An agent communication-oriented C-APN semantic message header encapsulation and forwarding system

Abstract

The application discloses an agent communication-oriented C-APN semantic message header encapsulation and forwarding system and relates to the technical field of data communication. The system comprises a message header processing entity and a network device. The message header processing entity receives an agent communication message, acquires agent communication semantic input information, generates a C-APN semantic message header containing a flow identification field and a cognitive semantic field, and encapsulates the C-APN semantic message header outside the original message header. The network device analyzes the message header, obtains the flow identification and the cognitive semantic field, generates a record when no flow forwarding record is queried, reads the record when the record is queried, and performs forwarding processing according to the cognitive semantic field or the flow forwarding record. The application explicitly carries the cognitive semantic requirement in the message in agent communication, enables the network device to identify the communication requirement without deeply analyzing the application layer load, realizes the quick forwarding of subsequent messages of the same interactive flow through the flow identification, and supports the differentiated forwarding and resource guidance of requirements such as Token throughput, context dependence, inference chain position and power coordination.

Description

Technical Field

[0001] This invention relates to the field of data communication technology, and in particular to a C-APN semantic message header encapsulation and forwarding system for intelligent agent communication. Background Technology

[0002] In traditional data communication, data packets typically include the original header and payload. When forwarding data packets, network devices primarily determine the next-hop path based on fields such as source address, destination address, source port, destination port, and protocol type. While these fields can identify ordinary network sessions, they are insufficient to express the semantics of higher-layer tasks.

[0003] With the development of large language models and autonomous intelligent agent technology, multiple intelligent agents can perform planning, retrieval, reasoning, tool invocation, code execution, and result verification around complex tasks. Communication between different intelligent agents is no longer just ordinary data transmission, but also carries information such as task stage, context dependencies, token generation rate, inference chain position, semantic error tolerance, and computing power collaboration requirements.

[0004] If existing network devices forward messages based solely on the original packet headers, they cannot determine whether a particular agent's interaction flow is on a critical inference path, whether the interaction flow depends on a continuous context, or whether the interaction flow needs to be guided to an edge computing node or a specific model service node. If network devices obtain the above information by parsing application layer content, it will increase processing overhead and may be affected by encrypted transmission, protocol diversity, and changes in application payload format.

[0005] Existing semantic forwarding schemes include approaches that involve adding a semantic header to the outer layer of the original packet header and using semantic tags and flow identifiers for forwarding. These schemes typically record the next-hop route when the first packet of a data stream arrives, and subsequent packets are forwarded quickly based on the flow record, thereby reducing redundant parsing of the original header.

[0006] However, existing semantic forwarding schemes are mostly geared towards general network attributes, such as security requirement level, user level, traffic type, or quality of service level. They have not yet established a dedicated message header carrying and forwarding mechanism for the model inference characteristics, context passing characteristics, inference chain position characteristics, and computing power collaboration requirements in agent communication.

[0007] In intelligent agent communication scenarios, even if the upper-layer system can identify communication processing requirements such as task stages, context dependencies, token throughput requirements, inference chain positions, or computing power collaboration requirements, network devices still need a message-level bearer mechanism that can be identified and processed in the data plane. This allows cognitive semantic requirements to be transmitted with communication messages and play a role in processes such as initial message flow establishment, subsequent message table lookup and forwarding, QoS queue selection, network slice selection, and computing power node guidance.

[0008] Therefore, a C-APN semantic header encapsulation and forwarding system for agent communication is needed to convert the cognitive semantic requirements in agent communication into header fields that network devices can parse, and to perform efficient flow identification, fast forwarding and differentiated resource guidance in the data plane. Summary of the Invention

[0009] The purpose of this invention is to provide a C-APN semantic header encapsulation and forwarding system for agent communication to solve the problems of existing network devices having difficulty in recognizing the cognitive semantic needs of agent communication in the data plane, high overhead of repeated parsing, and inability to perform differentiated forwarding based on agent task requirements.

[0010] To achieve the objective of this invention, the technical solution is as follows:

[0011] A C-APN semantic header encapsulation and forwarding system for intelligent agent communication includes a header processing entity and network devices:

[0012] The message header processing entity is used to receive communication messages initiated by the agent, obtain the agent communication semantic input information corresponding to the communication message, generate a flow identifier field and a cognitive semantic field according to the agent communication semantic input information, assemble a C-APN semantic message header according to a preset message header format, and encapsulate the C-APN semantic message header outside the original message header of the communication message to obtain an agent communication message carrying the C-APN semantic message header;

[0013] The network device is used to parse the agent communication message carrying the C-APN semantic message header, obtain the flow identifier field and the cognitive semantic field, generate a flow forwarding record when no flow forwarding record corresponding to the flow identifier field is found, and read the flow forwarding record when a corresponding flow forwarding record is found, and perform forwarding processing according to the cognitive semantic field or the flow forwarding record;

[0014] The C-APN semantic header encapsulation and forwarding system operates according to the following steps:

[0015] S101, Receive communication messages initiated by the intelligent agent;

[0016] S102. Obtain agent communication semantic input information corresponding to the communication message. The agent communication semantic input information includes flow identifier generation information for determining the agent interaction flow to which the communication message belongs, and communication requirement indication information for characterizing the communication processing requirements of the agent interaction flow.

[0017] S103. Generate a C-APN semantic message header based on the agent communication semantic input information. The C-APN semantic message header includes at least an identifier field for indicating that the communication message carries a C-APN semantic message header, a flow identifier field for identifying the agent interaction flow to which the communication message belongs, and a cognitive semantic field for characterizing the cognitive communication needs of the agent interaction flow. Each field is arranged according to a preset message header format.

[0018] S104. Encapsulate the C-APN semantic header outside the original header of the communication message to obtain an agent communication message carrying the C-APN semantic header;

[0019] S105. The network device parses the agent communication message carrying the C-APN semantic message header to obtain the flow identifier field and the cognitive semantic field;

[0020] S106. Perform forwarding processing on the agent communication message according to the flow identifier field and the cognitive semantic field;

[0021] C-APN stands for Cognitive-semantic Agent-oriented Perception Network. The C-APN semantic header is not limited to a specific physical network protocol header.

[0022] The agent communication message includes at least one of the following:

[0023] HTTP request messages, RPC call messages, WebSocket messages, message queue messages, event bus messages, multi-agent cooperation protocol messages, model call request messages, tool call request messages, context synchronization messages, and result verification messages.

[0024] Furthermore, the stream identifier generation information includes at least two of the following: source agent identifier, destination agent identifier, task identifier, session identifier, protocol type, communication direction, interaction stage, message sequence information or time window information;

[0025] The communication requirement indication information includes at least one of the following: task priority, token throughput requirement, context dependency level, semantic error tolerance, inference chain position, latency requirement, service quality requirement, computing power collaboration requirement, model service type, tool call type, data retrieval requirement, or context synchronization status.

[0026] Furthermore, the cognitive semantic field includes at least one of the following:

[0027] The Token throughput field is used to characterize the model's input or output rate requirements;

[0028] Context-dependent fields are used to characterize the continuity requirements of cross-turn dialogues, context windows, shared memory, or intermediate inference states;

[0029] The semantic error tolerance field is used to characterize the degree to which the agent's interaction flow tolerates semantic deviations caused by latency, jitter, out-of-order delivery, packet loss, or missing context.

[0030] The task priority field is used to characterize the importance of the current task in the multi-agent collaboration process;

[0031] The inference chain position field is used to characterize the position or scope of influence of the current communication message in the cooperative inference chain;

[0032] The computing resource requirement field is used to indicate whether the current communication message needs to be guided to a model inference node, edge computing node, tool execution node, or data retrieval node.

[0033] Furthermore, the C-APN semantic header includes an identifier field, a flow identifier field, and a cognitive semantic field, and also includes a header length field, a header type field, and a checksum field; generating the C-APN semantic header includes the following steps:

[0034] The set of fields for the C-APN semantic header is determined according to the preset header format, and the identifier field is generated according to the preset C-APN protocol identifier value;

[0035] The process involves extracting flow identifier generation information from the agent communication semantic input information and generating the flow identifier field based on the flow identifier generation information; extracting communication requirement indication information from the agent communication semantic input information and mapping the communication requirement indication information to one or more cognitive semantic fields according to a preset semantic mapping rule; generating the header length field based on the total length of the field set and generating the header type field based on the encapsulation type or bearer protocol type; writing the identifier field, header type field, header length field, flow identifier field, and cognitive semantic field in the field order specified by the preset header format; calculating the check field based on at least part of the header content excluding the check field; and writing the check field to the corresponding position to obtain the C-APN semantic header.

[0036] The preset semantic mapping rules include at least one of the following:

[0037] Map token throughput requirements to low, medium, or high throughput levels;

[0038] Map the degree of context dependency to a weak dependency level, a normal dependency level, or a strong dependency level;

[0039] Map semantic error tolerance to high tolerance level, medium tolerance level, or low tolerance level;

[0040] Map task priorities to normal task level, priority task level, or critical task level;

[0041] Map the computing power collaboration requirements to no-computing-power guidance, edge inference guidance, model service guidance, tool execution guidance, or data retrieval guidance.

[0042] Further, in step S104, the C-APN semantic header is encapsulated in addition to the original header of the communication message, including:

[0043] The C-APN semantic header is generated in the agent runtime component, proxy gateway, service mesh data plane proxy, network edge device, or computing network scheduling gateway located between the agent application and the network;

[0044] The C-APN semantic message header is added to the communication message as one of the outer extension header, tunnel encapsulation header, transport layer extension header, application layer metadata header, or intelligent agent communication protocol extension header;

[0045] The C-APN semantic header is encapsulated in any of the following locations: between the outer network encapsulation header and the original network layer header; between the transport layer header and the application layer payload; in the metadata field of an HTTP, RPC, WebSocket, message queue, or event bus message; or in the extended protocol header carrying the agent invocation request.

[0046] Furthermore, the step of forwarding the agent communication message based on the flow identifier field and the cognitive semantic field includes the following steps:

[0047] Determine whether the agent communication message is the first message of the agent interaction stream;

[0048] If it is the first message of the intelligent agent interaction flow, the basic next-hop path is obtained according to the original message header of the communication message, the semantic forwarding strategy is determined according to the cognitive semantic field, and a flow forwarding record is generated according to the basic next-hop path and the semantic forwarding strategy.

[0049] If it is a subsequent message of the intelligent agent interaction stream, then query the stream forwarding record according to the stream identifier field, and forward the subsequent message according to the forwarding action in the stream forwarding record;

[0050] The flow forwarding record includes at least one of the following: flow identifier, source agent identifier, destination agent identifier, source address, destination address, source port, destination port, protocol type, next-hop path, QoS queue identifier, network slice identifier, computing node identifier, model service identifier, tool service identifier, data service identifier, cognitive semantic field, forwarding action, record generation time, and record aging time.

[0051] Furthermore, a semantic forwarding strategy is determined based on the cognitive semantic field, including at least one of the following:

[0052] When the Token throughput field indicates a high throughput level, select a high-bandwidth link, increase the queue scheduling weight, or forward to a high-throughput service instance;

[0053] When the context dependency field indicates a strong dependency level and the semantic error tolerance field indicates a low tolerance level, select a low latency, low jitter, or low packet loss path.

[0054] When the task priority field indicates a high-priority task, the agent communication message is mapped to a high-priority QoS queue or a critical task network slice.

[0055] When the inference chain position field indicates that the current message is at a critical inference chain node, the execution path of the agent's communication message is maintained, resources are reserved, or priority is forwarded.

[0056] When the computing power resource requirement field indicates that computing power collaboration is needed, the intelligent agent communication message is directed to the corresponding edge computing power node, model inference node, tool execution node, or data retrieval node.

[0057] The system also includes: when the network device cannot recognize the C-APN semantic header, performing basic forwarding based on the original header of the communication message; or, the network device removing, transparently transmitting, or converting the C-APN semantic header before performing forwarding.

[0058] Further, in step S105, the step of parsing the agent communication message carrying the C-APN semantic header by the network device to obtain the flow identifier field and the cognitive semantic field includes:

[0059] After receiving the agent communication message carrying the C-APN semantic message header, the network device identifies the C-APN semantic message header; when the C-APN semantic message header includes a check field, the integrity of the C-APN semantic message header is checked according to the check field.

[0060] When the C-APN semantic header does not include a verification field or the integrity verification passes, the C-APN semantic header is parsed to obtain the flow identifier field and the cognitive semantic field;

[0061] When the C-APN semantic header includes a verification field and the integrity verification fails, the C-APN semantic header is discarded, basic forwarding is downgraded, or an exception record is generated.

[0062] Furthermore, after generating the flow forwarding record based on the basic next-hop path and the semantic forwarding strategy, the method further includes: setting an aging time for the flow forwarding record; updating the status information or aging time of the flow forwarding record when a subsequent message belonging to the same agent interaction flow is received and the flow forwarding record is found; and deleting the flow forwarding record when the aging time expires and no subsequent message belonging to the same agent interaction flow is received.

[0063] Furthermore, the network device in the system includes: a receiving unit, used to receive an agent communication message carrying a C-APN semantic message header; a parsing unit, used to parse the C-APN semantic message header to obtain a flow identifier field and a cognitive semantic field; and a flow record management unit, used to query flow forwarding records according to the flow identifier field, and generate flow forwarding records according to the flow identifier field, the cognitive semantic field, and the original message header when no corresponding flow forwarding record is found.

[0064] The flow record management unit is also used to read the flow forwarding record when a corresponding flow forwarding record is found, and to update the status information or aging time of the flow forwarding record;

[0065] The forwarding unit is used to perform forwarding processing on the agent communication message according to the cognitive semantic field or the stream forwarding record.

[0066] Compared with the prior art, the significant advantages of this invention are: First, this invention explicitly carries the cognitive semantic requirements in agent communication in the message through the C-APN semantic message header, enabling network devices to identify agent communication requirements without deeply parsing the application layer load.

[0067] Second, by using the flow identifier field and flow forwarding record, this invention enables network devices to avoid repeatedly parsing the original packet header and cognitive semantic field when processing subsequent packets of the same intelligent agent interaction flow, thereby improving forwarding efficiency.

[0068] Third, this invention incorporates the unique requirements of agent communication, such as token throughput, context dependency, semantic error tolerance, inference chain position, and computing power collaboration, into the message header field, enabling the data plane to perform differentiated forwarding tailored to the needs of agent tasks.

[0069] Fourth, this invention can guide agent communication messages to appropriate QoS queues, network slices, forwarding paths or computing nodes, providing a message-level entry point for joint scheduling of computing and network in agent communication scenarios.

[0070] Fifth, this invention supports pass-through, removal, conversion, or basic forwarding in network devices that cannot recognize C-APN semantic headers, and has good protocol compatibility and engineering deployability. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments are briefly described below. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0072] Figure 1 is a schematic flowchart of the C-APN semantic header encapsulation and forwarding method provided in Embodiment 3 of the present invention;

[0073] Figure 2 is a schematic diagram of the intelligent agent communication message structure carrying the C-APN semantic message header provided in Embodiment 3 of the present invention;

[0074] Figure 3 is a schematic diagram of the C-APN semantic message header field structure provided in Embodiment 1 of the present invention;

[0075] Figure 4 is a schematic diagram of the initial message flow establishment record and subsequent message lookup and forwarding process provided in embodiments 3 and 4 of the present invention;

[0076] Figure 5 is a schematic diagram of the process of differentiated forwarding based on C-APN cognitive semantic fields provided in embodiments 3 and 4 of the present invention;

[0077] Figure 6 is a schematic diagram of the structure of the C-APN semantic header encapsulation and forwarding device provided in an embodiment of the present invention;

[0078] Figure 7 is a schematic diagram of the network device structure provided in an embodiment of the present invention. Detailed Implementation

[0079] The present invention will now be clearly and completely described with reference to the accompanying drawings and embodiments. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0080] In this embodiment of the invention, the term "intelligent agent" can refer to a software entity that possesses the capabilities of task understanding, reasoning, planning, execution, tool invocation, memory access, result verification, or autonomous decision-making. Intelligent agents can be implemented based on large language models, multimodal models, rule engines, retrieval systems, tool execution systems, or other intelligent algorithms.

[0081] The term "agent interaction flow" refers to a set of related messages formed when two or more agents communicate around the same task, the same session, the same inference chain, or the same collaborative process. The agent interaction flow can be determined by one or more of the following: source agent, destination agent, task identifier, session identifier, protocol type, and time window.

[0082] The term "agent communication semantic input information" refers to the communication control information used to generate the C-APN semantic message header and corresponding to the current communication message. The agent communication semantic input information includes at least flow identifier generation information and communication requirement indication information. The flow identifier generation information is used to determine the agent interaction flow to which the current communication message belongs; the communication requirement indication information is used to characterize the processing requirements of the agent interaction flow during forwarding, scheduling, queue allocation, path selection, or computing power guidance. The flow identifier generation information may include at least two of the following: source agent identifier, destination agent identifier, task identifier, session identifier, protocol type, communication direction, interaction stage, message sequence information, or time window information. The communication requirement indication information may include at least one of the following: task priority, token throughput requirement, context dependency degree, semantic error tolerance, inference chain position, latency requirement, quality of service requirement, computing power collaboration requirement, model service type, tool invocation type, data retrieval requirement, or context synchronization state.

[0083] The term "C-APN Semantic Header" refers to an extended header used to carry the cognitive semantic requirements of intelligent agent communication. C-APN is an abbreviation for Cognitive-semantic Agent-oriented Perception Network. The C-APN semantic header is not limited to a specific physical network protocol header, but rather refers to an extended header, metadata header, or protocol extension header used to carry cognitive semantic information of intelligent agent communication and capable of being parsed by network devices or intermediate agent components. The C-APN semantic header includes at least an identifier field, a flow identifier field, and a cognitive semantic field. The C-APN semantic header may carry fields such as flow identifier, semantic tag, task priority, token throughput requirements, context dependency, semantic error tolerance, inference chain position, quality of service level, computing resource requirements, lifecycle, and verification information, enabling network devices or intermediate agent components to perform flow identification, path selection, queue allocation, computing node guidance, and rapid forwarding of subsequent packets based on these fields.

[0084] Example 1: C-APN Semantic Header Field Structure

[0085] As shown in Figure 3, the C-APN semantic header may include basic control fields, flow identification fields, cognitive semantic fields, resource guidance fields, and reliability fields.

[0086] In this embodiment, the C-APN semantic header includes at least an identifier field, a flow identifier field, and a cognitive semantic field. In a preferred implementation, the C-APN semantic header further includes a verification field. The identifier field indicates that the current packet carries a C-APN semantic header; the flow identifier field identifies the agent interaction flow to which the current packet belongs; the cognitive semantic field characterizes the communication processing requirements of the agent interaction flow; and the verification field checks whether the C-APN semantic header fields are complete.

[0087] In one implementation, the C-APN semantic header can be arranged in the following order: basic control field, flow identifier field, cognitive semantic field, resource bootstrapping field, and reliability field. The basic control field, located at the beginning of the header, enables network devices to identify the header version, length, type, and whether it carries a C-APN semantic header. The flow identifier field, following the basic control field, supports flow establishment for the first packet of an interaction flow within the same agent and table lookup and forwarding for subsequent packets. The cognitive semantic field, following the flow identifier field, supports network devices in determining differentiated forwarding strategies based on the agent's communication processing needs. The resource bootstrapping field, following the cognitive semantic field, further indicates QoS queues, network slices, computing nodes, or service entry points. The reliability field, located at the end of the header, limits the validity period of the header or flow forwarding record and performs integrity verification or protection on the header.

[0088] The basic control fields may include a version field, a header length field, a header type field, and an identifier field. The version field supports the evolution of the C-APN semantic header format; the header length field indicates the length of the C-APN semantic header, enabling network devices to determine the boundary between the C-APN semantic header and existing headers or application payloads; the header type field distinguishes between request packets, response packets, context synchronization packets, tool call packets, or result verification packets; and the identifier field indicates that the current packet carries a C-APN semantic header, and this identifier field is generated based on a preset C-APN protocol identifier value. For example, the preset C-APN protocol identifier value can be a 16-bit hexadecimal number 0xCA50. When the network device reads that the identifier field is 0xCA50, it determines that the current packet carries a C-APN semantic header and continues to parse the header length field, header type field, flow identifier field, and cognitive semantic field according to the preset header format. The field length and specific value of the protocol identifier can be predetermined by the C-APN protocol or determined by negotiation between the communicating parties. This embodiment does not impose any limitations on this.

[0089] The flow identifier field is used to uniquely or nearly uniquely identify the agent interaction flow to which the current communication message belongs. An agent interaction flow refers to a set of related communication messages generated by two or more agents around the same task, the same session, the same inference chain, or the same tool invocation process. Network devices can determine whether the current message is the first message or a subsequent message in a certain agent interaction flow based on the flow identifier field, and accordingly perform flow establishment for the first message or table lookup and forwarding for subsequent messages.

[0090] The flow identifier field can be generated based on at least two of the following: source agent identifier, destination agent identifier, task identifier, session identifier, protocol type, communication direction, interaction stage, message sequence information, or time window information. Alternatively, it can be assigned by the upstream agent orchestration component or network control component. For example, in a scenario where a planning agent invokes a retrieval agent to execute task T001, the flow identifier field can be generated based on the source agent identifier, destination agent identifier, task identifier, session identifier, and protocol type. Subsequent messages belonging to the same task invocation process carry the same flow identifier field, allowing the network device to query the generated flow forwarding record and perform fast forwarding based on this flow identifier field.

[0091] Cognitive semantic fields are used to characterize the communication processing requirements of the current agent interaction flow in network forwarding, queue scheduling, path selection, or computer-network collaborative processing. These cognitive semantic fields do not carry complete application-layer business content; instead, they convert the semantic requirements related to agent communication forwarding into level values, enumeration values, or identifier values ​​that can be recognized by network devices.

[0092] The cognitive semantic fields may include at least one of the following: task priority field, token throughput level field, context dependency level field, semantic error tolerance level field, inference chain position field, and computing resource requirement field. The task priority field indicates whether the current task is a normal task, a priority task, or a critical task; the token throughput level field characterizes the current agent interaction flow's requirement for model input or output rate; the context dependency level field characterizes the degree of dependence of the current agent interaction flow on cross-turn dialogue, shared memory, or the continuity of intermediate inference states; the semantic error tolerance level field characterizes the degree of tolerance of the current agent interaction flow for semantic shifts caused by latency, jitter, out-of-order delivery, packet loss, or missing context; the inference chain position field characterizes the position of the current communication message in the collaborative inference chain; and the computing resource requirement field characterizes whether the current communication message needs to be guided to a model inference node, edge computing node, tool execution node, or data retrieval node.

[0093] For example, when the Token throughput level field is high throughput, the network device can choose a high-bandwidth link or increase the queue scheduling weight; when the context dependency level field is strong dependency and the semantic error tolerance level field is low tolerance, the network device can choose a low-latency, low-jitter, or low-packet-loss path; when the inference chain position field indicates that the current packet is at a critical inference chain node, the network device can perform path preservation, resource reservation, or priority forwarding; when the computing resource requirement field indicates tool execution guidance, the network device or computing network gateway can guide the packet to the tool execution node. The resource guidance field can include a service quality level field, a network slice identifier field, a model service identifier field, a tool service identifier field, and a data service identifier field. These fields are used to further indicate the specific QoS queue, network slice, model service, tool service, or data service entry point, based on the communication processing requirements determined by the cognitive semantic fields.

[0094] Reliability fields can include lifecycle fields, check fields, and integrity protection fields. Lifecycle fields are used to limit the validity period of C-APN semantic headers or flow forwarding records; check fields are used to check the integrity of header fields; and integrity protection fields are used to prevent tampering with critical semantic fields.

[0095] Example 2: System Architecture

[0096] In one embodiment, the system provided by this invention includes a message header processing entity and a network device. The agent application, serving as the source of communication messages and agent communication semantic input information, can be located inside or outside the system.

[0097] The intelligent agent application may include a planning intelligent agent, a retrieval intelligent agent, a reasoning intelligent agent, a tool invocation intelligent agent, a code execution intelligent agent, a visual understanding intelligent agent, or a result verification intelligent agent.

[0098] The header processing entity is used to acquire the agent's communication semantic input information, generate a flow identifier field and a cognitive semantic field based on the agent's communication semantic input information, and assemble and encapsulate a C-APN semantic header according to a preset header format. The header processing entity is a functional entity that can be implemented through software, hardware, or a combination of both, and can be deployed in the agent runtime, proxy gateway, service mesh data plane proxy, network edge device, or computing network scheduling gateway.

[0099] The network device is used to receive agent communication messages carrying C-APN semantic message headers and parse the C-APN semantic message headers. The network device can be an edge gateway, virtual switch, router, service mesh data plane proxy, SDN switching node, computing network scheduling gateway, or agent communication proxy node.

[0100] The network device includes a flow forwarding record management function, which is used to generate a flow forwarding record when no flow forwarding record corresponding to the flow identifier field is found, and to read, update and age the flow forwarding record when a corresponding flow forwarding record is found, thereby enabling fast forwarding of subsequent packets.

[0101] Through the above system architecture, the cognitive semantic requirements generated by the application layer of the intelligent agent can be transmitted to the network device in the form of a message header. This enables the network device to complete the identification, flow establishment, table lookup and forwarding, and differentiated resource guidance of the intelligent agent communication message without repeatedly performing deep parsing of the application layer content.

[0102] Example 3: C-APN Semantic Header Encapsulation and Forwarding Method

[0103] As shown in Figure 1, the C-APN semantic header encapsulation and forwarding method of the system in Embodiment 2 provided by this invention may include the following steps:

[0104] S101 receives communication messages initiated by the intelligent agent.

[0105] In this embodiment, the communication message can be a request message sent by the source agent to the destination agent, a response message returned by the destination agent to the source agent, or a message message used by multiple agents for context synchronization, tool invocation, model invocation, result verification, or task status feedback.

[0106] The communication messages can take the form of HTTP, RPC, WebSocket, message queues, event buses, or multi-agent cooperation protocols. This invention does not limit the specific protocol type of the communication messages.

[0107] S102, Obtain the intelligent agent communication semantic input information corresponding to the communication message.

[0108] In this embodiment, the agent communication semantic input information is communication control information used to generate the C-APN semantic message header. It can be obtained by the agent runtime component, proxy gateway, service mesh data plane proxy, network edge device, or computing network scheduling gateway based on the current communication message and its communication context.

[0109] The semantic input information for agent communication includes flow identifier generation information and communication requirement indication information. The flow identifier generation information is used to determine the agent interaction flow to which the current communication message belongs, and may include at least two of the following: source agent identifier, destination agent identifier, task identifier, session identifier, protocol type, communication direction, interaction stage, message sequence information, or time window information.

[0110] The communication requirement indication information is used to characterize the communication processing requirements of the agent's interaction flow, and may include at least one of the following: task priority, token throughput requirement, context dependency degree, semantic error tolerance, inference chain position, latency requirement, quality of service requirement, computing power collaboration requirement, model service type, tool call type, data retrieval requirement, or context synchronization state.

[0111] It should be noted that this embodiment does not limit the specific method of obtaining the semantic input information of the intelligent agent communication, as long as the information can be used to generate the flow identifier field and cognitive semantic field in the C-APN semantic message header.

[0112] S103, Generate C-APN semantic message header based on the intelligent agent communication semantic input information.

[0113] In this embodiment, generating a C-APN semantic header may include three processes: field determination, field mapping, and field assembly.

[0114] First, the header processing entity determines the set of fields for the C-APN semantic header based on a preset header format. The identifier field, flow identifier field, and cognitive semantic field are the core fields. In this embodiment, the field set also includes a header length field, a header type field, and a checksum field. Specifically, the identifier field is generated based on a preset C-APN protocol identifier value; the header length field represents the total length of the C-APN semantic header; the header type field represents the encapsulation type or bearer protocol type used; and the checksum field is used to perform integrity checks on at least a portion of the header content, excluding the checksum field itself.

[0115] Secondly, the message header processing entity extracts flow identifier generation information from the agent communication semantic input information and generates a flow identifier field based on the flow identifier generation information. The flow identifier generation information may include at least two of the following: source agent identifier, destination agent identifier, task identifier, session identifier, protocol type, communication direction, interaction stage, or message sequence information. In one implementation, the message header processing entity may concatenate or encode the source agent identifier, destination agent identifier, task identifier, session identifier, and protocol type, and perform hash calculation on the concatenated or encoded result to obtain the flow identifier field.

[0116] Next, the message header processing entity extracts communication requirement indication information from the agent's communication semantic input information and generates cognitive semantic fields according to preset semantic mapping rules. The communication requirement indication information may include at least one of the following: task priority, token throughput requirement, context dependency degree, semantic error tolerance, inference chain position, latency requirement, quality of service requirement, computing power collaboration requirement, model service type, tool call type, data retrieval requirement, or context synchronization status.

[0117] The cognitive semantic fields include at least one or more of the following: task priority field, token throughput level field, context dependency level field, semantic error tolerance level field, inference chain position field, and computing power resource requirement field.

[0118] For example, when the communication requirement indication information includes a token throughput requirement, the token throughput requirement is mapped to a token throughput level field; when the communication requirement indication information includes a context dependency degree, the context dependency degree is mapped to a context dependency level field; when the communication requirement indication information includes a semantic error tolerance, the semantic error tolerance is mapped to a semantic error tolerance level field; when the communication requirement indication information includes a task priority, the task priority is mapped to a task priority field; when the communication requirement indication information includes an inference chain position, the inference chain position is mapped to an inference chain position field; when the communication requirement indication information includes a computing power collaboration requirement, a model service type, a tool call type, or a data retrieval requirement, it is mapped to a computing power resource requirement field or a corresponding service identifier field.

[0119] Finally, the header processing entity generates a header length field based on the total length of the field set, and a header type field based on the encapsulation type or bearer protocol type used. It then writes the identifier field, header type field, header length field, flow identifier field, and cognitive semantic field in the field order specified by the preset header format. Finally, it calculates a checksum field based on at least a portion of the header content (excluding the checksum field) and writes the checksum field to the corresponding position, thus obtaining the C-APN semantic header. The preset header format specifies the field type, field length, and arrangement order of each field, which can be pre-defined by the C-APN protocol or pre-negotiated by the communicating parties.

[0120] S104, the C-APN semantic message header is encapsulated outside the original message header of the communication message, and then the intelligent agent communication message carrying the C-APN semantic message header is generated.

[0121] like Figure 2As shown, after encapsulating the C-APN semantic header outside the original header of the communication message, an agent communication message carrying the C-APN semantic header is obtained. The agent communication message carrying the C-APN semantic header may include the C-APN semantic header, the original network layer or transport layer header, application layer control information, and application data payload. The C-APN semantic header, located outside the original header, is used to explicitly provide the network device with the flow identifier of the agent interaction flow, communication processing requirements, and forwarding control information.

[0122] In this embodiment, different encapsulation methods can be adopted depending on the network environment and the location of the protocol stack.

[0123] In one implementation, the C-APN semantic header can be encapsulated as an outer extension header between the outer network encapsulation header and the original network layer header. This approach is suitable for network edge devices, tunnel gateways, or SDN forwarding scenarios.

[0124] In another implementation, the C-APN semantic header can be encapsulated between the transport layer header and the application layer payload. This approach is suitable for agent communication scenarios that require carrying semantic information in the transport channel.

[0125] In another implementation, the C-APN semantic header can be encapsulated in the metadata fields of HTTP, RPC, WebSocket, message queue, or event bus messages. This approach is suitable for proxy gateways, service mesh data plane proxies, or application layer middleware scenarios.

[0126] In another implementation, the C-APN semantic header can be used as an extension header for the agent communication protocol. This approach is suitable for communication protocols specifically designed for multi-agent cooperation.

[0127] In this way, the communication requirements of intelligent agents can be encapsulated into message header fields that can be recognized by network devices or intermediate proxy components without changing the semantics of application layer business load.

[0128] S105, the network device parses the agent communication message carrying the C-APN semantic message header to obtain the flow identifier field and the cognitive semantic field.

[0129] After receiving an agent communication message carrying a C-APN semantic header, the network device first determines whether the message carries a C-APN semantic header based on the identifier field. If it does, it further parses the flow identifier field and the cognitive semantic field.

[0130] During the parsing process, when the C-APN semantic header includes a checksum field, the network device can perform integrity verification on the C-APN semantic header based on the checksum field. If the checksum field is not included or the verification passes, the cognitive semantic field is read and used for forwarding decisions; if the verification fails, the C-APN semantic header can be discarded, basic forwarding can be downgraded, or an exception record can be generated.

[0131] If the network device does not support recognizing the C-APN semantic header, it can perform basic forwarding according to the original header, or it can forward the C-APN semantic header after passing it through, removing it, or converting it. This ensures the compatibility of the method of this invention in heterogeneous networks.

[0132] S106, perform forwarding processing on the agent communication message according to the flow identifier field and the cognitive semantic field.

[0133] In this embodiment, the network device determines whether the agent communication message is the first message of the corresponding agent interaction stream.

[0134] If the agent communication message is the first message, the network device obtains the basic next-hop path based on the original message header and determines the semantic forwarding strategy based on the cognitive semantic field in the C-APN semantic message header. Subsequently, the network device generates a flow forwarding record based on the basic next-hop path and the semantic forwarding strategy.

[0135] The flow forwarding record may include flow identifier, source agent identifier, destination agent identifier, source address, destination address, source port, destination port, protocol type, next-hop path, QoS queue identifier, network slice identifier, computing node identifier, model service identifier, tool service identifier, data service identifier, cognitive semantic field, forwarding action, record generation time, and record aging time.

[0136] If the agent communication message is a subsequent message, the network device queries the flow forwarding record based on the flow identifier field and forwards the subsequent message according to the forwarding action in the flow forwarding record. Therefore, subsequent messages do not need to repeatedly parse all the original message headers and cognitive semantic fields, which improves forwarding efficiency.

[0137] It should be noted that step S106 can further include two levels of processing. The first level is the process of establishing a flow for the first message and looking up and forwarding subsequent messages based on the flow identifier field. This process is used to determine whether the current agent communication message is the first message or a subsequent message, and to decide whether to generate a flow forwarding record or query an existing flow forwarding record. This process can be found in [reference needed]. Figure 4The second layer is the process of determining differentiated forwarding strategies based on cognitive semantic fields. This process determines specific forwarding actions based on fields such as token throughput level, context dependency level, semantic error tolerance level, task priority, inference chain position, and computing resource requirements. This process can be found in [reference needed]. Figure 5 .

[0138] In other words, Figure 4 The flowchart shown focuses on describing the processing relationship between the first message and subsequent messages in the same agent's interaction flow. Figure 5 The illustrated process focuses on describing how to determine a differentiated forwarding strategy based on the C-APN cognitive semantic field when generating or updating flow forwarding records. These two factors together constitute the specific implementation of step S106.

[0139] In step S106 above, the differentiated forwarding strategy determination process is used to determine the specific forwarding action during the first packet processing and write the forwarding action into the flow forwarding record; subsequent packets that hit the flow forwarding record are forwarded according to the forwarding action in the record. The processing flow for establishing a flow for the first packet and looking up tables for subsequent packets is described below. Figure 4 For the process of determining the differentiated forwarding strategy, please refer to [link / reference]. Figure 5 .

[0140] like Figure 5 As shown, after receiving an agent communication message carrying a C-APN semantic message header, the network device parses the C-APN semantic message header to obtain semantic tags and cognitive semantic fields. Then, based on the semantic tags, it determines the differentiated forwarding strategy type corresponding to the message, and combines fields such as token throughput level, context dependency level, semantic error tolerance level, task priority, inference chain position, and computing power resource requirements to determine specific path selection, queue allocation, slice selection, or computing power node guidance actions.

[0141] In one implementation, when the Token throughput field indicates a high throughput level, the network device can select a high-bandwidth link, increase the scheduling weight of the corresponding queue, or forward communication packets to a high-throughput model service instance to ensure the continuity of model output.

[0142] In one implementation, when the context dependency field represents a strong dependency level and the semantic error tolerance field represents a low tolerance level, the network device can choose a low latency, low jitter, or low packet loss path to reduce the risk of semantic shifts in the upper-layer inference results caused by context transmission anomalies, data out-of-order, or latency fluctuations.

[0143] In one implementation, when the task priority field indicates a high-priority task, the network device can map the agent's communication message to a high-priority QoS queue or to a critical task network slice to reduce task execution blocking.

[0144] In one implementation, when the inference chain position field indicates that the current packet is at a critical inference chain node, the network device can perform path preservation, resource reservation, or priority forwarding to prevent abnormal responses from that node from affecting multiple downstream agent tasks.

[0145] In one implementation, when the computing resource requirement field indicates the need for computing power collaboration, the network device can guide the agent's communication message to the corresponding edge computing node, model inference node, tool execution node, or data retrieval node, thereby achieving collaborative selection of communication path and computing resources.

[0146] In this way, network devices can classify communication messages of different agents into throughput-sensitive, low-latency and high-reliability, high-priority guarantee, critical link guarantee, or computing power-guided types based on the cognitive semantic fields carried in the C-APN semantic message header, and execute the corresponding forwarding strategies accordingly.

[0147] Example 4: Generation, Query, and Aging Methods of Flow Forwarding Records

[0148] This embodiment is a further refinement of step S106 in embodiment three above, used to illustrate how network devices distinguish between the first message and subsequent messages when processing the same intelligent agent interaction stream, and how to reduce the overhead of repeated parsing through flow forwarding records.

[0149] like Figure 4 As shown, after receiving an agent communication message carrying a C-APN semantic header, the network device parses the C-APN semantic header, extracts the flow identifier field and cognitive semantic field, and queries the corresponding flow forwarding record based on the flow identifier field. If no corresponding flow forwarding record is found, the agent communication message is treated as the first message: the basic next-hop path is determined based on the original message header, the semantic forwarding policy is determined based on the cognitive semantic field, and a flow forwarding record is generated based on the basic next-hop path and the semantic forwarding policy. The flow forwarding record may include forwarding actions such as next-hop path, QoS queue, network slice, computing node or service entry, and set corresponding aging time. If a corresponding flow forwarding record is found, forwarding is performed according to the forwarding action in the flow forwarding record.

[0150] For example, when the first message corresponds to a high token throughput requirement, the network device can write a high-bandwidth path or a high-priority queue in the flow forwarding record; when the first message corresponds to a strong context dependency and low semantic error tolerance, the network device can write a low-latency path in the flow forwarding record; when the first message corresponds to a computing power collaboration requirement, the network device can write the target computing power node or target model service identifier in the flow forwarding record.

[0151] When subsequent packets belonging to the same agent's interaction flow arrive, the network device parses the flow identifier field of the subsequent packet and queries the flow forwarding record based on the flow identifier field. When the corresponding flow forwarding record is found, the network device does not need to re-execute the complete semantic forwarding policy calculation. Instead, it performs forwarding according to the next-hop path, QoS queue, network slice, computing node identifier, or other forwarding actions in the flow forwarding record, and updates the status information and / or aging time of the flow forwarding record.

[0152] When no corresponding flow forwarding record is found based on the flow identifier field, the network device processes the current agent communication packet as the first packet, re-parses its C-APN semantic header and the original header, determines the semantic forwarding strategy, and establishes a new flow forwarding record. This processing method can restore forwarding even when the original flow forwarding record has aged, the device has restarted, the forwarding table has been lost, or the path has been rebuilt.

[0153] To prevent flow forwarding records from occupying device resources for extended periods, network devices can set an aging time for these records. When the aging time expires and no subsequent packets belonging to the same agent's interaction flow are received, the network device deletes the flow forwarding record.

[0154] In another implementation, the network device can also dynamically adjust the aging time of the flow forwarding record based on the packet arrival frequency, task status, session status, or the lifecycle field in the C-APN semantic header. For example, a longer aging time can be set for agent interaction flows that are on the critical inference chain; a shorter aging time can be set for low-priority or one-time tool call packets.

[0155] Example 5: Workflow Example

[0156] In a multi-agent collaborative question-answering scenario, the planning agent sends a retrieval request to the retrieval agent. The message header processing entity, deployed in the agent runtime, proxy gateway, or network edge node, obtains the agent communication semantic input information corresponding to the request, determines that the request has a high context dependency and a high task priority, and then generates a C-APN semantic message header carrying a strong context dependency level and a high task priority level.

[0157] After receiving the initial packet of the request, the network device parses the C-APN semantic header to determine that the request should prioritize low-latency paths and high-priority queues, and generates a flow forwarding record. When subsequent packets of the same retrieval interaction flow arrive, the network device queries the flow forwarding record based on the flow identifier field and forwards the packets quickly according to the forwarding action in the record.

[0158] In another scenario involving the inference agent's tool invocation, the inference agent needs to invoke the code execution agent. The computational resource requirement field in the C-APN semantic header indicates that the request needs to be routed to the tool execution node. After parsing the initial packet, the network device can route the corresponding traffic to an edge node with code execution capabilities and write the node's identifier into the flow forwarding record. Subsequent packets are then forwarded directly to this node based on the flow record.

[0159] Through the above process, the underlying network no longer processes agent communication traffic based solely on address, port, or protocol, but is able to perform differentiated forwarding based on the task semantics and cognitive communication requirements in the C-APN semantic header.

[0160] Example 6: Device Structure

[0161] As shown in Figure 6, the device referred to in this embodiment is a collection of functional modules deployed in the same network edge device, proxy gateway, or computing network scheduling gateway, which can be implemented through software, hardware, or a combination of both. The device may include a message receiving module, an information acquisition module, a semantic message header generation module, a message header encapsulation module, a semantic message header parsing module, a first message processing module, a flow forwarding record management module, a subsequent message processing module, a differentiated forwarding control module, and a message sending module.

[0162] The message receiving module is used to receive communication messages initiated by the agent, or to receive agent communication messages carrying C-APN semantic message headers from other network nodes.

[0163] The information acquisition module is used to acquire the agent communication semantic input information corresponding to the communication message. The agent communication semantic input information includes flow identifier generation information and communication requirement indication information. The flow identifier generation information is used to determine the agent interaction flow to which the current communication message belongs, and the communication requirement indication information is used to characterize the communication processing requirements of the agent interaction flow.

[0164] The semantic header generation module is used to generate a C-APN semantic header based on the agent's communication semantic input information. The C-APN semantic header may include a flow identifier field, a semantic tag field, a task priority field, a token throughput level field, a context dependency level field, a semantic error tolerance level field, an inference chain position field, a quality of service level field, a computing resource requirement field, a lifecycle field, and a verification field.

[0165] The message header encapsulation module is used to encapsulate the C-APN semantic message header outside the original message header of the communication message, so as to obtain an agent communication message carrying the C-APN semantic message header.

[0166] The semantic header parsing module is used to parse agent communication messages carrying C-APN semantic headers to obtain the flow identifier field and cognitive semantic field. The semantic header parsing module can also be used to perform integrity verification on the C-APN semantic header.

[0167] The first message processing module is used to determine the next-hop forwarding path based on the original message header and the C-APN semantic message header when the agent communication message is the first message of the agent interaction flow, and to trigger the generation of the flow forwarding record.

[0168] The flow forwarding record management module is used to create, query, update, and age flow forwarding records. The flow forwarding record may include a flow identifier, source agent identifier, destination agent identifier, next-hop path, QoS queue identifier, computing node identifier, cognitive semantic field, forwarding action, and record aging time.

[0169] The subsequent message processing module is used to query the flow forwarding record based on the flow identifier when the communication message of the agent is a subsequent message of the same agent interaction flow, and determine the forwarding action based on the query result.

[0170] The differentiated forwarding control module is used to perform path selection, QoS queue allocation, network slice selection, computing node guidance, or service entry selection based on the semantic tags in the C-APN semantic header or the forwarding actions in the flow forwarding record.

[0171] The message sending module is used to send agent communication messages according to the forwarding strategy determined by the differentiated forwarding control module.

[0172] The modules described above can be implemented in software, hardware, or a combination of both. Each module can be deployed in an intelligent agent operating platform, proxy gateway, service mesh component, network edge device, network controller, computing network scheduling device, or other devices with communication processing capabilities.

[0173] Example 7: Network Device

[0174] As shown in Figure 7, the network device provided in this embodiment of the invention may further include a processor, a memory, a communication interface, a bus, and a forwarding chip or a switching chip. The processor, memory, communication interface, and forwarding chip or switching chip can be connected via a bus.

[0175] The memory can store computer program instructions and a flow forwarding record table. When the computer program instructions are executed by the processor, the network device receives agent communication messages carrying C-APN semantic message headers, parses the C-APN semantic message headers, generates or queries flow forwarding records, and performs forwarding processing on the agent communication messages.

[0176] The communication interface is used to receive intelligent agent communication message input and output intelligent agent communication messages after forwarding processing. The forwarding chip or switching chip is used to perform message forwarding, queue scheduling, port selection, or path selection according to the forwarding action determined by the processor.

[0177] In one implementation, the network device may include a message receiving unit, a semantic message header parsing unit, a first message processing unit, a subsequent message processing unit, a flow forwarding record management unit, a differentiated forwarding control unit, and a message sending unit.

[0178] The message receiving unit receives agent communication messages carrying C-APN semantic message headers. The semantic message header parsing unit parses the C-APN semantic message header to obtain flow identifiers and semantic tags. The first message processing unit generates a flow forwarding record when the agent communication message is the first message. The subsequent message processing unit queries the flow forwarding record when the agent communication message is a subsequent message. The flow forwarding record management unit maintains the flow forwarding record table. The differentiated forwarding control unit determines the forwarding action based on the semantic tags or flow forwarding records. The message sending unit sends the agent communication message according to the forwarding action.

[0179] The network device may be an edge gateway, virtual switch, router, service mesh data plane proxy, SDN switching node, computing network scheduling gateway, intelligent agent communication proxy node, or other device capable of performing packet parsing and forwarding.

[0180] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the above embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A C-APN semantic message header encapsulation and forwarding system for intelligent agent communication, characterized in that: The system includes message header processing entities and network devices: The message header processing entity is used to receive communication messages initiated by the agent, obtain the agent communication semantic input information corresponding to the communication message, generate a flow identifier field and a cognitive semantic field according to the agent communication semantic input information, assemble a C-APN semantic message header according to a preset message header format, and encapsulate the C-APN semantic message header outside the original message header of the communication message to obtain an agent communication message carrying the C-APN semantic message header; The network device is used to parse the agent communication message carrying the C-APN semantic message header, obtain the flow identifier field and the cognitive semantic field, generate a flow forwarding record when no flow forwarding record corresponding to the flow identifier field is found, and read the flow forwarding record when a corresponding flow forwarding record is found, and perform forwarding processing according to the cognitive semantic field or the flow forwarding record; The C-APN semantic header encapsulation and forwarding system operates according to the following steps: S101, Receive communication messages initiated by the intelligent agent; S102. Obtain agent communication semantic input information corresponding to the communication message. The agent communication semantic input information includes flow identifier generation information for determining the agent interaction flow to which the communication message belongs, and communication requirement indication information for characterizing the communication processing requirements of the agent interaction flow. S103. Generate a C-APN semantic message header based on the agent communication semantic input information. The C-APN semantic message header includes at least an identifier field indicating that the communication message carries a C-APN semantic message header, a flow identifier field identifying the agent interaction flow to which the communication message belongs, and a cognitive semantic field characterizing the cognitive communication needs of the agent interaction flow. Each field is arranged according to a preset message header format. S104. Encapsulate the C-APN semantic header outside the original header of the communication message to obtain an agent communication message carrying the C-APN semantic header; S105. The network device parses the agent communication message carrying the C-APN semantic message header to obtain the flow identifier field and the cognitive semantic field; S106. Perform forwarding processing on the agent communication message according to the flow identifier field and the cognitive semantic field; C-APN is an abbreviation for Cognitive-semantic Agent-oriented Perception Network, meaning a cognitive semantic perception network oriented towards intelligent agents; the intelligent agent communication messages include at least one of the following: HTTP request messages, RPC call messages, WebSocket messages, message queue messages, event bus messages, multi-agent cooperation protocol messages, model call request messages, tool call request messages, context synchronization messages, and result verification messages.

2. The system according to claim 1, characterized in that: The stream identifier generation information includes at least two of the following: source agent identifier, destination agent identifier, task identifier, session identifier, protocol type, communication direction, interaction stage, message sequence information, or time window information; The communication requirement indication information includes at least one of the following: task priority, token throughput requirement, context dependency level, semantic error tolerance, inference chain position, latency requirement, service quality requirement, computing power collaboration requirement, model service type, tool call type, data retrieval requirement, or context synchronization status.

3. The system according to claim 1, characterized in that: The cognitive semantic field includes at least one of the following: The Token throughput field is used to characterize the model's input or output rate requirements; Context-dependent fields are used to characterize the continuity requirements of cross-turn dialogues, context windows, shared memory, or intermediate inference states; The semantic error tolerance field is used to characterize the degree to which the agent's interaction flow tolerates semantic deviations caused by latency, jitter, out-of-order delivery, packet loss, or missing context. The task priority field is used to characterize the importance of the current task in the multi-agent collaboration process; The inference chain position field is used to characterize the position or scope of influence of the current communication message in the cooperative inference chain; The computing resource requirement field is used to indicate whether the current communication message needs to be guided to a model inference node, edge computing node, tool execution node, or data retrieval node.

4. The system according to claim 3, characterized in that: The C-APN semantic header also includes a header length field, a header type field, and a checksum field; generating the C-APN semantic header includes the following steps: The set of fields for the C-APN semantic header is determined according to the preset header format, and the identifier field is generated according to the preset C-APN protocol identifier value; Extract stream identifier generation information from the agent communication semantic input information, and generate the stream identifier field based on the stream identifier generation information; The communication requirement indication information is extracted from the intelligent agent's communication semantic input information, and mapped to one or more cognitive semantic fields according to a preset semantic mapping rule; the message header length field is generated according to the total length of the field set, and the message header type field is generated according to the encapsulation type or bearer protocol type used; the identifier field, message header type field, message header length field, flow identifier field and cognitive semantic field are written in the field order specified by the preset message header format, and the check field is calculated based on at least part of the message header content except for the check field, and the check field is written to the corresponding position to obtain the C-APN semantic message header. The preset semantic mapping rules include at least one of the following: Map token throughput requirements to low, medium, or high throughput levels; Map the degree of context dependency to a weak dependency level, a normal dependency level, or a strong dependency level; Map semantic error tolerance to high tolerance level, medium tolerance level, or low tolerance level; Map task priorities to normal task level, priority task level, or critical task level; Map the computing power collaboration requirements to no-computing-power guidance, edge inference guidance, model service guidance, tool execution guidance, or data retrieval guidance.

5. The system according to claim 1, characterized in that: In step S104, the C-APN semantic header is encapsulated in addition to the original header of the communication message, including: The C-APN semantic header is generated in the agent runtime component, proxy gateway, service mesh data plane proxy, network edge device, or computing network scheduling gateway located between the agent application and the network; The C-APN semantic message header is added to the communication message as one of the outer extension header, tunnel encapsulation header, transport layer extension header, application layer metadata header, or intelligent agent communication protocol extension header; The C-APN semantic header is encapsulated in any of the following locations: between the outer network encapsulation header and the original network layer header; between the transport layer header and the application layer payload; in the metadata field of an HTTP, RPC, WebSocket, message queue, or event bus message; or in the extended protocol header carrying the agent invocation request.

6. The system according to claim 1, characterized in that: The step of forwarding the agent communication message based on the flow identifier field and the cognitive semantic field includes the following steps: Determine whether the agent communication message is the first message of the agent interaction stream; If it is the first message of the intelligent agent interaction flow, the basic next-hop path is obtained according to the original message header of the communication message, the semantic forwarding strategy is determined according to the cognitive semantic field, and a flow forwarding record is generated according to the basic next-hop path and the semantic forwarding strategy. If it is a subsequent message of the intelligent agent interaction stream, then query the stream forwarding record according to the stream identifier field, and forward the subsequent message according to the forwarding action in the stream forwarding record; The flow forwarding record includes at least one of the following: flow identifier, source agent identifier, destination agent identifier, source address, destination address, source port, destination port, protocol type, next-hop path, QoS queue identifier, network slice identifier, computing node identifier, model service identifier, tool service identifier, data service identifier, cognitive semantic field, forwarding action, record generation time, and record aging time.

7. The system according to claim 6, characterized in that: Determining a semantic forwarding strategy based on the cognitive semantic field includes at least one of the following: When the Token throughput field indicates a high throughput level, select a high-bandwidth link, increase the queue scheduling weight, or forward to a high-throughput service instance; When the context dependency field indicates a strong dependency level and the semantic error tolerance field indicates a low tolerance level, select a low latency, low jitter, or low packet loss path. When the task priority field indicates a high-priority task, the agent communication message is mapped to a high-priority QoS queue or a critical task network slice. When the inference chain position field indicates that the current message is at a critical inference chain node, the execution path of the agent's communication message is maintained, resources are reserved, or priority is forwarded. When the computing power resource requirement field indicates that computing power collaboration is needed, the intelligent agent communication message is directed to the corresponding edge computing power node, model inference node, tool execution node, or data retrieval node.

8. The system according to claim 1, characterized in that: In step S105, the process of parsing the agent communication message carrying the C-APN semantic header by the network device to obtain the flow identifier field and the cognitive semantic field includes: After receiving the intelligent agent communication message carrying the C-APN semantic message header, the network device identifies the C-APN semantic message header; when the C-APN semantic message header includes a check field, the integrity of the C-APN semantic message header is checked according to the check field. When the C-APN semantic header does not include a verification field or the integrity verification passes, the C-APN semantic header is parsed to obtain the flow identifier field and the cognitive semantic field; When the C-APN semantic header includes a verification field and the integrity verification fails, the C-APN semantic header is discarded, basic forwarding is downgraded, or an exception record is generated.

9. The system according to claim 6, characterized in that: After generating the flow forwarding record based on the basic next-hop path and the semantic forwarding strategy, the method further includes: setting an aging time for the flow forwarding record; updating the status information or aging time of the flow forwarding record when a subsequent message belonging to the same agent interaction flow is received and the flow forwarding record is found; and deleting the flow forwarding record when the aging time expires and no subsequent message belonging to the same agent interaction flow is received.

10. The system according to claim 1, characterized in that: Network equipment, including a receiving unit, a parsing unit, and a stream record management unit; The receiving unit is used to receive agent communication messages carrying C-APN semantic message headers; The parsing unit is used to parse the C-APN semantic header to obtain the flow identifier field and the cognitive semantic field; The flow record management unit is used to query flow forwarding records based on the flow identifier field, and generate flow forwarding records based on the flow identifier field, the cognitive semantic field, and the original packet header when no corresponding flow forwarding record is found. The flow record management unit is also used to read the flow forwarding record when a corresponding flow forwarding record is found, and to update the status information or aging time of the flow forwarding record; The forwarding unit is used to perform forwarding processing on the agent communication message according to the cognitive semantic field or the stream forwarding record.

Images