openEuler unveils agentic infrastructure OS with 20 ms sandbox startup, targets heterogeneous AI workloads

At the Agentic Infra OS Conference and openEuler Summit 2026, the openEuler project detailed a major architectural evolution aimed at transforming the operating system into infrastructure purpose-built for agentic AI workloads. Rather than treating agents as ordinary processes, openEuler's design introduces an Agent Kernel layer for foundational system capabilities and an Agent Service layer exposing sandboxing, memory management, session handling, and tool invocation as reusable services. The stack targets heterogeneous environments spanning x86, Arm, RISC-V, LoongArch, and multiple AI accelerators.

The architecture rests on five technical pillars: lightweight sandboxed execution via StratoVirt and the Conch runtime; multi-tier KV cache reuse through LMCache and openYuanrong for inference acceleration; persistent agent memory using OpenViking, oGMemory, and openGauss; serverless database branching on openGauss for elastic data services; and agent-aware scheduling across the stack. openEuler cites measured performance figures including 20 ms sandbox startup latency, 2,000-plus sandbox density per host, 100,000 sandboxes launched in 10 seconds, 1.5x–3x inference throughput gains, and 30%–70% reductions in both time-to-first-token and token costs.

Agentic OS architecture and the five infrastructure pillars

• Sandbox layer (StratoVirt/Conch): Combines lightweight virtualization, snapshot-based recovery, resource prewarming, and high-density scheduling to achieve 20 ms sandbox startup, 2,000-plus density, and 100,000 concurrent launches in 10 seconds.
• KV Cache subsystem (LMCache/openYuanrong): Implements multi-tier caching with global indexing, NUMA-aware placement, and cross-node transfer; agent-aware scheduling maps context to cache chunks for prefetch and eviction, yielding 1.5x–3x throughput improvement and 30%–70% TTFT reduction.
• Memory Agent (OpenViking/oGMemory/openGauss): Provides structured context extraction, hybrid vector/full-text retrieval, and cross-agent knowledge sharing, cutting token costs 30%–70% and improving task success 1.5x–3x in targeted scenarios.
• Serverless database (openGauss-based): Delivers elastic compute with copy-on-write storage, database branching, point-in-time recovery, and branch comparison/merge; instance startup is in milliseconds, branch creation and rollback in seconds.
• Heterogeneous target support: The full stack is explicitly designed for x86, Arm, RISC-V, and LoongArch general-purpose processors alongside multi-vendor AI accelerators, reflecting enterprise multi-architecture deployment realities.

Why it matters

openEuler's agentic infrastructure push represents one of the first comprehensive attempts to redesign a Linux distribution's core architecture around AI agent execution rather than merely hosting AI frameworks on a conventional OS. For Linux distribution maintainers and enterprise infrastructure teams, this establishes a reference model for how sandboxing, inference caching, persistent agent memory, and elastic data services can be integrated below the application layer. The explicit support for RISC-V and LoongArch alongside x86 and Arm also signals growing investment in non-x86 AI infrastructure, potentially accelerating toolchain and accelerator driver maturity on those architectures.