, and the similarity score S calculated by the evaluator 12 of "Node A" for the agent 11 of "Node B" is shown. a,b This indicates the recorded state.

[0037] Returning to Figure 2, the similarity scores are converted into a matrix format and recorded, and then the processing of the next agent to be evaluated, 11, is moved on (S09, S01). The loop process ends when processing for all other agents 11 is complete. Figure 3(b) shows the state of the matrix at the end of the loop process in tabular format. The similarity scores S calculated for "Node A" for the other agents 12 of "Node B", "Node C", and "Node D" are calculated respectively. a,b S a,c , and S a,dThis indicates the recorded state for each. Note that for "Node A," its own agent 11 is not subject to evaluation or scoring, so its score in the matrix remains "0" (it is not forbidden to evaluate itself, and as will be explained later, it may evaluate itself).

[0038] Returning to Figure 2, similarity scores are calculated for all other agents 11 to obtain a matrix like the example in Figure 3(b), which is then sent to the evaluators 12 of all other agents 11 (S10). Note that in practice, not the entire matrix is ​​sent, but rather a vector relating to the similarity scores calculated by the evaluator 12 being processed (the row vector of the first row in the example in Figure 3(b)). It may also be sent in a predetermined format such as JSON-LD (JSON (JavaScript Object Notation) for Linking Data).

[0039] Since the same processing is performed by the evaluators 12 of other agents 11 besides the agent 11 being processed, these evaluators 12 of other agents 11 also transmit the matrix (vector) relating to the similarity score they calculated to the evaluators 12 of the other agents 11 (S11). As a result, the evaluator 12 of the agent 11 being processed receives the matrix (vector) relating to the similarity score evaluated by each of the other evaluators 12 of the other agents 11 and records it (S12). Figure 3(c) shows the state of the matrix in tabular form when vectors relating to the similarity score have been received from the evaluators 12 of all other agents 11 (nodes 10).

[0040] Returning to Figure 2, the matrix relating to the similarity scores is obtained from the evaluators 12 of all other agents 11 to obtain a matrix like the example shown in Figure 3(c). Then, the identity score for each agent 11 is calculated (S13), and the process ends. The identity score can be calculated, for example, by summing the similarity scores for each column in the matrix shown in the example in Figure 3(c). Figure 3(d) shows the identity score T obtained by summing the similarity scores S for each column (each node 10) in the matrix example in Figure 3(c).a , T b , T c , and T d This indicates that each of these has been calculated. Note that the calculation of the identity score is not limited to a simple sum of the similarity scores calculated by the evaluators 12 of each node 10, but may also be a simple or weighted mean, or normalization may be applied, or corrections may be made based on prior knowledge.

[0041] By scoring the identity of each agent 11 using the method described above, it becomes possible to verify and evaluate identity from multiple perspectives, not only based on the subjective view of the agent 11 being processed, but also through the mutual perspectives of each connected agent 11.

[0042] In the example described above, a fully connected configuration is assumed in which the evaluators 12 of each agent 11 (node ​​10) can evaluate and score all other agents 11. However, in an actual agent system 1, it is not always the case that the evaluators 12 of all agents 11 can evaluate and score all other agents 11. Rather, it is assumed that in many cases the agent network 2 will be sparse, with no links to other agents 11 (or agents not having their own evaluators 12).

[0043] Figure 4 is a diagram illustrating an example of a sparse agent network 2 in one embodiment of the present invention. Figure 4(a) shows an example configuration of the sparse agent network 2, in which nodes A(10a) to D(10d) are all fully connected, each having a link to another node 10, but the overall configuration including nodes E(10e) to H(10h) is not fully connected.

[0044] In the example in Figure 4(a), there is no link from node A (10a) to node H (10h), and no links from node B (10b) to nodes E (10e), F (10f), and G (10g). Similarly, there are no links from node C (10c) to nodes E (10e) and H (10h), and no links from node D (10d) to nodes E (10e) and F (10f). Furthermore, there are no links from nodes E (10e) through H (10h) to other nodes 10.

[0045] In other words, nodes E(10e) to H(10h) will be the ones being evaluated for identity, but will not be the ones doing the evaluating. Such a sparse agent network 2 configuration may occur, for example, when nodes A(10a) to D(10d) are the user's own agents 11, and nodes E(10e) to H(10h) are services or tools provided by providers such as Google or Apple(registered trademark).

[0046] In this configuration, as shown in the example in Figure 4(b), the matrix resulting from the identity score evaluation also functions as a matrix showing the link relationships between nodes 10 (adjacency matrix, or an adjacency matrix that allows self-loops) (where the value is zero, it indicates that there is no link). In this embodiment, even with such a sparse agent network 2 configuration, the identity score of each agent 11 can be calculated using the method described above.

[0047] <Correction of identity score> The identity score of each agent 11 calculated using the method described above indicates that a higher score indicates higher identity over time, and can be used as reference information when each agent 11 selects an agent 11 from among the other agents 11 to perform a predetermined task.

[0048] On the other hand, a case may arise where an agent 11 with a high identity-related score is not unconditionally a good agent. For example, in a sparse agent network 2 as shown in the example of Fig. 4 described above, the links between agents 11 (nodes 10) may have varying weights, that is, there may be nodes 10 (evaluators 12) that evaluate the self, with some having a large number and others having a small number. It can be said that the greater the number of other nodes 10 (evaluators 12) that evaluate the self, the higher the reliability of the identity score of the agent 11.

[0049] Accordingly, in the present embodiment, the value of the identity score is corrected based on its reliability, that is, the certainty of evaluation based on the number of verifications and evaluations of the identity score for the agent 11. For example, the identity score is corrected using the concept of "PageRank (registered trademark, the same applies hereinafter)" which was used as one of the indices for prioritizing display of search results in Google's search service. "PageRank" is an algorithm that calculates the importance of a Web page based on the number of incoming links (the number of links referring to the page) and the like, and is an index based on the concept that the greater the number of incoming links, the higher the attention level and the higher the value.

[0050] This concept is applied to the mechanism of multi-perspective mutual evaluation of identity scores between agents 11 as in the present embodiment, and correction is performed such that the greater the number of incoming links = the number of times the identity score has been verified and evaluated, the higher the value of the identity score becomes.

[0051] One of the definitions of the formula for calculating the PageRank PR(A) of a certain Web page A is,

[0052] [Formula] is represented by. Here, PR(T i ) is the PageRank of page T that links to Web page A i , and C(T i ) is page T iThis is the total number of links to other web pages contained within. N is the total number of web pages, and d is the dumping factor (usually 0.85, but a smaller value is used for those who artificially inflate their PageRank). For example, each PR(T i After setting a uniform initial value for ), the calculation of equation 1 above is repeated until the value of PR(A) converges.

[0053] In this embodiment, for the PageRank calculation formula, the identity score AR(A) corrected by PageRank for a certain agent A is used.

[0054]

number

[0055] Here, PR(T i ) is each Agent T adjacent to Agent A i This is the PageRank, and the ID(T i ) is Agent T i This is the identity score. Also, C(T i ) is each agent T adjacent to agent A. i This is the number of agents adjacent to a given agent, where N is the total number of agents and d is the dumping factor. Note that "agents adjacent to a given agent" may be excluded by specification or the range may be limited by conditions. For example, it may be limited to agents (nodes) within a predetermined number of hops from the target agent (node).

[0056] In this embodiment, the method for calculating identity scores from multiple perspectives is described as one in which the evaluator 12 of each node 10 evaluates the agents 11 of other nodes 10. However, the evaluator 12 of each node 10 may also perform a self-evaluation of the agent 11 of its own node 10. In that case, for example, for a certain agent A, the identity score AR(A) obtained by performing the calculation of equation 1 until convergence is performed, and then the identity score ID(A) obtained from the self-evaluation of agent A is used. AR'(A) = AR(A) × ID(A) Alternatively, the identity score AR'(A) may be obtained by correcting it using the formula shown. This can be interpreted as correcting the identity score based on Agent A's own thoughts and values.

[0057] In this embodiment, as shown in the example of node D (10d) in Figure 1, node 10, which does not have agent 11 and is located outside the agent network 2 using the A2A protocol and has only an evaluator 12d, can also function as an evaluator. For example, this can be achieved by treating the evaluator 12d as equivalent to agent 11. In that case, it is required to output and publish the evaluation criteria of the evaluator 12d via an API. That is, the evaluator 12d behaves as both an evaluator 12d and agent 11 (or a proxy for agent 11), and the output of the evaluator 12 can be obtained from others via an API. This allows the above-described processes and mechanisms to be applied as is.

[0058] Agent system 1 may contain multiple evaluators 12 (or just one), and although they share outputs with each other via API, evaluators 12 are independent of the A2A protocol and exist independently of agent network 2. Evaluator 12 only calculates the identity score based on the identity score information obtained from all evaluators 12. If the identity score is calculated using only evaluators 12 outside the A2A protocol (agent network 2), only the identity score calculation is performed, and the correction based on PageRank described above may be skipped.

[0059] It is also possible for an evaluator 12 outside of agent network 2 to coexist and cooperate with an evaluator 12 at node 1 within agent network 2. In that case, when calculating the value to be used as the PageRank for the evaluator 12 outside of agent network 2 when correcting using PageRank, for example, N is the sum of the total number of agents 11 that have evaluators 12 within agent network 2 and the total number of evaluators 12 outside of agent network 2. Then, by setting the PageRank of the evaluator 12 outside of agent network 2 to a fixed value of 1 / N, and treating the node 10 related to that evaluator 12 as also being included in agent network 2, the method of correcting identity scores using PageRank can be applied.

[0060] <Extension of Autonomous Form Network 2> As an example of the application of the agent system 1 of this embodiment, which has the identity scoring mechanism of agent 11 described above, it is possible to realize a function that newly searches for necessary agents 11 according to the content of the task to be processed and automatically adds them to the agent network 2 to expand it.

[0061] For example, in the configuration example shown in Figure 1, the agent search unit 30 stores attribute information for each available agent (such as "url", "description", and "method" in the A2A protocol) in a database, or makes it searchable via a peer-to-peer (P2P) method. Agent 11 expresses the content of the task it wants to execute in text or the like, and searches for an agent using the agent search unit 30. The agent search unit 30 sorts the search results in descending order of score and outputs them. Agent 11 registers the agent with the highest score in its agent list.

[0062] Furthermore, if allowing unlimited conversations with other new agents in this process is deemed dangerous, the authorization unit 20 in the configuration example of Figure 1 may be configured to allow registration only after user approval of the authorization. In this case, the user may also be required to approve the authorization after undergoing user authentication according to a predetermined protocol such as OAuth.

[0063] One possible use case scenario for extending agent network 2 is as follows: Up until now, all tasks that a certain agent 11 has received from a user have been completed using only text and images / videos, and the other agents 11 that agent 11 can collaborate with are also only those that can handle text and images / videos, and there have been no particular problems. However, in this situation, the user requests that "create a video with audio," but this request cannot be processed or resolved using only the agents 11 that agent 11 knows.

[0064] At this time, agent 11 can use the search function of agent search unit 30 to find an agent that can handle voice. However, if multiple search results are obtained, without a criterion for deciding which agent to prioritize, it is impossible to identify the agent that should be used in collaboration. To address this problem, if an agent with a high identity score, as described above, can be selected, it is possible to connect with a high-quality agent for processing (and if user authorization is obtained, it can be made into an even more secure system).

[0065] Figure 5 is a sequence diagram illustrating an example of the processing (preparation) for extending the agent network 2 in one embodiment of the present invention. Here, as preparation, agent 11 (host agent 11m in the figure), which receives task processing requests from the user, performs the process of calculating the identity score of each other known agent 11n, that is, the scoring process shown in the example of Figure 2 above. Note that the scoring process shown in the example of Figure 2 has been performed once, and subsequent processing is assumed.

[0066] In the example shown in Figure 5, first, the host agent 11m sends the identity score it calculated for other known agents 11n to each known agent 11n (and its evaluator 12) (S10). Here, known agents 11n are either agents 11 that have an evaluator 12 or evaluators 12 that do not have an agent 11. This process corresponds to step S10 in Figure 2. Each known agent 11n similarly scores the identity of each other agent 11 in the agent network 2 or the network between evaluators 12 and responds (S11). At this time, a list of references of each other agent 11 is collected and this information is also included in the response. This process corresponds to step S11 in Figure 2.

[0067] The host agent 11m collects the identity score and reference information of each agent 11 responded to by each known agent 11n (S12), calculates the identity score of each known agent 11n and stores it (S13). These processes correspond to steps S12 and S13 in Figure 2, respectively. Furthermore, the host agent 11m applies the PageRank correction described above to the calculated identity score and stores it (S14). In the example in Figure 5, the identity score value is corrected by PageRank, but the original identity score may be used as is without correction by PageRank, or conversely, the agent 11 may be selected using only PageRank without using the identity score.

[0068] Figure 6 is a sequence diagram illustrating an example of the process for extending the agent network 2 in one embodiment of the present invention. This example shows a case where user authorization is not required when the host agent 11m connects with a new agent 11x.

[0069] First, the user (user agent 3) requests a task from the host agent 11m (for example, "create a video with audio" in the use case example above) (S21). For example, the request can be made by entering text via the chat interface. The host agent 11m determines whether the requested task can be executed by itself and known agents 11n connected to it. If it cannot be executed, it identifies the capabilities that each known agent 11n lacks and outputs the requirements for the necessary agent 11 (S22). Then, it requests the agent search unit 30 to search for an agent 11 that meets the requirements (S23).

[0070] The agent search unit 30 searches for agents 11 that meet the requirements from among the agents 11 that are database-driven or searchable via a P2P protocol, and responds with the search results (S24). The search results may include multiple agents 11, which are sorted in descending order of their identity score (which may be corrected by PageRank or not) or PageRank score. The results may also be based on a score evaluated by a third party using a predetermined method. These may include unknown or new agents 11 that have no connection to the host agent 11m. These identity scores are not calculated by the host agent 11m, but are calculated by other agents 11 through the pre-preparation process shown in Figure 6 above.

[0071] Subsequently, the host agent 11m retrieves information on the references of the top n agents 11 from the search results and registers them in its agent list (S25). Basically, n=1, and the one with the highest score is registered as the new agent 11x, but multiple agents 11 may be registered if n>1. Then, the input from the user agent 3 (for example, the execution of the task "create a video with audio") is sent to the newly connected agent 11x (S26), and the new agent 11x executes the requested task (S27). Through this process, the agent network 2 can be autonomously expanded.

[0072] Figure 7 is a sequence diagram outlining another example of the process for extending the agent network 2 in one embodiment of the present invention. Here, similar to the example in Figure 6 described above, an example is shown where user authorization is required when the host agent 11m newly connects with a new agent 11x. Note that the processes from steps S21 to S24 in the example in Figure 6 are the same and are therefore not illustrated or described.

[0073] The host agent 11m can only collaborate with and enlist the help of agents 11 that are registered in its agent list and are known to it, in order to process tasks. If an unauthorized agent 11 is registered in this agent list, threats such as information leakage or inaccurate or inappropriate results being output, which could affect the progress of tasks, may arise. In contrast, in the example in Figure 7, when registering new agents 11x to the agent list, only those that have been "authorized" by the user are registered, thereby preventing unauthorized agents 11 from infiltrating the agent list.

[0074] The host agent 11m obtains the search results for agents 11 that meet the requirements from the agent search unit 30, then identifies n (basically n=1) from the top of the search results and obtains the reference information for these new agents 11x (S31). Then, it requests the authorization unit 20 to register the new agents 11x (S32).

[0075] The authorization unit 20 sends an authorization request to the user (user agent 3) (S33), and when user agent 3 grants authorization (S34), it responds to the host agent 11m granting permission to register the new agent 11x (S35). It is assumed that the user has already been authenticated by the authorization unit 20 before the processing in step S33. Furthermore, the authorization in step S34 can be triggered, for example, by the user pressing the OK button.

[0076] Upon receiving permission from the authorization unit 20, the host agent 11m registers the information of the new agent 11x in its agent list (S36) and sends the input from user agent 3 (for example, the task "Create a video with audio") to the new agent 11x (S37). The new agent 11x then executes the requested task (S38). Through this process, the agent network 2 can be autonomously expanded in a more secure manner.

[0077] As described above, according to the agent system 1, which is one embodiment of the present invention, the identity of each agent 11 over time can be scored and understood, making it possible to select and process a more appropriate agent 11 when processing a task. Furthermore, by correcting the identity score with "PageRank", the reliability of the identity score calculated through peer evaluation and mutual evaluation among agents 11 can be further improved. And by using such a mechanism, for example, it is possible to realize a mechanism that searches for and selects new agents necessary for processing a task, adds them to the agent list, and autonomously expands the agent network 2.

[0078] The present inventors have described the invention in detail based on embodiments above, but it goes without saying that the present invention is not limited to the above embodiments and can be modified in various ways without departing from its essence. Furthermore, the above embodiments are described in detail for the purpose of explaining the present invention in an easy-to-understand manner and are not necessarily limited to those having all the configurations described. In addition, it is possible to add, delete, or replace some of the configurations of the above embodiments with other configurations.

[0079] Furthermore, each of the above configurations, functions, processing units, and processing means may be implemented in hardware, in whole or in part, for example, by designing them as integrated circuits. Alternatively, each of the above configurations, functions, and means may be implemented in software by having the processor interpret and execute programs that implement each function. Information such as programs, tables, and files that implement each function can be stored in memory, hard disks, SSDs, or other recording devices, or in recording media such as IC cards, SD cards, or DVDs.

[0080] Furthermore, in the diagrams above, the control lines and information lines shown are those deemed necessary for explanation and do not necessarily represent all control lines and information lines that would be present in the actual implementation. In reality, it can be assumed that almost all components are interconnected. [Industrial applicability]

[0081] This invention can be used in development support systems that assist development work using generative AI. [Explanation of symbols]

[0082] 1... Agent system, 2... Agent network, 3... User agent 10...Node, 10a...Node A, 10b...Node B, 10c...Node C, 10d...Node D, 10e...Node E, 10f...Node F, 10g...Node G, 10h...Node H 11, 11a, 11b, 11c... Agents, 11m... Host Agent, 11n... Known Agent, 11x... New Agent 12, 12a, 12b, 12d... evaluators, 20...Authorization Department, 30...Agent Search Department

Claims

1. An agent system in which multiple agents, including those that utilize generative AI, constitute nodes in an agent network, and each of the agents cooperates with each other to execute processing requested by a user, Each node includes an evaluator for each agent. An agent system comprising: a first evaluator for each first agent among the agents, which aggregates the first scores relating to the similarity of the outputs calculated for each of the other second agents using a predetermined first method and transmits them to each second evaluator for each of the second agents; a second score transmitted from each second evaluator, which is calculated by each second evaluator using the first method; and an identity score for each agent based on the first and second scores for each agent.

2. In the agent system described in claim 1, The first evaluator is an agent system that corrects the identity score relating to the first agent based on an index calculated by a predetermined third method based on the number of identity scores the first agent has received from each of the other second evaluators.

3. In the agent system described in claim 1, An agent system including the node consisting solely of the evaluator, which also functions as an agent.

4. In the agent system described in claim 1, Furthermore, the agent search unit has an agent search unit that searches for each of the agents included in the agent system. The first agent is an agent system that adds a predetermined number of reference destinations related to a third agent to a list of agents with which it can collaborate, in order of the identity score in the search results obtained by the agent search unit.

5. In the agent system according to claim 4, Furthermore, it has an authorization unit that accepts authorization requests from users, The first agent is an agent system that, when adding a reference to the third agent to the list, performs authorization processing by the user via the authorization unit.