AXP gives Chinese infrastructure providers a reported benchmark for coordinating AI agents at scale, bringing network reliability into enterprise agent planning.
Purple Mountain Laboratories has unveiled AXP, a network platform intended to connect AI agents, or software systems that carry out tasks, across different systems. The October 16 release at Nanjing’s 10th Future Network Development Conference also involved China Mobile, Jiangsu Future Network Group and Jiangsu Future Networks Innovation Institute. Xinhua reports that the four organizations designed AXP to translate protocols among heterogeneous agents while scheduling computing and network resources together.
The release puts an infrastructure question alongside the rapid expansion of agent software: how systems developed by different parties can discover, address and coordinate with one another. AXP is built on deterministic networking, an approach intended to make a network’s performance more predictable. Pandaily reports that the materials give no commercial availability date or pricing, while Purple Mountain Laboratories says it has begun joint testing and industrial validation across several settings.
AXP puts agent coordination onto a deterministic network
The four-developer structure brings together Purple Mountain Laboratories, China Mobile and two organizations focused on future networks. Xinhua identifies Purple Mountain Laboratories, China Mobile, Jiangsu Future Network Group and Jiangsu Future Networks Innovation Institute as the joint developers, with AXP presented at a conference focused on the next generation of communications networks. That arrangement places the project at the intersection of research infrastructure, telecom operations and potential industrial use.
The platform’s stated function is to convert the different protocols used by heterogeneous agents and to schedule their computing and network resources as a single system. Pandaily describes deterministic networking as managing traffic within set performance bounds in a network domain instead of leaving delivery to the ordinary best-effort treatment of internet traffic. In practical terms, AXP is intended to provide a common network layer for agents that must find one another and exchange work across organizations or regions.
The resource-scheduling role is central to the developers’ pitch. An agent workflow can involve a sequence of software services, each needing a connection and computing capacity at a particular point in the task. AXP’s reported design would place protocol translation, agent discovery and joint compute-network scheduling in one infrastructure layer. Applications could draw on that layer for their connections.
The reported capacity figures describe the platform’s intended scale
The platform was released with reported support for one million registered agents and 100,000 agents online at once. Xinhua also reports claims of fault sensing and switchover within seconds, plus semantic-address responses in about 100 milliseconds. Semantic addressing means that a system can locate an agent by the function it performs, a capability that could matter when many agents must locate the right service without relying solely on a fixed network address.
The developers say that, using the national CENI future-network testbed, AXP completed high-precision agent addressing across domains more than 4,000 kilometers apart and reduced end-to-end task execution time by more than 20%. The published accounts omit the workload, comparison baseline and test method for that reduction, and they have not identified the agent protocols that AXP translates. Those details will determine how readily enterprises can compare its performance with their own agent systems.
The reported figures describe several connected operational tasks: keeping a large registry, maintaining simultaneous connections, detecting faults and locating services by their purpose. Each has a different relevance for an enterprise system. A manufacturer coordinating plant systems, for example, would need performance information for its own locations and workloads before treating the conference figures as a usable capacity measure.
Deployment disclosures will determine AXP’s commercial shape
Purple Mountain Laboratories says it has started joint testing and industrial validation with telecom operators, communications-equipment makers and a large steel manufacturer. Junction News reports that Liu Yunjie said China Mobile had deployments in 13 cities and Zhongtian Steel was using the work at three locations in Jiangsu. The reports point to early operating settings for the platform, alongside the conference’s broader technology claims.
For network operators, cloud providers and industrial users, the useful follow-up will be evidence on the protocols, workloads and operating conditions behind those deployments. Earlier EastFrontier reporting on China Telecom’s agent-compute forecast showed why widespread agent use could raise demand for distributed computing and network capacity. An AXP update that specifies how the 13-city deployments operate would provide a practical benchmark for China’s AI-agent infrastructure.