Network Working Group S. Wang Internet-Draft L. Wang Intended status: Standards Track Y. Zhang Expires: 20 May 2027 L. Ma Y. Cao Y. Liu China Unicom 16 November 2026 Research on service semantic identification and routing technology for Internet of agents draft-wang-agent-semantic-routing-00 Abstract With the deep integration of artificial intelligence and network technologies, the Internet of agents has become an important direction for the evolution of the next-generation Internet. The connection paradigm of the traditional Internet centered on “location addressing” can hardly meet the demands of the Internet of agents for “business-driven and service-as-connection”. The service semantic routing technology system, paradigm evolution, core collaboration requirements in the Internet of agents, and the limitations of traditional routing technologies were systematically elucidated. It was clarified that semantic routing was the key to realizing the transformation from “service demand” to “resource matching”. The solutions for four core semantic routing mechanisms in the Internet of agents, namely fuzzy search based on tag matching and flexible adaptation, precision-named DNS (domain name system) for “service- domain name” mapping, business-layer network identification based on Anycast, and URL (uniform resource locator) semantic routing with application-layer semantic encapsulation were analyzed in depth. Finally, the technical characteristics of each solution were summarized and compared, providing an important reference for subsequent research. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at https://datatracker.ietf.org/drafts/current/. Wang, et al. Expires 20 May 2027 [Page 1] Internet-Draft Service Semantic Routing for IoA November 2026 Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 20 May 2027. Copyright Notice Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/ license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License. Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 2. Routing Requirements Analysis for the Internet of Agents . . 4 2.1. Core Characteristics and Requirements of the Internet of Agents . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.2. Limitations of Traditional Routing . . . . . . . . . . . 5 3. State of the Art in Service Semantic Identification and Routing . . . . . . . . . . . . . . . . . . . . . . . . . 7 3.1. Overview of Service Semantic Identification and Routing . . . . . . . . . . . . . . . . . . . . . . . . . 7 3.2. Four Representative Categories of Service Semantic Identification and Routing . . . . . . . . . . . . . . . 8 3.2.1. Fuzzy Search . . . . . . . . . . . . . . . . . . . . 8 3.2.2. DNS-Based Precise Naming . . . . . . . . . . . . . . 9 3.2.3. Anycast-Based Data Plane Network Identification . . . 10 3.2.4. Application-Layer URL-Based Service Semantic Identification . . . . . . . . . . . . . . . . . . . 12 4. Service Semantic Identification and Routing for the Internet of Agents . . . . . . . . . . . . . . . . . . . . . . . . . 13 4.1. Fuzzy Search for the Internet of Agents . . . . . . . . . 13 4.2. DNS-Based Precise Naming for the Internet of Agents . . . 16 4.3. Anycast Identification and Routing for the Internet of Agents . . . . . . . . . . . . . . . . . . . . . . . . . 18 4.3.1. Application-Layer Service Semantic Identification and Routing Based on URLs for the Agent Internet . . . . 21 5. Security Considerations . . . . . . . . . . . . . . . . . . . 25 Wang, et al. Expires 20 May 2027 [Page 2] Internet-Draft Service Semantic Routing for IoA November 2026 6. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 25 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 25 1. Introduction The rapid advancement of the digital economy and artificial intelligence is driving the evolution of the Internet from information interconnection toward intelligent interconnection. Agent technology powered by large language models is precipitating a fundamental transformation in Internet architecture, in which the Internet will evolve into an "Internet of Agents" (IoA) — a network of intelligent entities capable of perception, decision-making, execution, and interaction. The central requirement of the IoA is service-driven interconnection: agents must autonomously discover, match, and invoke heterogeneous service resources — including other agents, large models, knowledge bases, and compute nodes — based on high-level business objectives. In this paradigm, the core interaction unit shifts from hosts or service endpoints to agents, and the mode of accomplishing complex tasks evolves from user-to-service access to agent-to-agent collaborative service. Traditional Internet routing is centered on location-based addressing, relying on physical identifiers such as IP addresses to establish connectivity between nodes. It cannot directly interpret the business semantics or service requirements of agents, creating an efficiency bottleneck in the demand-expression-to-resource-matching pipeline for inter-agent coordination. For example, when an agent requests "a low-latency Chinese-language LLM inference service," traditional routing requires multiple layers of adaptation and translation and cannot directly map the request to an appropriate service instance in semantic form, increasing coordination complexity and interaction latency. Building an identification and routing framework that supports semantic understanding and precise matching is therefore a critical enabler for the large-scale development of the IoA. This paper investigates service semantic identification and routing techniques for the IoA. The main topics covered are: the core requirements of the IoA and the semantic identification paradigm; the principles, agent-oriented designs, and technical details of four categories of semantic routing mechanisms — fuzzy search, DNS-based precise naming, anycast network identification, and application-layer URL-based semantic routing; a comparative analysis of these four approaches; and an outlook on future directions. Wang, et al. Expires 20 May 2027 [Page 3] Internet-Draft Service Semantic Routing for IoA November 2026 2. Routing Requirements Analysis for the Internet of Agents The Internet of Agents is an emerging Internet paradigm in which highly autonomous AI agents serve as the primary interaction entities, and represents key infrastructure for supporting future large-scale societal-level collaboration. Its central objective is to connect agents, data, models, and tools across cloud, edge, and end-device tiers into an open, extensible, and secure intelligent collaboration network through standardized protocols and open architectures. In this paradigm, agents act as reliable proxies for users, autonomously performing service discovery, task negotiation, and resource scheduling, and accomplishing complex tasks through cross-domain, cross-platform collaboration. The core interaction unit shifts from traditional hosts to autonomous agents, and the primary network function advances from supporting human information exchange to driving efficient and precise inter-agent coordination and value creation. While service discovery and routing have attracted research attention in both academia and industry, a significant gap remains: traditional network technologies lack semantic capability, and AI-oriented research tends to overlook the underlying network support layer. This section systematically characterizes the core features and requirements of the IoA and analyzes the limitations of traditional routing mechanisms, providing the foundation for subsequent investigation. The evolution path from the traditional Internet to the IoA is illustrated in Figure 1. \`\`\`text Boundary Expansion Human-to-Human Communication & Collaboration ===========================> Human-Agent, Agent-Agent Communication & Collaboration Human Agent Agent / / / Traditional Internet \ Vehicle / Agent Internet \ Laptop \ / \\-------/ Human Human Human +Agent Industrial Agent+Human Robot Figure 1: Evolution Path from Traditional Internet to Agent Internet ``` Wang, et al. Expires 20 May 2027 [Page 4] Internet-Draft Service Semantic Routing for IoA November 2026 2.1. Core Characteristics and Requirements of the Internet of Agents In the IoA, although agents possess a closed-loop autonomous cycle of perception, decision-making, and execution, the realization of genuine value and the completion of complex tasks fundamentally depend on large-scale, high-efficiency multi-agent and multi-resource coordination. The core of such coordination is precise matching of service demands to resources: starting from the high-level business service requirements of a user or a leading agent, the network system dynamically discovers and schedules service agents with the corresponding capabilities, ultimately achieving the objective through collaboration. Realizing this coordination process places elevated requirements on network identification and routing: First, the IoA requires strong semantic understanding capability, with the primary objective of accommodating agents' potentially imprecise service demands and supporting demand-to-resource matching in exploratory inter-agent collaboration. Second, a unified, human-readable semantic identifier framework is needed so that the service demand semantics of agents across different domains and types can be effectively recognized, and the service capabilities of various resources can be clearly characterized, enabling interoperability and precise matching between service demands and heterogeneous resources. Third, real-time awareness of service instance operational state is required to dynamically adapt to changing agent service demands and ensure continuous real-time alignment between demand and resource state. Finally, extremely low interaction overhead and millisecond-level response capability are required, primarily to accommodate the high-frequency, short-session, massively concurrent service demand interaction patterns characteristic of inter-agent communication, ensuring rapid resource matching and response following demand initiation. Each of these core characteristics imposes specific design requirements on network identification and routing schemes. 2.2. Limitations of Traditional Routing IP, as the de facto standard for global network communication, identifies network device interfaces by destination IP address and forwards packets end-to-end on a shortest-path-first basis. It offers broad compatibility and generality for traditional data transfer scenarios. However, its inherent design limitations become apparent when faced with the low-latency, high-concurrency, service-direct-delivery requirements of the IoA, and it cannot satisfy the core demand-to-resource-matching objective. Wang, et al. Expires 20 May 2027 [Page 5] Internet-Draft Service Semantic Routing for IoA November 2026 In terms of addressing, the granularity of network-layer IP addresses does not align with that of application-layer service interfaces. IP identifies only network device interfaces or the networks to which they belong, and cannot match the service-invocation granularity driven by agent business logic, preventing routing decisions from precisely addressing business requirements. In terms of scheduling, IP routing defaults to best-effort forwarding; while optimizations such as Equal-Cost Multi-Path (ECMP) and Constrained Shortest Path First (CSPF) improve path selection, IP addresses cannot express complex service requirements, making it difficult to satisfy the differentiated demands of agent workloads for latency, bandwidth, and security isolation, thereby reducing scheduling efficiency. In terms of cross-domain coordination, networks are partitioned into multiple administrative and technical domains whose routing schemes and semantics differ. IP addresses lack a unified information interpretation framework, and cross-domain resource scheduling without complex Service Level Agreement (SLA) support is prone to policy inconsistency and performance instability. Traditional semantic routing refers to the practice of augmenting IP routing decisions with additional information carried in packets or configured into routers. It can be broadly categorized into overlay-based semantic routing and intra-domain semantic routing. Overlay architectures construct an independent overlay network layer on top of the existing IP underlay to carry service semantic routing functions; intra-domain architectures confine semantic routing to a designated physically isolated network domain, with uniform routing policies applied within the domain and inter-domain communication handled through tunnels or application gateways for semantic adaptation. However, the "additional information" carried by traditional semantic routing typically consists of general network attributes and does not incorporate service-specific semantic dimensions, making it unable to support precise mapping from service demands to resource capabilities. Furthermore, traditional semantic routing architectures are designed around data transmission path optimization rather than service resource scheduling, and are not suited to short-duration, high-frequency, heterogeneity-compatible workloads. In summary, existing technologies suffer from identifier schemes that lack semantic expressiveness, routing decisions that are not semantics-driven, and insufficient adaptability and cross-domain coordination capability. They cannot meet the requirements of the IoA for service heterogeneity, dynamism, and high-frequency coordination. There is therefore an urgent need to construct a complete identification and routing framework for the IoA that integrates semantic understanding, heterogeneous compatibility, dynamic adaptation, and efficient response, enabling efficient Wang, et al. Expires 20 May 2027 [Page 6] Internet-Draft Service Semantic Routing for IoA November 2026 inter-agent coordination and optimized resource scheduling. The remainder of this paper focuses on this core requirement, analyzing various service semantic identification and routing approaches in depth, and deriving customized service semantic identification and routing schemes for the IoA based on the analysis and the characteristic fit of each approach. 3. State of the Art in Service Semantic Identification and Routing 3.1. Overview of Service Semantic Identification and Routing Service semantic routing is an evolutionary advancement of traditional semantic routing. Its core innovation is embedding service semantic information — such as service type and QoS requirements — directly into packets, transforming the generic auxiliary information of traditional routing into service-specific attributes tightly bound to business logic, and encoding service demands as routable identifiers so that the routing system can directly interpret business requirements. This represents a fundamental shift in routing decisions from location-based addressing to service-demand-driven forwarding. The objective is efficient and precise service-direct delivery; the architecture can be flexibly adapted to deployment requirements and business scenarios, yielding a variety of implementation approaches. Service semantic routing constructs a unified semantic identifier framework that standardizes the description of interfaces and quality differences across heterogeneous service resources, enabling one-request-multiple-fit and resolving the core tension between service heterogeneity and coordination efficiency. It relies on dynamic synchronization with a service registry to support resource state awareness and precise instance scheduling, with low parsing and scheduling overhead to meet the millisecond-level response requirements of short-duration, high-frequency interactions. The service semantic identifier is the foundational enabler. Its multi-dimensional semantic properties can precisely encode agent business demands and capability parameters, accommodating the coordination characteristics of agents — imprecise demands, service heterogeneity, and dynamic state — and facilitating dynamic scheduling and cross-domain coordination in the IoA while reducing the barrier to heterogeneous collaboration. Existing solutions are largely focused on general-purpose compute scheduling or traditional service scenarios, and a customized technical framework tailored to the IoA has yet to emerge; scenario-specific development is the key to overcoming current adaptation bottlenecks. Wang, et al. Expires 20 May 2027 [Page 7] Internet-Draft Service Semantic Routing for IoA November 2026 3.2. Four Representative Categories of Service Semantic Identification and Routing Service semantic routing techniques in current industry practice primarily differentiate along the dimension of the carrier and parsing logic for semantic information, covering four major directions: fuzzy search based on resource tag matching, DNS-based semantic encoding, anycast identification using data plane network IP, and application-layer URL-based service semantic identification. Each approach is suited to different business scenarios and precision requirements, together forming a multi-layer, full-scenario service semantic routing framework. The four approaches are described in turn below. 3.2.1. Fuzzy Search Fuzzy search is a service and resource retrieval technique based on tag matching and flexible constraint satisfaction. It does not require users to specify precise parameters; instead, it parses user-supplied constraints — such as "domestically produced," "throughput ≥ 10 Gbit/s," or "memory ≥ 64 GB" — and performs similarity computation and semantic association against predefined multi-dimensional resource tags (e.g., vendor, network throughput, memory capacity, storage type) to rapidly filter candidate services or resources from a large heterogeneous pool that satisfy the user's core requirements. In this paper, services are treated as a category of resource, ultimately supporting automated service and resource provisioning decisions in data centers. The core value of fuzzy search lies in lowering the operational barrier for users, accommodating the imprecision of demands and the heterogeneity of resources, and improving resource matching efficiency. The fuzzy search mechanism is illustrated in Figure 2. Wang, et al. Expires 20 May 2027 [Page 8] Internet-Draft Service Semantic Routing for IoA November 2026 \`\`\`text Fuzzy Search Information Collection, Crawler | v +--------+ Query +----------------------------------------------------------------------------------------+ +-------------+ | |---------------->| |---->| Terminal 1 | | User | | Fuzzy Search Platform | | |<----------------| |---->| Terminal 2 | +--------+ Return URL | +----------------+ +-----------------------+ +---------------------+ | ^ | | Accept User | | Retrieval Algorithm | | Registry List |---->| | | | Request |------->| Fuzzy Matching |<-------| List of Content & | | Terminal 3 | | | +----------------+ +-----------------------+ | URLs | | | +---------------------+ | +----------------------------------------------------------------------------------------+ | ^ | | +-----------------------------------------------------------------------------+ Figure 2: Fuzzy Search Mechanism ``` In the fuzzy search mechanism, the platform first collects multi-dimensional attributes and access indices — such as URLs and query protocols — from various resources, and constructs a unified information resource repository through normalization. User requests expressed in non-standardized natural language are then parsed using natural language processing techniques to extract core requirements and constraints, which are mapped to a standard tag vocabulary and transformed into structured queries. Flexible matching is then performed against the resource repository: text attributes are evaluated using semantic similarity algorithms, numerical constraints are handled by range matching, and layered weight scoring is applied to rank and select the top-N optimal resources. Finally, the access method for the selected resources is returned to the user, while the platform monitors access status, collects feedback, and updates resource attributes and service state to close the fuzzy search routing loop. 3.2.2. DNS-Based Precise Naming DNS-based precise naming is a semantic routing technique that encodes core service semantics — such as service type, functional scenario, QoS requirements, and resource attributes — into standardized domain names. Its defining characteristic is the one-to-one correspondence between a domain name and a specific service: the domain name itself carries the business requirements, enabling DNS to go beyond Wang, et al. Expires 20 May 2027 [Page 9] Internet-Draft Service Semantic Routing for IoA November 2026 traditional "domain name → IP" location addressing and directly achieve precise mapping from "service demand → target service." The mechanism is illustrated in Figure 3. \`\`\`text +---------+ +------------------------| DNS |<--------------------------+ | http://mypaSERVICE +---------+ | | 203.0.113.0 | | | | v v v +------------------+ +--------+ +-------------+ | Service Semantic | | Client |<------------ | Global Net |----------------> | Routing | +--------+ +-------------+ +------------------+ Figure 3: Precise Naming DNS Mechanism ``` In this approach, the system first establishes a unified semantic encoding rule set, embeds core service semantic fields into domain names to form a standardized semantic domain name scheme, and builds a global semantic domain name-to-service mapping table. Service providers register their semantic domain names, service instance IPs, and attribute metadata with the DNS server via Dynamic DNS (DDNS), with automatic updates upon instance changes to maintain dynamic consistency of the mapping. The DNS server, acting as the control plane core, receives a user's semantic domain name query, decomposes and parses the semantic fields in the domain name for validation, performs A/AAAA record lookup to match the optimal target server IP, and returns it to the requester. The user encapsulates the request packet destined for that IP and sends it to the target server, where a web service proxy forwards it to the corresponding service instance for processing; the result is returned along the same path. Access status is collected and fed back to the DNS server to optimize subsequent selection policies, forming a complete closed-loop workflow. 3.2.3. Anycast-Based Data Plane Network Identification The anycast-based data plane network identification mechanism, along with the application-layer URL semantic identification mechanism, is illustrated in Figure 4. Anycast-based service-layer network identification is a network identification and routing technique centered on semantic anycast IP addresses. Its core logic is "same service, same IP": services that are functionally equivalent and semantically identical are assigned a single dedicated anycast IP address. This IP serves not only as the unified addressing entry Wang, et al. Expires 20 May 2027 [Page 10] Internet-Draft Service Semantic Routing for IoA November 2026 point for anycast routing but also carries explicit business semantics, establishing a strong one-to-one mapping of "one IP = one service category." This approach overcomes the limitation of traditional IP, which identifies only physical nodes, elevating IP addresses from location-oriented to business-semantics-oriented. Clients need not be aware of the physical location, real IP, or deployment count of service instances; they simply direct requests to the semantic anycast IP, and the network automatically schedules the request to the optimal service instance via anycast routing. This is applicable to scenarios requiring flexible scheduling of equivalent services, such as compute-network convergence, cloud services, and distributed systems, as shown by Client 2 in Figure 4. ```text http://IP/ai/chat/ <--------------------------+ Client 1 | \ Service Instance 1 \ Request / \ / v / +-------------------------------+ / | Control Plane | v | | | +-----------------------+ | < ------ Service Instance 2 | | Business Logic | | | | [ ] * [ ] [ ] | | ^ | +-----------------------+ | \ | <------- Response | \ +-------------------------------+ \ ^ Service Instance 3 / | / | Client 2 (IP) Business Plane | <----------------------------------+ Figure 4: Mechanism of Business Plane IP Identifier & Application Layer URL Service Semantic Identifier ``` In the anycast IP-based service-layer routing scheme, the system first establishes binding rules between service categories and their dedicated anycast IPs, associating semantically equivalent service clusters with a single anycast IP that carries the business semantics, and constructs a global mapping table. All service instance information associated with each anycast IP is then collected via standardized protocols and synchronized to the service registry, enabling dynamic association between anycast IPs and instances with automatic updates upon instance changes. When the gateway receives a client request, it extracts the destination Wang, et al. Expires 20 May 2027 [Page 11] Internet-Draft Service Semantic Routing for IoA November 2026 anycast IP and forwards it to the parsing module, which retrieves service type, compute requirements, and other information and passes them to the scheduling module. The scheduling module queries the registry based on the parsed results, evaluates real-time QoS metrics and operational status of candidate instances, selects the optimal instance, and generates a scheduling plan containing the real IP and an SRv6 path, which is returned to the gateway. The gateway encapsulates the request packet and forwards it along the specified path; the server-side gateway parses the packet and delivers it to the target service instance. The service instance executes the task and generates the result, which is returned to the client along the reverse SRv6 path, completing the service loop. 3.2.4. Application-Layer URL-Based Service Semantic Identification Application-layer URL-based service semantic routing is an application-layer routing technique that uses the URL as the semantic carrier and HTTP as the transport protocol. Its core logic is to encode service semantics into a URL field — such as the URL-PATH or query parameters — and to use HTTP header fields for auxiliary information, so that the URL functions not merely as a resource locator but as a structured description of service requirements. The client sends an HTTP request carrying a semantic URL; the network side parses the semantic information in the URL and the header fields to achieve precise routing from "service demand → target service instance," without relying on underlying IP addresses or network-layer configuration — semantic-driven service scheduling is performed entirely at the application layer. Whereas a traditional URL answers only "which resource to access," this technique additionally encodes "what type of service, what functionality, and which version," so that the HTTP request itself carries the information needed for routing decisions, as shown by Client 1 in Figure 4. In application-layer URL-based service semantic routing, the system predefines a unified encoding rule set that embeds core service semantic fields — service type, compute specifications, version information, etc. — into URLs, establishing a "one service category per dedicated URL" mapping. The control plane collects, via standardized protocols, all service instance information associated with each URL — including real IP, deployment location, resource configuration, QoS capability, and supported semantic tags — and synchronizes this information to the service registry. When the client-side gateway receives an HTTP request carrying a semantic URL, it first parses the semantic fields in the URL and query parameters to extract core characteristics such as service type, compute requirements, and version constraints, then parses the HTTP header fields, and integrates the results into a structured request. Based Wang, et al. Expires 20 May 2027 [Page 12] Internet-Draft Service Semantic Routing for IoA November 2026 on this structured information, the gateway queries the control plane service registry to filter candidate instances that fully match the URL semantics and header constraints, and then evaluates real-time instance state — load, availability — and QoS metrics — response latency, throughput — to select the optimal service instance, forming the forwarding target. The gateway encapsulates the HTTP request packet and forwards it to the address of the optimal instance. The target instance executes the service task, generates the result, encapsulates the response according to the HTTP protocol with the source address of the original request, and returns it to the client-side gateway, which verifies response integrity and forwards it to the client. 4. Service Semantic Identification and Routing for the Internet of Agents In the collaborative paradigm of the IoA, service semantics refers to the business intent expressed by an agent in order to accomplish a specific task. Its core is the target service the agent seeks to discover and invoke — such as DeepSeek LLM inference or face recognition — with the goal of achieving precise demand-to-resource matching. A complete service semantic description comprises two layers: the "functional intent," which identifies the core capability and category of the target service and can be expressed via a service name, model ID, or API identifier; and the "enhancement constraints," which define delivery quality, performance, and contextual requirements such as compute specifications, latency, cost, and session context. Based on the specificity of the description, service semantics can be classified as either precise (with a well-defined functional intent) or fuzzy (lacking a clear functional intent but carrying certain constraint conditions), a distinction that directly influences the choice of routing mechanism. In technical implementation, service semantics must be transformed into computable, routable identifiers; common representation forms include structured tag sets, key-value collections, and encoded information embedded in URL fields or dedicated protocol fields. The four routing mechanisms examined in this section differ fundamentally in how they carry service semantic information, the granularity at which it is parsed, and the matching strategy applied. 4.1. Fuzzy Search for the Internet of Agents The fuzzy search approach for the IoA constructs a closed-loop mechanism of "demand parsing → semantic matching → dynamic binding," enabling efficient and precise mapping of an agent's imprecise business requirements — such as "I need a fast text-to-image service that handles Chinese" — to capability-compatible, state-compliant agent instances or service chains, without requiring precise Wang, et al. Expires 20 May 2027 [Page 13] Internet-Draft Service Semantic Routing for IoA November 2026 technical parameters. This realizes the initial connectivity discovery from business demand to resource entity. The mechanism is illustrated in Figure 5 and proceeds in four steps. ```text +---------------------------+ +---------------------------+ | User Input Agent Query Req| | Global Agent Resource Pool| +-------------+-------------+ +-------------+-------------+ | | v v +---------------------------+ +---------------------------+ | Intent Understanding Mod | | Capability Tag Vector, | +-------------+-------------+ | Trust Domain | | +-------------+-------------+ v | +---------------------------+ | | Structured Query Vector | | +-------------+-------------+ | | v | +---------------------------+ +------> | Vector Database | +-------------+-------------+ | v +-----------------------------------------------------------+ | Preliminary Filter: Semantic Similarity via Capability | | Tags & Trust Domain Compliance | +-------------------------+---------------------------------+ | +----------------+----------------+ | | v (Complex Intent) | (Simple Intent) +--------------------------------+ | | Refinement & Composition: | | | Auto Compose Service Chain | | +---------------+----------------+ | | | | | v | +---------------------------+ | | Return Agent List |<---+ +---------------------------+ Figure 5: Fuzzy Search Mechanism for Agent-Oriented Internet ``` Step 1: Unified registration of agent semantic identifiers. To enable precise business matching, standardized semantic descriptions must be constructed for all schedulable agents in the network. Agent Wang, et al. Expires 20 May 2027 [Page 14] Internet-Draft Service Semantic Routing for IoA November 2026 semantic descriptions and multi-dimensional attributes are registered actively or collected via crawling through standardized protocols, forming a semantic registry that covers capability attributes (core functionality), performance and state attributes (load rate, response latency), and trust and collaboration attributes (trust domain, communication protocol), constituting a globally maintained, dynamically updated agent resource repository. Step 2: Parsing and structured encoding of fuzzy demands. When a leading agent or user submits a fuzzy collaboration request in natural language or similar form, the semantic parsing module performs tokenization, disambiguation, and entity recognition on the raw request to extract core task atoms and implicit constraints. The encoding module then converts these into a structured query vector — encompassing capability requirements, QoS constraints, data formats, and session context — according to predefined service semantic templates, serving as the basis for subsequent matching. Step 3: Multi-dimensional semantic matching and candidate set generation. The matching module retrieves and ranks the global agent resource repository upon receiving the structured query vector. It first applies preliminary filtering based on capability tag semantic similarity, task match degree, and trust domain compliance to form an initial candidate set; for complex requirements, it automatically composes and validates service chains. A multi-objective scoring model then comprehensively evaluates each candidate's semantic match score, real-time load, historical success rate, and network overhead, outputting a priority-ranked optimal match list. For text tag matching, cosine similarity based on bag-of-words models such as TF-IDF, or semantic vector comparison using pretrained language models, is commonly used. For numerical QoS constraints, range matching or threshold filtering algorithms are typically applied. In generating the overall ranking, a common engineering practice is to construct a multi-attribute weighted scoring function that integrates semantic match score, real-time load, historical success rate, and network overhead into a final priority-ranked output. Step 4: Result feedback and client-side dynamic binding. The platform returns to the requester an optimal match list containing access information such as IP:Port and service identifiers. This step constitutes a recommendation rather than mandatory routing; the requester may independently select the target, initiate the connection, and configure communication parameters based on its own policies. The platform additionally provides dynamic binding assistance: if performance degradation of the bound agent is detected during task execution, an alert is triggered and retargeting is recommended, with the final switching decision left to the requester, balancing flexibility with service elasticity and robustness. Wang, et al. Expires 20 May 2027 [Page 15] Internet-Draft Service Semantic Routing for IoA November 2026 By combining natural language parsing with flexible tag matching, this technique precisely accommodates the fuzzy service demand characteristics of the IoA and provides efficient resource discovery for exploratory inter-agent collaboration, substantially lowering the barrier to initial coordination among heterogeneous agents. It is best suited to initial collaboration scenarios with imprecise and exploratory business requirements. Its technical characteristics complement those of the anycast approach and the application-layer URL approach, addressing the problem of resource discovery when the precise service name or address is unknown, and providing the discovery entry point for subsequent high-efficiency invocation via more precise routing mechanisms. 4.2. DNS-Based Precise Naming for the Internet of Agents In the IoA, DNS-based precise naming deeply accommodates the "business-driven, service-heterogeneous, dynamically collaborative" characteristics by extending the "one domain name per service" principle into a dynamic scheduling mechanism for "semantic service demand → optimal agent instance." It upgrades traditional "location-addressing" DNS into a routing hub that integrates "semantic addressing and dynamic scheduling," providing a standardized, high-throughput semantic addressing entry point for large numbers of heterogeneous agents. The implementation proceeds in four steps. Step 1: Semantic encoding rule definition and agent service registration. A network-wide unified semantic domain name encoding specification is established, defining a standardized mapping from service semantics to domain name fields and specifying encoding formats for core fields such as service capability, QoS requirements, and version information to form structured domain names. When service-capable agents join the network, they register with the coordination platform via an enhanced DDNS protocol, providing the reachable IP address and semantic attribute metadata including capability descriptions, performance thresholds, and SLA parameters. The global registry establishes precise bindings between "semantic domain name — attribute set — dynamic instance list," with automatic DDNS updates upon instance changes. The semantic domain name encoding rules for agent-oriented DNS are illustrated in Figure 6. Wang, et al. Expires 20 May 2027 [Page 16] Internet-Draft Service Semantic Routing for IoA November 2026 ```text +-----------------------------------------------------------------+ | | | service-performance-version.trust-domain | | |------| |----------| |------| |----------| | | Service Performance Version Trust Domain | | Capability | | | | Example: speech-tts-lowlatency-v2.agent-net | | | +-----------------------------------------------------------------+ Figure 6: Precise Naming DNS Domain Encoding Rule for Agents ``` Step 2: Request reception and domain name semantic parsing. Users or leading agents can initiate standard DNS query requests carrying semantic domain names representing the desired service, without prior knowledge of instance addresses. The enhanced DNS server, acting as the control plane core, receives the request and decomposes the domain name according to predefined rules, parsing and extracting key constraints such as service type and quality requirements. Step 3: Dynamic instance matching and optimal IP selection. After completing semantic parsing, the DNS server queries the global registry to retrieve the list of matching online, normally loaded service instances. Through a built-in dynamic scheduling module, it integrates real-time data on instance load, network latency, and historical success rate, applies a scheduling algorithm to select the current optimal instance, and returns the real IP to the requester via a DNS response. Step 4: Direct client access and service loop optimization. The requester uses the returned IP as the destination address, encapsulates the service request packet, establishes a network connection, and completes the service invocation with the target agent instance. Data plane interaction is conducted autonomously between the two parties, with no further DNS involvement. By providing a unified semantic domain name encoding framework, this technique enables precise matching of agent service demands to resources, accommodating the core IoA requirement for a unified semantic identifier scheme. It provides standardized, highly reliable addressing and scheduling for large-scale coordination of heterogeneous agents across domains, deeply integrating semantic identification with network scheduling. Its technical characteristics complement those of the anycast approach and the Wang, et al. Expires 20 May 2027 [Page 17] Internet-Draft Service Semantic Routing for IoA November 2026 application-layer URL approach, providing a standardized, high-performance discovery mechanism for scenarios in which the semantic requirement is known but the instance location is not. It serves as a foundational routing solution for large-scale service scheduling in the IoA. 4.3. Anycast Identification and Routing for the Internet of Agents The anycast identification and routing approach for the IoA associates equivalent functional agent services through a unified anycast identifier, integrating capability awareness with dynamic scheduling to enable efficient communication, precise discovery, optimal routing, and reliable forwarding among agents. The mechanism is illustrated in Figure 7 and proceeds as follows. Wang, et al. Expires 20 May 2027 [Page 18] Internet-Draft Service Semantic Routing for IoA November 2026 ```text Receive service instance info | v +-----------------------------------------------------------------------------------------------------------+ | Control Plane: Agent Service Orchestration | | | | +--------------------------------+ +----------------------------------------+ | | | Agent Demand Management | | Agent Service Identifier | | | | v | | | | | | (Compute) |<----------------------->| +----------------+ +----------------+ | | | | v | | |Terminal & Subnet| |Service Instance| | | | | Service Instance Table | | |Management | |Management | | | | | v | | +----------------+ +----------------+ | | | | (Compute) |<----------------------->| v | | | | v | | SID Management | | | | Policy Management | | v | | | +--------------------------------+ | Gateway Management | | | | +----------------------------------------+ | | | | | | | Network State Management <-------------+ | +-----------------------------------------------------------------------------------------------------------+ | v +-----------------------------------------------------------------------------------------------------------+ | Business Plane | | | | Intelligent Agent -----> +------------------------+ +------------------------+ | | | Protocol Gateway | | | | | Smart Device ----------> | +----------------+ | | Agent Service Instance/AGENT | | | | Request Parse | | | Agent Service Instance/Tool | | | +--------------+ | | Agent Service Instance/LLM | | | | | Agent Service Instance/Knowledge DB | | | +----------------+ | | Computing Service Instance | | | | Policy Execute| | | Computing Service Instance | | | +--------------+ | | | | +------------------------+ +------------------------+ | | ^ v | +------------------------+ | | Network Controller |---------------------------+ +------------------------+ Figure 7: Anycast Identifier and Routing Method for Agent-Oriented Internet ``` Wang, et al. Expires 20 May 2027 [Page 19] Internet-Draft Service Semantic Routing for IoA November 2026 Step 1: Anycast identifier definition and assignment. Standardized mapping rules are established, binding semantically equivalent agent service clusters to dedicated anycast IPs, achieving deep integration of service semantics with agent identification. Step 2: Agent service instance registration. The system collects and registers the full set of agent instance information associated with each anycast IP via standardized protocols, including real IP, deployment location, compute configuration, response latency, load state, and QoS support capability, enabling dynamic association between anycast IPs and instances. Upon instance scale-out, migration, deregistration, or upgrade, the system automatically triggers a registry update to ensure that the anycast IP is always associated with valid instances. Step 3: Request reception and semantic parsing. The initiating agent directs its request to the corresponding semantic anycast IP without needing to know the target instance location or real IP. The gateway receives the packet, extracts the destination IP, and forwards it to the parsing module. The parsing module decodes the business semantics using the global mapping table, identifies the service type, compute requirements, and request core content, and validates the initiating agent's authorization and QoS requirements. Step 4: Network scheduling and optimal instance selection. The scheduling module queries the registry based on the parsed results, first filtering candidate instances that satisfy the semantic, compute, and baseline QoS requirements, then performing comprehensive evaluation of real-time monitoring data — load rate, latency, bandwidth, and availability — using a weighted algorithm to select the optimal instance (e.g., Instance A) and generate a scheduling plan containing the instance's real IP (IP1(a)) and an SRv6 path. Step 5: Path establishment and task forwarding. The initiating gateway processes the packet according to the scheduling plan, embeds the SRv6 path identifier, and forwards it along the specified path. Intermediate routing nodes perform fast forwarding based on the path identifier without re-parsing the anycast IP. The gateway at the target Instance A resolves the packet via the local Segment Identifier (SID), confirms attribution, and delivers the packet to Instance A. Step 6: Result generation and reverse return. Instance A executes the service task — such as data computation or collaborative decision-making — and generates the response. The gateway maintains path consistency, ensuring that the response is returned accurately and efficiently to the initiating party, completing the full anycast communication loop. Wang, et al. Expires 20 May 2027 [Page 20] Internet-Draft Service Semantic Routing for IoA November 2026 By leveraging semantic anycast IPs to achieve deep integration of business logic with the network, associating equivalent agent services through dedicated anycast IPs, sensing service instance operational state in real time, and performing dynamic scheduling during data plane packet forwarding to realize agent service semantic routing, this technique belongs to the same IoA semantic routing framework as the approaches in previous sections, and complements them in enriching the range of applicable routing scenarios. Its core role is to support agent coordination at scale under dynamic load conditions, providing reliable routing support for efficient communication among heterogeneous agents. 4.3.1. Application-Layer Service Semantic Identification and Routing Based on URLs for the Agent Internet The proposed application-layer semantic identification scheme for the Agent Internet is designed around two objectives: precise semantic service representation and efficient routing. The scheme accommodates heterogeneous protocols used in agent collaboration environments. Service semantics are embedded into protocol-native fields or key business fields, enabling a direct mapping between service requirements and computing resources. This approach supports flexible and efficient routing decisions. For agent interactions with Large Language Model (LLM) APIs, a two-level optional identification mechanism can be employed. The URL path serves as the primary identifier, where structured path encoding (e.g., "/agent/model/gpt-4-v/v1") conveys core semantic information such as model category and version. Gateways can extract these semantics directly from the URL without inspecting the message payload, making the approach suitable for high-frequency and low-latency agent interactions. However, URL-path parsing can represent only coarse-grained semantics such as model categories and versions. For LLM invocation scenarios that require finer-grained semantic descriptions, URL-level information alone is insufficient for accurate service selection. Additional payload-level semantics are therefore required. One possible enhancement is to use the "model" field in the JavaScript Object Notation (JSON) payload as a secondary identifier. This field can directly represent detailed model attributes such as "llama3-70b-fp16". By parsing the model field, gateways can obtain more precise model information and attributes. This paper uses a representative Agent-to-Agent (A2A) protocol as an example. Under the A2A protocol, URLs remain the primary carrier of service semantics. In addition, session affinity must be considered. Fields such as contextId and taskId can be used to determine whether Wang, et al. Expires 20 May 2027 [Page 21] Internet-Draft Service Semantic Routing for IoA November 2026 a request belongs to an existing session or is associated with a previously established task. The application-layer semantic routing architecture for the Agent Internet is illustrated in Figure 8 and operates as follows. ```text +-------+--------+ | Agent | +-------------+ +----------------+ | Controller | | +------+------+ | / / | \ | \ / / | \ | \ v v v v v v +----------+ +----------+ +----------+ +----------+ +----------+ -------+--------+ | Router | | Router | | Router | | Router | | Router |<---| Agent | +----------+ +----------+ +----------+ +----------+ +----------+ +----------------+ | | | │ | | +-----------+------------+---------│--+-----------+ (Cloud Network Fabric) | | v v +-------------+ +-------------+ | Agent | | Agent | +-------------+ +-------------+ | +-----------------------+|+-----------------------+ | | | +--------+--------+ +--------+--------+ +-------+--------+ | Large Model | | Tools | | Knowledge Base | +-----------------+ +------------------+ +----------------+ Figure 8: Application Layer Service Semantic Routing Architecture for Agent-Oriented Internet ``` Step 1: Agent Registration and Mapping Table Construction Before participating in A2A-based service interactions, an agent registers its service information with the network system. The registration process primarily involves publishing an AgentCard that contains the agent's service endpoint URL and supported capabilities. Based on the registration information, the network system maintains a mapping table between semantic service identifiers and active agent instances. The mapping table supports dynamic updates and stale-entry removal to ensure accuracy and timeliness. Wang, et al. Expires 20 May 2027 [Page 22] Internet-Draft Service Semantic Routing for IoA November 2026 Step 2: Request Reception and Service Requirement Extraction All service requests initiated by agents are directed to the application-layer semantic routing gateway. The gateway receives the request, performs protocol-compliance validation, and parses the message content. Information extracted includes the URL from the HTTP layer and Task and Message objects from the A2A layer. Relevant fields such as contextId and taskId are also extracted for subsequent routing decisions. Step 3: Determination of Routing Requirements Based on the extracted fields, the gateway determines whether the request belongs to an existing stateful interaction, represents a new task, or continues a previously established task. For task-continuation requests, the gateway uses taskId, contextId, or related identifiers to locate the previously selected backend service instance. In such cases, the request is forwarded directly without invoking a new routing procedure. Otherwise, the full semantic-routing process is initiated. Step 4: Semantic Routing for New Tasks For stateless requests or newly created tasks that do not require session affinity, the gateway identifies candidate agents according to the service semantics represented by the URL. It then selects the optimal target agent according to predefined routing policies. The request is forwarded to the selected agent, and the mapping between the task and the chosen backend instance is cached to maintain continuity for subsequent requests. Step 5: Response Generation and Delivery After receiving the request, the target service agent executes the requested task and generates a response. The result is encapsulated according to the A2A protocol and returned to the client-side gateway together with the original source-address information. After validating response integrity, the gateway forwards the response to the requesting client. Using the A2A protocol as an example, this section presents an application-layer service semantic identification and routing mechanism. Through cooperation between control-plane functions and data-plane gateways, an independent semantic-routing plane is established between the network layer and the application layer. The architecture decouples agent services from network addresses while supporting centralized registration, Layer-7 (L7) semantic parsing and scheduling, and Quality of Service (QoS) enforcement. As a result, fine-grained cross-layer and cross-domain semantic optimization can be achieved without affecting the agent interaction experience. Wang, et al. Expires 20 May 2027 [Page 23] Internet-Draft Service Semantic Routing for IoA November 2026 To ensure feasibility and robustness, several engineering considerations must also be addressed. For example, session affinity requires an efficient cache that maintains mappings between context identifiers and service instances. Cache expiration may be implemented through aging mechanisms or integrated with instance health monitoring. In the event of routing conflicts, such as the same session identifier being associated with multiple service instances, predefined arbitration policies (e.g., first-match or best-match selection) should be applied. The specific design of these mechanisms depends on the system's reliability, consistency, and performance requirements. By embedding service semantics directly into application-layer protocol fields, the proposed approach enables task-level semantic scheduling and routing. The routing granularity can be reduced to individual service requests, allowing the architecture to adapt to the highly dynamic nature of Agent Internet services. The scheme is compatible with heterogeneous protocols and provides a flexible and extensible routing framework for high-frequency and short-lived agent interactions across distributed computing-network environments. Using URLs as semantic carriers and leveraging A2A fields such as contextId and taskId for session-affinity management, the approach achieves deep integration between service semantics and network routing. It complements the mechanisms described in previous sections and can be deployed in combination with them to further enhance routing support in cross-domain computing-network collaboration scenarios. For other agent interoperability protocols, such as the Agent Network Protocol (ANP) and the Model Context Protocol (MCP), application-layer semantic identification and routing mechanisms can be defined according to protocol-specific characteristics while following the same architectural principles. To enable semantic interoperability and coordinated scheduling across vendors and protocols, a unified service-semantic metadata framework and protocol adaptation gateways may be introduced.Specifically, the control plane may define a common semantic description framework that maps protocol-specific elements, such as contextId in A2A, session identifiers in ANP, and resource names in MCP, into a common set of abstract service-routing attributes representing service types, capabilities, and session contexts. Data-plane gateways can then incorporate protocol-specific parsing modules to perform semantic extraction and translation during forwarding. This approach ensures that requests originating from different protocols are processed using a consistent semantic model and routing policy framework, thereby enabling accurate service matching and routing while preserving protocol heterogeneity. Detailed implementation mechanisms are outside the scope of this paper. Wang, et al. Expires 20 May 2027 [Page 24] Internet-Draft Service Semantic Routing for IoA November 2026 5. Security Considerations Semantic-routing-enabled agent-to-agent interactions bring new security risks including semantic identifier spoofing, fake service registration, routing hijacking, and service deception. Mechanisms for semantic identifier authentication and integrity validation, trust-domain isolation, and access-control policies are required. Malicious agents may forge service semantic tags to attract requests; registration validation and runtime anomaly detection should be deployed. Privacy-preserving matching techniques are needed to avoid disclosure of agent task intent and capability metadata during routing procedures. 6. IANA Considerations This document does not require any IANA-assigned code points or registry actions. Authors' Addresses Shiji Wang China Unicom Email: wangsj7236@chinaunicom.cn Liwen Wang China Unicom Email: wanglw97@chinaunicom.cn Yan Zhang China Unicom Email: zhangy1156@chinaunicom.cn Lijun Ma China Unicom Email: malj105@chinaunicom.cn Yunfei Cao China Unicom Email: caoyf34@chinaunicom.cn Yongsheng Liu China Unicom Email: liuys170@chinaunicom.cn Wang, et al. Expires 20 May 2027 [Page 25]