Skip to content

Tools

openEuler outlines AI agent infrastructure, with SP5 planned for late December

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. OpenEuler presented its operating system for AI agents and announced the 24.03 LTS SP5 service pack in Beijing.
  2. The project described sandboxing, inference caching, agent memory and workload security.
  3. Test builds are public. The plan calls for a December 30–31 stable release if it passes review on December 24–25.
PracticalDevelopers
Server chassis line a dark bench, with a metal tray partly extended in the foreground.
AI-generated conceptual image: Server hardware and a partly extended tray represent agent infrastructure and a service pack still being prepared. · AI Ping

The openEuler community presented an operating system for running AI agents and announced its 24.03 LTS SP5 service pack at its Beijing summit on November 13. The project described support for isolating code in sandboxes, keeping intermediate model computations in a cache, handling agent memory and securing agent workloads. Its release plan schedules the stable SP5 build for December 30–31.

Developers can find existing releases on the project's download page and SP5 test builds in a separate repository. For teams planning an upgrade to a supported stable build, the project has scheduled a release review before the end-of-December window. The package set for that release goes through its own review, which determines what teams will get in the final distribution.

Operating system support for code execution and inference

OpenEuler is targeting applications in which an agent carries out a series of steps. An app that gathers internal documents and drafts a report, for example, can move between sending requests to a model and running code. A sandbox limits how that code affects other workloads. A key-value cache holds intermediate computations from processing text, allowing a later model request to reuse some of that work.

The project says its approach builds on support for a SuperPoD, in which applications can draw on computing and memory resources across several machines. Once a workload uses more than one server, the system has to manage communication between nodes and contain failures that could otherwise spread. The project's earlier technical description discussed detecting and isolating such faults. Its summit announcement extended that approach to infrastructure that supports both model inference and the applications using it.

Some of the pieces are already listed for other releases. The existing long-term support release, SP4, includes E2B sandboxes and support for recovering inference services. The separate 26.09 innovation release, published October 1, lists monitoring for agents and cache management for inference. Its documentation describes using an agent's context to determine which cached material may be useful in a subsequent request.

The 26.09 server documentation says the cache retains a shared chunk when one session ends and another still uses it. The software records how the chunks relate to sessions and can rebuild that mapping if the cache manager restarts. That description covers an existing implementation in 26.09. The project tracks the status of SP5 features separately in the service pack's release plan.

Two black modules occupy separate compartments beside a green memory board.
AI-generated conceptual image: Separate tray compartments and a memory board represent isolated code execution and reuse of model computations. · AI Ping

A November test build and a December release review

OpenEuler's build team posted to its developer list that it would start building the fourth SP5 release candidate at noon Beijing time on November 12. The team builds it from the development branch for the long-term support line. Changes merged after the cutoff would generally miss that round of testing, the notice said. The public repository lists a directory for that release candidate alongside daily builds dated November 13 and 14. Teams reporting test results may therefore be working with different snapshots of the service pack.

The release plan describes a service pack based on the Linux 6.6 kernel for servers, clouds, edge systems and embedded devices. November tests check individual modules. In early December, the project plans to test the system as a whole and restrict changes to the release branch to bug fixes.

A tabbed notebook, clipped papers and a small circuit board sit on a light desk.
AI-generated conceptual image: Notebook tabs and a test board represent testing and preparations for the stable release after the announcement. · AI Ping

The plan schedules a review of which packages to ship for December 1–3, with the SP5 branch due to be created in the December 4–8 window. The project then lists two rounds of regression tests. Its planning document records whether proposed features are still in development, being tested or accepted, and identifies the form in which individual items would ship, such as an installation image or a separate package.

The community plans to make its release decision on December 24–25. If that review passes, the release plan calls for the stable build to go out on December 30–31.

Sources