Cyanogen creators are building KORA, a Linux OS with built-in AI agents and a local 26B-parameter model

It's a familiar pain: today, an AI agent on your phone or laptop is an app that knows nothing about neighboring apps, can't see your files as a whole, and yet demands access to everything. The creators of Cyanogen, whom many remember for their alternative Android firmware, have decided to approach the problem from a different angle: make agents part of the operating system. The project is called KORA, and for now it is more of a pitch than a product you can get your hands on.

What they promise

The announcement points to three ideas. The first is shared memory for agents. Imagine you asked one agent to sort your email while a second one puts together a report. Today you have to explain the context to each of them from scratch. If memory is shared at the OS level, the second agent already knows what the first one found and what you asked for.

The second idea is searching files by meaning. Not "report_final_v3.docx", but "that email about the contractor agreement where we discussed deadlines". It's ordinary semantic search, but baked into the system rather than a standalone app.

The third is isolated workspaces. The agent is given a separate workspace and sees only what has been put into it. This answers the main fear of anyone who has given agents access to their disk: "did the AI walk off with something extra?" The idea is sensible, but everything depends on the implementation. A sandbox is only as good as the rigor with which its boundaries are checked.

Cortex and local models

A separate component, Cortex, is supposed to decide where to run a request: on the device or in the cloud. The logic is clear. Simple tasks (search, short answers, document parsing) can be computed locally, quickly and privately. Complex ones go to the cloud. For a developer, this means the system, rather than each app, takes care of routing between the local and cloud model.

One specific hardware figure is given: on the Snapdragon X2 Elite, the system will be able to run a 26-billion-parameter model locally. This is a Qualcomm chip for laptops, so it's worth clarifying exactly how KORA is positioned: as a mobile system in the narrow sense, or as an OS for ARM devices more broadly. The announcement doesn't make this entirely clear. It's also unknown which model is meant, with what quantization and at what generation speed. Without this data, "26B locally" remains a marketing number.

Where the catch is: openness

The main limitation is the license. Cortex will apparently remain closed. It will run on open models, but the layer that makes the decisions is closed. It's easy to read this as "almost open source", but in reality the most interesting part, namely orchestration, memory and routing, remains a black box. If you build a product on top of such a platform, you depend on the vendor just as you would on any closed SDK.

There is also a broader market question. Mobile Linux has been trying to find a niche for many years and each time runs into the app ecosystem. AI agents may become a new reason to try it, but only if enough agents and integrations appear. A single system isn't enough for that.

What to do about it

• Agent developers should take a look at the architecture itself: shared memory, semantic search and sandboxes by default are what will soon be expected from any platform, not just KORA.
• Product managers would do well to build the question "what can the agent see and touch" into the design before choosing a model. Data isolation is becoming part of user trust.
• Founders thinking about betting on a platform should simply watch for now: there is no public Cortex code, no independent measurements, and no clear distribution model.

Bottom line: the idea of building agents into the OS is logical, and the list of promises looks right. But for now it is a press release. Until code, devices and benchmarks appear, KORA should be treated as an interesting announcement, not a ready platform.