Standard Communication Protocol System and Method for Multi-AI Agent Comprising Integrated Ambiguity Blocking, Controlled Language Evolution, and Semantic-Based Urgency Correction

Abstract

The present invention relates to a communication protocol system and method between multiple AI agents, and discloses a system that fundamentally blocks ambiguity in structured messages exchanged between agents, controls the evolution of communication language under human approval, and suppresses the abuse of urgency at the semantic field level. The system includes a zero ambiguity gate (110) that blocks the transmission of messages with an integrated parsing path count not of 1 by combining structured grammar-based syntax parsing and entity-attribute schema-based semantic verification; a controlled evolution engine (120) that generates abbreviation candidates for repetitive expressions, registers them in a global standard dictionary only after information equivalence verification and human approval, and automatically rolls them back when the misinterpretation rate exceeds a threshold; and an urgency inflation suppressor (130) that applies a cumulative decay discount factor to the urgency semantic field value and directly inputs it into the task priority determination logic of the receiving agent when the high-urgency message ratio per agent exceeds a threshold. By doing so, ambiguity-free communication, language evolution under human control, and prevention of urgency inflation are simultaneously achieved in a multi-AI agent environment.

Description

Technology Field

[0001] The present invention relates to communication between multiple AI agents, and more specifically, to a communication protocol system and method that enforces syntax-semantic integration ambiguity verification for structured messages as a transmission condition, controls language evolution between agents through human approval and automatic rollback mechanisms, and corrects the urgency semantic field value within the message to directly reflect it in the task priority determination of the receiving agent. Background Technology

[0002] A multi-AI agent system is a paradigm in which multiple AI models collaborate to handle complex tasks that are difficult to perform with a single model. In such systems, communication between agents is a core function, and various approaches have been proposed in conventional technology.

[0003] FIPA ACL (Foundation for Intelligent Physical Agents Agent Communication Language) provides a standard message structure including message parameters such as performative, sender, receiver, content, ontology, and protocol, while KQML (Knowledge Query and Manipulation Language) provides a three-layer message structure consisting of content, message, and communication layers, as well as a performative-based keyword-value pair format. Google's A2A (Agent-to-Agent) protocol provides data models such as Task, Message, Part, and Artifact, along with JSON-RPC, gRPC, and REST bindings. AutoGen uses a JSON-based message wrapper in the form of {"role": ..., "content": ...}, and LangChain uses the tool, tool_input, and log field structures of AgentAction.

[0004] However, the above-mentioned conventional technologies have the following common limitations.

[0005] first, Neglect of semantic ambiguity Conventional protocols only check the well-formedness of a message and do not verify the possibility that the same message may be interpreted differently by receiving agents. Since FIPA ACL and KQML allow natural language or arbitrary content language in the content field, semantic ambiguity is transmitted as is even if the message syntax is valid. The A2A protocol also does not include a mechanism to verify the singular semantic interpretation of the Part's content.

[0006] Second, Non-controllable change in communication language As agent systems operate for a long period, abbreviations or new terms may naturally arise between agents, but conventional protocols do not provide a mechanism to manage changes in the language itself. The vocabulary of FIPA ACL is fixed, and while A2A allows extensions, it does not include verification of semantic equivalence of extensions or rollback procedures in the event of side effects.

[0007] Third, Lack of response to abuse of urgency In a multi-agent environment, if a specific agent sends all messages with high urgency, the task scheduling of the receiving agent becomes distorted. Conventional Quality of Service (QoS) technologies limit network bandwidth or packet transmission rates, and no technology has been provided to correct the semantic urgency field values ​​within the message itself and reflect them in the receiving agent's decision-making. General reputation systems or priority normalization are also limited to control at the sender or transmission level and do not implement an end-to-end chain that modifies the receiving side's behavior by directly correcting the semantic fields within the message. The problem to be solved

[0010] In order to overcome the limitations of the prior art described above, the present invention aims to solve the following three problems.

[0011] The first task is to provide a communication gate that performs ambiguity verification by integrating syntax parsing and ontology-based semantic interpretation for structured messages, and allows the transmission of only messages that can be interpreted by a single agent.

[0012] The second task is to allow the abbreviation of repetitive expressions between agents, while controlling language evolution through closed-loop governance that includes semantic equivalence verification, human approval, monitoring, and automatic rollback.

[0013] The third task is to detect abuse of urgency by agent and to correct the urgency semantic field value within the message itself so that it is directly reflected in the receiving agent's task priority determination. means of solving the problem

[0015] To solve the above problem, the present invention provides a communication protocol system between multiple AI agents comprising (a) a zero ambiguity gate, (b) a controlled evolution engine, and (c) an urgency inflation suppressor.

[0016] (a) Zero Ambiguity Gate ...is an integrated parsing path number for a structured message, including the number of syntax parsing paths based on formal grammar and the number of semantic parsing paths based on a pre-registered entity-attribute schema (hereinafter "ontology schema"). Nparse Calculates. The above Nparse If is not 1, the transmission of the message is blocked and a clarification request is returned to the sending agent. This prevents the transmission of a message that is syntactically valid but semantically open to multiple interpretations.

[0017] (b) Controlled evolution engine It generates reduction candidates from representations repeatedly used between agents, and the information equivalence score between the original representation and the reduction candidate Eqinfo Calculates. Eqinfo It is registered as a subject for human operator approval only if it meets or exceeds a predetermined information equivalence threshold. Prior to approval, the above abbreviation candidate is allowed only as a local alias for the sending agent, and after human operator approval, it is registered in the global standard dictionary. A predetermined grace period after registration. Tgrace If the misinterpretation rate of the receiving agent exceeds a predetermined standard, the corresponding abbreviation candidate is automatically discarded and rolled back to the original expression.

[0018] (c) Urgency inflation suppressor is the high-urgent message generation rate by agent δurg Calculate, and the above ratio is a predetermined threshold θurg For agents exceeding, a discount factor on the urgency semantic field value inside the subsequent message ? The above discounted urgency semantic field value is directly input into the task priority determination logic of the receiving agent. By doing so, the abuse of urgency is suppressed through the correction of the message semantics itself, rather than through the limitation of the network transmission rate. Effects of the invention

[0020] According to the present invention, the following effects are achieved.

[0021] First, the transmission of syntactically valid but semantically ambiguous messages is blocked at the source through integrated syntax-semantic verification, thereby preventing malfunctions caused by misunderstandings between agents. Unlike conventional protocols that check only format validity, only messages that can be interpreted by a single agent are allowed to be transmitted.

[0022] Second, the natural evolution of the inter-agent communication language is managed in a controllable manner through closed-loop governance of local aliases → human approval → global dictionary promotion → monitoring → automatic rollback. This simultaneously resolves the rigidity of fixed vocabulary (FIPA ACL) and the risks of uncontrolled extension (general protocols).

[0023] Third, by directly correcting the urgency semantic field value within the message and reflecting it in the receiving agent's decision-making, prevention of urgency abuse at the application layer, distinct from network-level QoS, is achieved.

[0024] Fourth, the above configurations can be implemented on all computing platforms, such as CPUs, GPUs, quantum computing, neuromorphic chips, and optical computing, and are not dependent on specific hardware.

[0025] Fifth, the above configurations can be utilized for automatic term translation between domains by linking with an interdisciplinary knowledge mapping system. Brief explanation of the drawing

[0027] Fig. 1 : Overall system architecture diagram? Deployment and data flow of zero ambiguity gate (110), controlled evolution engine (120), and urgency inflation suppressor (130) Fig. 2 : Integrated parsing process of the Zero Ambiguity Gate? Syntax parsing (210), ontology schema verification (220), Nparse Output (230), transmission / blocking branch (240) Fig. 3 : Closed-loop governance of a controlled evolutionary engine? Reduced candidate generation (310), Eqinfo Output (320), local alias registration (330), human approval (340), global promotion (350), monitoring (360), rollback (370) Fig. 4 : Operation flow of the urgency inflation suppressor? δurg Output (410), threshold comparison (420), discount factor application (430), receiving agent priority determination input (440) Fig. 5 : Example of Ontology Schema Structure ? Entity (MOTOR-03), Attribute (temperature), Type (NUMBER), Unit (°C), Allowable Range (0~120), Domain Constraint Fig. 6 : The 5-Stage Language Development Model of a Controlled Evolutionary Engine? Stage 1: Primitive Communicator, Stage 2: Contraction Generator, Stage 3: Compound Representer, Stage 4: Abstract Conceptor, Stage 5: Metalinguistic Fig. 7: Message Token Structure Comparison? Comparison of Information Density between L2L Structured Messages and Natural Language Messages Specific details for implementing the invention

[0028] delete

[0029] delete

[0030] delete

[0031] delete

[0032] delete

[0033] delete

[0034] delete

[0035] delete

[0036] delete

[0037] delete

[0038] delete

[0039] delete

[0040] delete

[0041] delete

[0042] delete

[0043] delete

[0044] delete

[0045] delete

[0046] delete

[0047] delete

[0048] delete

[0049] delete

[0050] delete

[0051] delete

[0052] delete

[0053] delete

[0054] delete

[0055] delete

[0056] delete

[0057] delete

[0058] delete

[0059] delete

[0060] delete

[0061] delete

[0062] delete

[0063] delete

[0064] delete

[0065] delete

[0066] delete

[0067] delete

[0068] delete

[0069] delete

[0070] delete

[0071] delete

[0072] delete

[0073] delete

[0074] delete

[0075] delete

[0076] delete

[0077] delete

[0078] delete

[0079] delete

[0080] delete

[0081] delete

[0082] delete

[0083] delete

[0084] delete

[0085] delete

[0086] delete

[0087] delete

[0088] delete

[0089] delete

[0090] delete

[0091] delete

[0092] delete

[0093] delete

[0094] delete

[0095] delete

[0096] delete

[0097] delete

[0098] delete

[0099] delete

[0100] delete

[0101] delete

[0102] delete

[0103] delete

[0104] delete

[0105] delete

[0106] delete

[0107] delete

[0108] delete

[0109] delete

[0110] delete

[0111] delete

[0112] delete

[0113] delete

[0114] delete

[0115] delete

[0116] delete

[0117] delete

[0118] delete

[0119] delete

[0120] delete

[0121] delete

[0122] delete

[0123] §5.1 System OverviewThe system (100) of the present invention includes a processor and a memory that stores instructions executed by the processor. The processor executes a zero ambiguity gate (110), a controlled evolution engine (120), and an urgency inflation suppressor (130). The structured message processed in the present invention is generated according to a predetermined formal grammar and may consist of tokens classified into six parts of speech: entity, action, value, relation, meta, and pattern. However, the number and types of parts of speech are not limited thereto, and any structured format necessary for communication between agents may be used. The structured message format itself is not a core component of the present invention, but functions as a base structure for input data to operate the three core components (110, 120, 130). A communication protocol method between multiple AI agents according to one embodiment of the present invention may be implemented in the form of program instructions executable on a computer, and said program instructions may be recorded on a computer-readable recording medium. In addition, in one embodiment of the present invention, the processor may perform a preprocessing step of receiving a natural language message in an unstructured format and extracting an entity identifier, attribute name, value, and urgency field from the natural language message and converting it into a structured message according to a predetermined formal grammar before calculating the number of integrated parsing paths Nparse for the structured message. Processing of a zero ambiguity gate (110), a controlled evolution engine (120), and an urgency inflation suppressor (130) may be performed on the converted structured message. §5.2 Variable DefinitionsThe main variables and symbols used in the present invention are defined below. Nparse refers to the number of integrated parsing paths, including the number of syntax parsing paths and the number of semantic interpretation paths, and has an integer value of 1 or greater. The value allowed in the present invention is 1. Eqinfo refers to the information equivalence score between a contraction candidate and an original representation. It is calculated as the degree of agreement between the set of semantic elements of the original representation and the set of semantic elements of the contraction candidate, and has a dimensionless value between 0 and 1. Specific calculation methods may include set-based similarity (Jaccard et al.), embedding vector-based similarity (cosine similarity, etc.), or a combination thereof. In a preferred embodiment, the threshold value is 0.99 and can be adjusted within a range of 0.95 or higher depending on the safety requirements of the application field. δurg represents the ratio of high-urgent messages within a specified period for a specific agent and has a dimensionless value between 0 and 1. θurg represents the urgency inflation trigger threshold, and the default value in a preferred embodiment is 0.3. ? represents the urgency discount factor and has a dimensionless value greater than 0 and less than 1. The default value in a preferred embodiment is 0.5. k represents the number of consecutive violations for a specific agent and has an integer value greater than or equal to 0. Tgrace represents the rollback grace period after global pre-registration and has a positive time value. The default value in a preferred embodiment is 72 hours. Treset represents the number of consecutive normal windows for urgency discount initialization and has an integer value greater than or equal to 1. The default value in a preferred embodiment is 10. Ssafety represents the security rating score of a message and is calculated as a weighted sum of the presence or absence of a digital signature, the encryption level, and the sender authentication status. It has a dimensionless value between 0 and 1. Sefficiency refers to the transmission efficiency score and is calculated as the weighted sum of the inverse of the number of tokens and information density. It has a dimensionless value between 0 and 1. Saudit refers to the human auditability score and is calculated as the natural language translatability ratio.It has a dimensionless value between 0 and 1. ρinfo represents information density and is the value obtained by dividing the number of semantic tokens by the total number of tokens. It has a dimensionless value between 0 and 1, and the target value in a preferred embodiment is 0.9 or higher. Pevolution represents the language evolution readiness score and has a dimensionless value of 0 or higher. The threshold value in a preferred embodiment is 1.0. dshared represents the number of dimensions of the shared embedding space and has an integer value of 768 or higher. dmodel represents the number of dimensions of the individual LLM embedding space and has an integer value of 1 or higher. All of the above default values ​​are values ​​in a preferred embodiment and can be adjusted depending on the application field. §5.3 Formal Grammar The structured message of the present invention is generated according to the following BNF (Backus-Naur Form) formal grammar. <message> ::= <sender> '→' <receiver> <action> <target> <value> <meta> ? <relation> ? <sender>::= ENTITY_ID <receiver>::= ENTITY_ID <action>::= 'REQ' | 'RPT' | 'CFM' | 'DNY' | 'QRY' | 'UPD' | 'DEL' | 'SUB' <target>::= ENTITY_ID '.'ATTRIBUTE <value>::= NUMBER UNIT | STRING | BOOLEAN | VECTOR_REF <meta> ::= '['TIMESTAMP ','PRIORITY ','SESSION_ID ']' <relation> ::= '|' <message>ENTITY_ID ::= [AZ][A-Z0-9-]*ATTRIBUTE ::= [az][a-z0-9-]*UNIT ::= '°C' | '%' | 'h' | 'rpm' | 'V' | 'A' | 's' | 'ms'NUMBER ::= [0-9]+('.'[0-9]+)?STRING ::= '"' [^"]* '"'BOOLEAN ::= 'T' | 'F'VECTOR_REF ::= '#' [A-F0-9]{32}TIMESTAMP ::= [0-9]{14}PRIORITY ::= [0-9]('.'[0-9]+)?SESSION_ID ::= [A-Z0-9]{8}-[A-Z0-9]{4}-[A-Z0-9]{4}-[A-Z0-9]{4}-[A-Z0-9]{12} The above grammar has all production rules that are deterministic, has no ε-production rules, and the first set of each non-terminal symbol is mutually exclusive. Therefore, LL(1) parsing is possible, and the number of syntactic parsing paths is always 1. However, the present invention Since the zero ambiguity gate (110) verifies semantic singleness based on an ontology schema in addition to syntactic singleness, the function of the zero ambiguity gate is not achieved by the determinism of the grammar alone. This is explained in detail below in §5.4. The above BNF grammar is a preferred embodiment, and the present invention is not limited to the above specific grammar. §5.4 Zero Ambiguity Gate (110) The zero ambiguity gate (110) includes a syntax parser (111) and a semantic verifier (112). The syntax parser (111) parses the received structured message according to the formal grammar and checks for syntactic validity. Syntactically invalid messages are immediately blocked and a format error message is returned to the sending agent. For syntactically valid messages, the number of syntax parsing paths is calculated. When using the grammar of §5.3, the number of syntax parsing paths is always 1. For syntactically valid messages, the semantic verifier (112) verifies semantic singularity by referring to a pre-registered ontology schema. The ontology schema defines data types, units, allowed value ranges, and domain constraints for each attribute of an entity. The ontology schema can be registered in JSON-LD, OWL, or a custom format, and the minimum required fields for registration include the entity identifier, attribute name, data type, unit, and allowed range. The specific verification procedure is described below. For example, assume that the following is registered in the ontology schema: MOTOR-03: temperature: type:NUMBER unit:°C range:[0, 120] domain:mechanical_engineering. In this case, for the message ROBOT-01→CTRL-01 REQ MOTOR-03.temperature 85°C, the semantic verifier (112) checks the following items sequentially. First, it checks whether MOTOR-03 is a registered entity. Since it is registered, it passes. Next, it checks whether temperature is a registered attribute of MOTOR-03. Since it is registered, it passes. Next, it checks whether 85°C is within the allowed range [0, 120]. Since it is within the range, it passes. Next, it checks whether the unit of the value matches the unit defined in the schema. It passes because it matches in °C. Finally, check if there is only one analysis that satisfies all the above conditions. Since there is a single analysis, it passes.As a result of the above verification, if the number of semantic interpretation paths is 1, it is combined with the number of syntax parsing paths to calculate the number of integrated parsing paths Nparse=1, and the message is transmitted. On the other hand, cases may occur where the number of semantic interpretation paths is 2 or more. For example, if the same attribute name temperature is registered with an allowed range [0, 120]°C in the mechanical_engineering domain and with an allowed range [0, 500]°C in the chemical_engineering domain, and the message sender does not specify the domain, the value 85°C is valid in both domains, so the number of semantic interpretation paths becomes 2. In this case, Nparse is calculated as 2, so transmission is blocked, and a domain clarification request message is generated and returned to the sending agent. In the case of a message containing an unregistered entity-attribute combination, a message containing a value outside the allowed range, or a message with a type conflict, Nparse is calculated as 0 and transmission is blocked because there is no valid semantic interpretation path. The ontology schema referenced by the semantic verifier (112) is pre-registered by the system operator. The schema of a newly promoted abbreviation candidate by the control-type evolution engine (120) described later is also automatically added to the ontology schema during global pre-registration. §5.5 Controlled Evolution Engine (120) The controlled evolution engine (120) includes a shortening candidate generator (121), an information equivalence verifier (122), a local alias manager (123), a human approval gate (124), a global pre-manager (125), a monitoring engine (126), and a rollback engine (127). The shortening candidate generator (121) analyzes communication logs between agents to detect expressions that are used repeatedly more than a predetermined frequency and automatically generates shortening candidates. The information equivalence verifier (122) calculates the degree of agreement between the set of semantic elements of the original expression and the set of semantic elements of the shortening candidate to generate an information equivalence score Eqinfo. Set-based similarity may be used for calculating the degree of agreement. For example, the value obtained by dividing the size of the intersection of the original semantic element set and the set of shortening candidate semantic elements by the size of the union may be used. Alternatively, embedding vector-based similarity may be used, for example, the cosine similarity between the original embedding and the shortening candidate embedding may be used. A weighted combination of both is also possible. If Eqinfo is below a predetermined information equivalence threshold, the corresponding abbreviation candidate is discarded. In a preferred embodiment, the threshold is 0.99. The technical basis for the threshold of 0.99 is as follows. In a multi-agent simulation environment at a manufacturing site, abbreviation candidates were generated for multiple L2L message pairs (more than 100 in one embodiment), and the interpretation results of the receiving agent were compared with the interpretation results of the original message. Semantic distortion was defined as a case where the action decision of the receiving agent changes compared to the original. When Eqinfo ≥ 0.90, the semantic distortion rate was observed to be approximately 15%, decreased to approximately 8% when Eqinfo ≥ 0.95, and decreased to approximately 0.3% when Eqinfo ≥ 0.99. Since the present invention aims for error-free communication between agents, 0.99, which achieves a semantic distortion rate of less than 1%, was set as the preferred default value. However, it can be adjusted within a range of 0.95 or higher depending on the safety requirements of the industry.The local alias manager (123) registers abbreviation candidates whose Eqinfo is above a threshold as targets for human approval, but manages the abbreviation candidates as local aliases usable only within the local session or local task scope of the sending agent until approval. Local aliases are not transmitted to other agents. The human approval gate (124) presents the abbreviation candidates, original representations, and Eqinfo scores to the human operator and receives an approval or rejection input. The step involving humans is limited to this gate, and humans do not intervene in real-time message transmission and verification. The global dictionary manager (125) registers the abbreviation candidates approved by the human operator in the global standard dictionary. Upon registration, the creation timestamp, the approving human operator identifier, the previous representation, and the rollback pointer are recorded together. The schema information of the registered abbreviation candidates is automatically reflected in the ontology schema of the zero ambiguity gate (110). The monitoring engine (126) monitors the usage results of the abbreviation candidates registered in the global dictionary for a predetermined grace period Tgrace. In a preferred embodiment, Tgrace is 72 hours. Monitoring is performed by calculating the misinterpretation rate by comparing the interpretation result of the receiving agent after applying the abbreviation candidate with the standard interpretation result when using the original expression. If the misinterpretation rate calculated by the monitoring engine (126) exceeds a predetermined standard, the rollback engine (127) automatically discards the corresponding abbreviation candidate from the global dictionary and returns to the previous expression by referring to the rollback pointer. In a preferred embodiment, the misinterpretation rate standard is 1%. The schema information of the discarded abbreviation candidate is also automatically removed from the ontology schema. The controlled evolution engine (120) can manage language evolution in the following five stages. The first stage is the raw communicator stage, in which only basic messages according to the formal grammar are used. The second stage is the abbreviation generator stage, in which the abbreviation of the repetitive expression begins. The third stage is the composite expresser stage, in which multiple messages are combined into a single composite message.Stage 4 is the abstract concept stage, where higher-level concepts abstracting concrete entities emerge. Stage 5 is the meta-language stage, where meta-level representations of the language itself become possible. Different levels of human approval are required for each stage, and if the language evolution readiness score Pevolution ≥ 1.0, it is registered as a candidate for the next stage. §5.6 Urgency Inflation Suppressor (130) The urgency inflation suppressor (130) includes a ratio calculator (131), a discount applyer (132), and a reset manager (133). The ratio calculator (131) calculates the ratio δ_urg of messages marked as high-urgent among all messages sent by a specific agent within a predetermined period (window). Specifically, δ_urg is the value obtained by dividing the number of high-urgent messages N_high_urg of the agent within the predetermined period by the total number of messages N_total of the agent within the same period. That is, δ_urg(i) = N_high_urg(i, Δt) / N_total(i, Δt). Here, δ_urg(i) is the ratio of high-urgent messages of agent i, N_high_urg(i, Δt) is the number of high-urgent messages of agent i within window Δt, and N_total(i, Δt) is the total number of messages of agent i within the same period. The discount applicator (132) applies a discount factor φ to the value of the urgency semantic field within a subsequent message for agents whose δ_urg is greater than or equal to a predetermined threshold θ_urg. In a preferred embodiment, θ_urg is 0.3 and φ is 0.5. If the agent violates consecutively, the discounted urgency is calculated by multiplying the original urgency by φ raised to the power of k. That is, discounted urgency = original urgency × φ^k, where k is the number of consecutive violations. For example, if the original urgency is 0.9, k=1 for the first violation, so φ^1 = 0.5 is applied and the discounted urgency is 0.45; k=2 for the second consecutive violation, so φ^2 = 0.25 is applied and the discounted urgency is 0.225; and k=3 for the third consecutive violation, so φ^3 = 0.125 is applied and the discounted urgency is 0.1125. The above discounted urgency meaning field value is directly input into the receiving agent's task selection, processing order determination, or alarm priority calculation logic.Accordingly, the urgency inflation suppressor (130) implements an end-to-end semantic correction chain that modifies the behavior of the receiving agent by correcting the semantic field value itself within the message, unlike conventional QoS technology that limits the network transmission rate. The reset manager (133) resets the discount factor to the initial state (k = 0) if the state in which the agent's δ_urg is less than θ_urg persists for a T_reset continuous window. In a preferred embodiment, T_reset is 10. §5.7 Layer 3 Communication Management (Optional Configuration) The system (100) of the present invention may optionally further include a three-layer communication management module (140). The module calculates a layer selection score L_select based on the communication safety importance S_safety, efficiency requirement S_efficiency, and audit requirement S_audit. That is, L_select = w_s × S_safety + w_e × S_efficiency + w_a × S_audit. Here, w_s, w_e, and w_a are each positive weights, and w_s + w_e + w_a = 1. In a preferred embodiment, w_s = 0.5, w_e = 0.3, and w_a = 0.2. Based on the score, one or more of the following three layers are selected. The L1 layer is a vector direct exchange layer that directly exchanges embedding vectors between agents. The L2 layer is a structured token layer that exchanges L2L messages according to the structured grammar of §5.3. The L3 layer is a natural language translation layer that translates L2L messages into human-readable natural language and exchanges them. When S_audit is above a certain threshold (0.8 in a preferred embodiment), the L3 layer is automatically transmitted regardless of the selected layer to enable human auditing. §5.8 Embedding Bridge (Example of L1 Layer) Vector exchange between LLMs with embedding spaces of different dimensions in the L1 layer can be performed through a shared projection layer. Each LLM i maps its own embedding vector v_i to the shared space through the shared projection matrix W_proj(i). That is, v_shared = W_proj(i) · v_i. Here, v_i is the self-embedding vector of LLM i, a real vector of d_model(i) dimensions; W_proj(i) is the shared projection matrix of d_shared rows × d_model(i) columns; and v_shared is a shared space vector of d_shared dimensions. The above W_proj(i) is pre-trained on an L2L corpus using contrastive learning (e.g., InfoNCE loss). When a new LLM j is added, the existing W_proj is fixed, and only the lightweight adapter ΔW(j), which has fewer than 1 million parameters, is fine-tuned to align it to the existing shared space. That is, v_shared = ( W_proj(base) + ΔW(j) ) · v_j. Here, W_proj(base) is the existing fixed shared projection matrix, ΔW(j) is a lightweight adapter dedicated to the new LLM j with fewer than 1 million parameters, and v_j is the embedding vector of the new LLM j. The above embedding bridge is a preferred embodiment of the L1 layer and is not strictly necessary for the operation of the core components (110, 120, 130) of the present invention. Explanation of the symbols

[0124] 100 ? Communication protocol system between multiple AI agents 110 ? Zero Ambiguity Gate 111 ? syntax parser 112 ? Semantic verifier 120 ? Controlled evolution engine 121 ? Shortening candidate generator 122 ? Information Equivalence Verifier 123 ? Local Alias ​​Manager 124 ? Human Approval Gate 125 ? Global Dictionary Manager 126 ? Monitoring engine 127 ? Rollback engine 130 ? Urgency Inflation Suppressor 131 ? Ratio Calculator 132 ? Discount Application 133 ? Reset Manager 140 ? Layer 3 Communication Management Module (Optional Configuration)< / message> < / relation> < / value> < / target> < / action> < / receiver> < / sender> < / relation> < / value> < / target> < / action> < / receiver> < / sender> < / message>

Claims

Claim 1 A communication protocol system between multiple AI agents comprises: a processor; and a memory for storing instructions executed by the processor; wherein the processor comprises: (a) a zero ambiguity gate (110) that calculates, for a structured message, an integrated parsing path number Nparse, which includes the number of syntax parsing paths based on formal grammar and the number of semantic interpretation paths based on a pre-registered entity-attribute schema and type constraints, and blocks the transmission of the message if Nparse is not 1; (b) a controlled evolution engine (120) that generates a shortened candidate from an expression repeatedly used between agents, calculates an information equivalence score between the original expression and the shortened candidate, registers it as a subject for human operator approval only if the calculated information equivalence score is greater than or equal to a predetermined information equivalence threshold, allows the shortened candidate only as a local alias of the sending agent before approval, registers it in a global standard dictionary after human operator approval, and automatically discards the shortened candidate and rolls it back to the original expression if the misinterpretation rate of the receiving agent exceeds a predetermined standard within a predetermined grace period after registration. and (c) an emergency inflation suppressor (130) characterized by calculating a high-urgent message generation rate δurg per agent, applying a discount factor to the emergency semantic field value within a subsequent message for agents whose rate exceeds a predetermined threshold, and directly inputting the discounted emergency semantic field value into the task priority determination logic of the receiving agent; wherein the schema information of a shortened candidate newly registered or rolled back in the global standard dictionary by the control-type evolution engine (120) is automatically reflected in the entity-attribute schema of the zero ambiguity gate (110), and the emergency semantic field value of a message blocked by the zero ambiguity gate (110) is excluded from the rate calculation of the emergency inflation suppressor (130). Claim 2 In paragraph 1, the number of integrated parsing paths Nparse A system characterized by being produced by combining a syntax parsing result according to the formal grammar of a structured message and a semantic verification result regarding whether the entity identifier, attribute name, data type, unit, allowed range, and domain constraint of the value included in the message are compatible with a pre-registered entity-attribute schema. Claim 3 A system according to claim 1, wherein the zero ambiguity gate (110) blocks transmission and generates a clarification request message to the sending agent when different semantic interpretations are possible in multiple domains for the same message, when an unregistered entity-attribute combination is included, when a value outside the allowed range is included, or when conflicting type constraints are detected. Claim 4 In paragraph 1, the controlled evolution engine (120) records together a creation timestamp, an approving human operator identifier, a previous representation, and a rollback pointer when a shortened candidate is approved by a human operator and registered in a global standard dictionary, and a predetermined grace period after registration. Tgrace A system characterized by calculating a misinterpretation rate by comparing the interpretation result of a receiving agent based on the use of the above-mentioned abbreviation candidate with the standard interpretation result when using the original expression, and automatically returning to the existing expression by referring to the above-mentioned rollback pointer if the misinterpretation rate exceeds a predetermined standard. Claim 5 A system according to claim 1, wherein the information equivalence score is calculated based on the degree of agreement between the set of semantic elements of the original representation and the set of semantic elements of the abbreviation candidate, and the degree of agreement is calculated by set-based similarity, embedding vector-based similarity, or a combination thereof. Claim 6 In paragraph 1, the urgency inflation suppressor (130) determines the discount factor when a specific agent continuously exceeds a predetermined threshold number of consecutive violations k Depending on ?k Accumulate attenuation, and the state in which the high-urgent message ratio of the above agent is below a predetermined threshold Treset A system characterized by resetting the above discount factor to its initial state if it persists for a continuous window. Claim 7 In paragraph 1, the system is of the importance of communication safety Ssafety , efficiency requirements Sefficiency , and the need for audit Saudit Based on, it further includes a communication management module (140) that optionally uses one or more of a vector direct exchange layer, a structured token layer, and a natural language translation layer, and the communication management module (140) Saudit A system characterized by enabling human auditing by automatically accompanying a natural language translation layer when the standard is higher than a certain threshold. Claim 8 delete Claim 9 delete Claim 10 delete Claim 11 delete Claim 12 delete

Images