Windows wants to decide where your AI thinks

Imagine asking Windows to organize a project and discovering that your new assistant likes to subcontract. Some work stays on your computer. Some goes elsewhere. Before handing it the folder, you would reasonably like to know who else is getting the job.

That is the interesting question behind Microsoft’s October 7 hybrid intelligence announcement: agents combining local and cloud computing. The attraction is obvious. Less fiddling with AI tools, more useful work finished. The catch is that convenience can make important decisions almost invisible.

The best version of this Windows would let you delegate the boring parts while keeping a clear view of the boundaries. The worst would turn “Auto” into a button you only understand after something surprising happens.

The same chat can involve different computers

GitHub’s HydraFusion already has a cloud-based research preview. It can use one model, escalate a task, or bring in another model to critique a draft. Its Windows local-model extension is scheduled for experimental preview later in October in GitHub Copilot’s app, CLI and VS Code.

GitHub describes two approaches to local models: automatic local/cloud selection or choosing a local model yourself, including through compatible local endpoints. Model inference and tool execution have separate boundaries. Selecting a local model does not make the session offline.

Picture a coding assistant preparing a small change on your PC, then seeking cloud help with a difficult problem. That is the kind of hybrid workflow worth asking for. The useful question is what crosses the boundary, and whether your settings can stop it.

There is a bill attached to some of that plumbing. GitHub’s pricing documentation ties cloud-model usage to tokens and the chosen model, with allowances depending on the plan. Local work could reduce cloud use, but “hybrid” tells you very little about the final bill. Ask for the total cost of finishing the task, including attempts that went nowhere.

A seamless interface should not require blind faith in the plumbing.

A sandbox needs to be switched on

Microsoft Execution Containers, or MXC, is now generally available. It lets developers define access to resources such as files and networks, with containment enforced outside the agent’s own decision-making. Microsoft’s separate Entra attribution and Intune management additions are still coming. Those layers should not be flattened into one giant “available now” sticker.

The immediate catch is refreshingly specific: GitHub says local sandboxing is off by default. Its CLI and app have separate settings. Built-in file tools check policy inside the application on a best-effort basis rather than through the operating-system sandbox; remote tools connected through MCP sit outside local isolation. Check the actual configuration before trusting the badge.

For a first experiment, enable sandboxing and restrict access to a disposable project folder and the tools the task needs. Review any request to relax those limits. “Please be careful” is a lovely sentiment. It is an ambitious substitute for access controls.

The everyday features have their own calendar

Copilot+ PCs are promised permission-based local context, PC actions and local models over the coming months. Windows Update is required, and timing varies by device, market and silicon. This is a rollout plan, not a reason to assume your current Copilot can already do everything in the presentation.

There is also a naming trap. Microsoft’s September introduction of Autopilot describes a cloud-hosted agent, with expansion to private preview. Adding local capabilities later does not establish that the entire assistant lives on your laptop. Nor does it establish that it will keep doing local work after you close the lid.

Developers have something more concrete to explore: experimental llama.cpp support in Windows ML, with GGUF and ONNX text generation and ONNX Whisper speech recognition. Models must be supplied, and Microsoft advises checking supported scenarios and limitations before production use. Better infrastructure could make useful local apps easier to build. It still takes somebody to build the app.

The new taskbar Search actions offer a smaller, relatable win: changing a setting or arranging windows without a tour of menus. Their gradual Experimental Insider preview began October 7, initially in English. Phone Link actions such as texting are still forthcoming. A search box that saves three clicks has a perfectly respectable reason to exist.

Gaming has an excellent test for all this ambition

Ask a PC gamer what should happen after pressing Play. You probably will not hear a request for more orchestration. Getting into the game without an opening ceremony about shaders would be a strong start.

Gears of War: E-Day launched worldwide on October 6, including on PC and Steam. It is a useful example because the game’s availability and the arrival of a Windows graphics feature are separate events. Owning the game is only the beginning of the compatibility checklist.

Advanced Shader Delivery already reached general availability on supported AMD hardware in June. Its basic approach is to deliver precompiled shaders with the download, reducing work at launch. The October expansion uses E-Day as its showcase. This is the sort of improvement that deserves attention precisely because it aims to make an annoying wait disappear.

For E-Day, Microsoft lists Windows 11 24H2 or newer, supported AMD or Snapdragon X2 hardware, and specified drivers. The Xbox PC app is the documented delivery route. Intel and NVIDIA RTX, including RTX Spark, are scheduled later in October. Microsoft claims roughly 95 percent less first-launch shader-compilation time; that is a vendor result for this scenario, not a universal gaming benchmark.

The store matters here. E-Day being sold on Steam does not prove that this Xbox PC app feature follows it there. The graphics hardware and driver matter too. Reading those details is less exciting than a cinematic trailer, but considerably cheaper than discovering them after buying the wrong setup.

Microsoft says Call of Duty is coming to RTX Spark in 2027. Keep future support in the future column. A compatibility announcement is encouraging; the game you and your friends can actually launch tonight has the stronger argument.

For a mixed gaming-and-AI PC, the sensible audition includes both jobs. Run the game you care about, then find out whether an assistant working alongside it affects the experience. Look at frame consistency, responsiveness and background resource use. Give the background assistant a place in the test plan before handing out medals for the whole platform.

Make the invisible decisions visible

Windows has a promising idea here. An assistant could save time by putting work on the right machine, and better system features could leave more of an evening for the game itself. Both deserve a fair trial.

The standard should stay wonderfully ordinary: show where the work runs, what it can access and what it costs. Make permissions understandable and stopping easy. If Windows wants to handle the subcontracting, it should let the person who owns the PC read the job sheet.