openEuler ThinkPro: the agent runtime moves into the OS

Every agent framework you use today reinvents the same three things: process state, memory, and scheduling. Those are operating system jobs. At HUAWEI CONNECT 2026 in Shanghai, Huawei said the quiet part out loud and put an agent runtime primitive into openEuler itself. Whatever you think of the vendor, the architectural direction is the interesting part — and it has a direct consequence for how portable your automation is.

What actually happened

Zhu Zhaosheng, Huawei's Chief Strategy Officer for Computing, presented on September 19. The headline piece is ThinkProcess — ThinkPro — introduced in openEuler as what Huawei calls the atomic unit of agent thinking, execution, exploration and evolution. In plain terms: a first-class OS object for an agent's work, with atomic-level APIs in both user space and kernel space that abstract state management and hardware resources like CPU and memory.

Read that again. Not a Python library. Not a framework. A process abstraction in the kernel and its userland, with APIs developers call directly.

The rest of the announcement is silicon-side. Huawei opened its tile-based PTO instruction set — more than 120 virtual instructions across eight categories, with Auto and Manual optimization modes and cross-generation operator compatibility. CANN upgraded Ascend C for the Ascend 950 generation with SIMD plus SIMT programming. On the developer-access side: a 100 NPU-Hour Program granting baseline compute, shared clusters at 10k-NPU scale, and a stated CNY 5 billion commitment over three years. Huawei also open-sourced CANNBot, Model Agent, MindStudio Agent and Solution Agent.

Zhu's framing was to "give developers an open foundation, the freedom to build on their own terms, and the space to innovate."

Why an OS-level agent runtime matters for your business

Almost none of you will run Ascend. Export controls, supply, and support make that a non-question for a Phoenix small business. We are not recommending the hardware. We are reading the roadmap.

The abstraction layer is where lock-in gets decided. Today your agents are processes your framework babysits — LangGraph, the Agents SDK, whatever you picked. When the OS starts offering state management, scheduling and resource limits natively, frameworks become thin wrappers, and the wrapper you chose stops mattering much. That is good news, as long as your business logic does not live inside the wrapper.

So separate the two now. Your prompts, tool definitions, approval rules and evaluation set are assets. The runtime that executes them is plumbing. If you cannot move an agent from one harness to another in an afternoon, you did not build an agent — you bought a subscription with extra steps. Keep tools behind MCP or plain HTTP, keep prompts in version control, keep the eval harness independent of the runner.

Kernel-level agent state cuts both ways. Better resource isolation and real scheduling would fix genuine problems — runaway token spend, agents starving each other, no clean way to cap memory. It also means the OS is now in your agent's trust boundary. Anyone shipping this should ask what a compromised ThinkProcess can reach before they ask how fast it is.

Watch whether Linux upstream follows. That is the signal worth tracking. If agent-process primitives show up in mainline kernels or in systemd, the pattern is real and vendor-neutral, and your hosting choice stops being a strategic decision.

Sources: Huawei, StorageNewsletter.