On September 21, Wu Hequan, an academician of the Chinese Academy of Engineering, attended the 2026 Beijing Internet Conference and delivered a speech. He pointed out that the evolution from large model "assistants" to AI agent "proxies" executing tasks is witnessing massive interconnection of AI agents, giving birth to the Intelligent Internet. This brings new opportunities and challenges to IPv6. The development of IPv6 and AI agents is experiencing a historic convergence, and both are entering their best window of development.
The Era of the Intelligent Internet Arrives, and AI Agent Identification Is Key
"AI Agent = Large Model + Memory System + Tool Invocation + Planning Capability." Wu Hequan stated that AI agents are usually manifested in the form of mini-programs and have currently become agents that perceive the environment, make autonomous decisions, and execute tasks. The task scope of a single AI agent is limited, and undertaking tasks often requires the combination of multiple AI agents. Two AI agents are not simply "1+1=2"; by sharing memory and tools, they can achieve an effect of "1+1≥2". When more AI agents are connected to the network, the Intelligent Internet is formed. The traditional Internet is about host interconnection and information interconnection, whereas the Intelligent Internet is about AI agent interconnection and capability interconnection.
How do users find AI agents? Wu Hequan explained that not all AI agents are deployed on the cloud, but cloud deployment is the most common form. Traditional clouds offer IaaS (Infrastructure as a Service), providing computing power, data, transmission, and cross-computing invocation; above that is PaaS, providing various tools and software; now there is MaaS offering Model as a Service, and AaaS offering Agent as a Service. Through cloud deployment, the industry has lowered the threshold for using AI agents.
Invoking tools among AI agents requires protocols. The currently widely used MCP protocol has been open-sourced, with thousands of open-source AI agents connected to data sources, making the ecosystem relatively mature. The typical process is that after the host starts, it queries tools and resources on multiple MCP service providers and servers via MCP, plans tasks according to customer requirements, and returns results.
AI agents also need to communicate with each other. Generally, they need to obtain the business card of the counterpart AI agent, send a contact request, start the task in the instance, manage and control the invoked tools and the entire process, and revoke it after the task is completed. AI agents can share tools and memory, but in principle, they are not required to expose which large model the other party uses or how it thinks, thus maintaining privacy. Mutual access involves the issue of whether the counterpart is willing to accept it, so protocols such as authentication, privacy protection, and encryption are needed for support.
In Wu Hequan's view, AI agent identification is the identity of the AI agent. If willing to share, it needs to present its basic identity, including a globally unique identifier, human-readable name, etc. In addition, AI agent identification also includes semantic capability modules, addressing modules, A2A interaction contract modules, and extended metadata.
IPv6 and AI Agents Converge, Reconstructing the Mobile Internet
The traditional Internet is human-driven, C/S oriented, features short sessions, and uses addresses only for reachability and positioning. Wu Hequan stated that considering the massive number of terminals within a single subnet in the traditional Internet, a 2^64 address capacity was designed. However, relying on the traditional IPv6 address architecture to identify AI agents faces four major challenges.
First, AI agents on open platforms such as public clouds can have independent IPv6 addresses, while AI agent mini-programs in closed sandboxes can only share one public IP. If the NAT/gateway proxy mode is adopted, it can address AI agents inward, but this will break end-to-end reachability and increase the latency and connection failure rate of AI agent discovery/P2P collaboration; moreover, a large number of AI agents are virtual instances without hardware MAC addresses.
Second, IPv6 can allocate independent public network routable addresses for each AI agent instance/container/computing power card, but there is no need to provide subnet addresses on the scale of 2^64; on the other hand, AI agent identification needs to provide attributes required for registration/discovery/addressing, such as capability description, discovery, policy, and audit.
Third, if the IPv6 address architecture is redefined to include identifiers such as AI agent attributes in the IID, when AI agents migrate across domains (e.g., drones), it will disrupt the SLAAC (Stateless Address Autoconfiguration) ecosystem, making it impossible to decouple identity and address. Solutions that do not change the address format need to be considered.
Fourth, during addressing, the semantic description of AI agent capabilities is carried by the payload and parsed by the application layer; during routing, AI agent attributes are carried in the IPv6 Hop-by-Hop Options Header/Destination Options Header or the SRv6 SRH extension header and SID (e.g., APN6). IPv6 provides network-layer positioning and reachability, while upper-layer identification resolves stable identity; the two are complementary rather than substitutive.
Wu Hequan believes that AI agent applications will accelerate the development of IPv6 single-stack. AI agents require a massive number of routable addresses; adding translation will increase forwarding latency points, O&M complexity, and fault points, and rewriting addresses during translation will also bring troubles to management and audit. At the same time, the IPv6 single-stack transition must also take into account the existing IPv4 base. In the future, IPv6 will also support new multicast (BIER), Service Function Chaining (SFC), and Deterministic IP; it will locate AI agents by defining containers, microservices, and processes.
Overall, in the AI agent era, IPv6's massive addresses, continuous reachability, address uniqueness, immediacy, mobility, and connectionless nature are more adaptable to the needs of AI agents. Meanwhile, AI agents also promote the enhancement of IPv6 capabilities, giving rise to various AI agents in massive cloud, edge, and end devices, featuring addressing autonomy, neighbor discovery self-healing, autonomous fault root cause localization and self-healing, dedicated scheduling for IPv4/IPv6 transition, protocol layer enhancement, and global security defense.
"The development of IPv6 and AI agents has formed a historic convergence. Now is the best period for the development of AI agents, and also the best period for the development of IPv6," Wu Hequan concluded.
Managing Editor/Layout: Zhu Wenfeng
Proofreader: Mei Yaxin
Supervisor: Liu Qicheng