Copilot can now see your files, act across Windows, and run tasks on local models. Here's what Microsoft's hybrid intelligence update actually adds.
What did Microsoft actually announce for Copilot? #
At a Windows and Surface event in San Francisco on October 7th, Microsoft gave Copilot on Copilot Plus PCs three new abilities: local context (reading your files and recent activity with permission), local actions (organizing files, running diagnostics, writing code, and executing workflows directly on your machine), and local models (routing certain tasks to AI models that run on your device instead of the cloud). Autopilot, Microsoft’s persistent background agent, now uses all three. The company is calling the broader framework “hybrid intelligence,” a system where Windows decides, task by task, whether work should happen on your PC or in the cloud.
TL;DR #
- Copilot can now read your files and recent activity (with permission) and take real actions across Windows, including organizing folders, running diagnostics, and writing code.
- A new hybrid intelligence toggle lets Copilot route tasks to local models running on your PC’s chip, falling back to the cloud only when a task needs more power.
- Microsoft says Copilot Plus PCs are already running over 2 trillion local inferences a month across search, imaging, and video calls, which is the economic logic behind pushing more work on-device.
- A new security layer called Microsoft execution containers (MXC) , now generally available on Windows 11, lets organizations restrict exactly which files and networks an agent can touch, with rules enforced live as the agent runs.
- Every agent gets its own identity through Agent 365 and Intune, separate from the human user, so actions can be audited and attributed correctly.
- The update leans on new Nvidia-powered hardware, including the RTX Spark chip, built to run large models (Microsoft claims over 120 billion parameters) entirely on a laptop.
- Microsoft is shrinking its own frontier models for local use, including a compressed coding model that loses less than two points of benchmark accuracy while running on-device.
How does local context and local actions actually work? #
Local context means Copilot, with the user’s permission, can look at what’s stored on the PC, including files, folders, and recent activity, and use that as grounding for its answers. Local actions extend this from reading to doing: Copilot can reorganize files, run device diagnostics, troubleshoot problems, write code, and carry out multi-step workflows without the user manually triggering each step in each app.
Microsoft demoed this with a tax-document scenario: a user receives an email from their accountant requesting tax paperwork, and Copilot searches across the PC (including downloads and cloud-synced folders), gathers the relevant files, sorts them, drafts a reply, and sends it. Satya Nadella described this as agents now using “the new computer” alongside humans, pulling from the same file system container and combining local and cloud models to complete a task end to end.
Microsoft also added search-based actions in the Windows taskbar. Typing commands like “switch to dark mode,” “mute,” “minimize all windows,” or “text someone I’ll be there soon” executes the action directly, without opening the relevant app first.
What is Copilot autopilot? #
Autopilot is Microsoft’s description for a persistent, proactive, personal agent layer inside Copilot. Instead of waiting for a prompt, it’s meant to act on standing goals or recurring tasks using the same local context, local models, and PC action capabilities described above. The tax-document demo is the clearest example Microsoft showed: the agent didn’t just answer a question, it noticed an email, pulled files from multiple locations, organized them, and completed the task without the user doing the manual legwork.
Why is Microsoft pushing local AI instead of just using the cloud? #
The short answer is cost. Every cloud AI interaction burns tokens, and tokens cost money. That’s manageable for occasional chatbot questions, but agents are a different animal: a single agent task might call dozens of tools, read hundreds of files, and make thousands of small decisions, each one metered. Windows chief Pavan Davaluri framed the core business question as getting the most out of a token budget without giving up frontier-level capability. Microsoft’s own messaging puts it plainly: organizations want every AI token to count.
Running a task locally, on the chip already inside the laptop, sidesteps the metering problem entirely. Those tokens are effectively free once the hardware is paid for. Microsoft says this is already happening at meaningful scale: Copilot Plus PCs are reportedly handling over 2 trillion local inferences per month across features like search, imaging, and video calls, and more than 40% of new business laptops are now Copilot Plus machines. The strategy splits work by nature: cheap, frequent, privacy-sensitive tasks stay local, harder reasoning gets routed to the cloud, and Windows handles the routing decision automatically.
There are secondary benefits too. Local processing reduces latency, works offline, and keeps sensitive files from leaving the device, which matters for businesses with compliance requirements around data residency.
How does Microsoft secure an AI agent with access to your whole PC? #
Seven tools to build an app. Or just Remy. #
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Giving an agent broad file access and the ability to act on your behalf creates an obvious security problem, and Microsoft built a dedicated layer to address it: Microsoft execution containers (MXC), now generally available on Windows 11. MXC lets organizations define exactly which files and network destinations an agent is permitted to touch, and those permissions are enforced while the agent is actively running, not just set once at install time. Unlike a static sandbox, the containment boundary can change dynamically with each prompt and each tool call the agent makes.
The second piece is identity. Through Agent 365 and Intune, every agent gets an identity distinct from the human user’s. That separation means actions taken by an agent show up as the agent’s actions, not the user’s, which makes auditing and governance possible: administrators can track what an agent did, apply policies to it, and monitor its token usage.
MXC isn’t exclusive to Microsoft’s own tools. It already supports agent harnesses from OpenAI’s Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, and Nvidia’s NVOpenShell, with support for Claude Code, Perplexity, Raycast, and Meta’s Muse coming later. Nadella explained the reasoning on stage: coding agents were the first to run locally even when using cloud models, which made the desktop file system the most exposed attack surface, so Microsoft worked with Nvidia to build these security primitives directly into Windows rather than leaving each tool to handle it independently.
What hardware makes this possible? #
Local models need local horsepower, and Microsoft’s answer leans heavily on Nvidia’s new RTX Spark chip, first shown at Computex in Taipei. It combines a Blackwell RTX GPU with 6,144 CUDA cores, a 20-core Nvidia Grace CPU, up to 128 GB of unified memory, and up to one petaflop of AI performance in a package small enough for a thin laptop. Jensen Huang noted the chip took over 4,000 engineering years to develop and compared its one-petaflop performance to Nvidia’s original DGX-1 supercomputer, a 500-pound, quarter-billion-dollar machine from the early days of the deep learning boom that put up similar numbers.
The first device built on RTX Spark is Microsoft’s Surface Laptop Ultra, starting at $2,599 and shipping October 16th, with a 15-inch MiniLED display and a claimed ability to run AI models with more than 120 billion parameters entirely on-device. Microsoft also announced a Surface RTX Spark Dev Box at $5,999 for November, plus RTX Spark laptops from Asus, Dell, HP, Lenovo, and MSI. At the high end, a DGX Station for Windows is coming later this year, built on Nvidia’s GB300 Grace Blackwell Ultra chip with 784 GB of memory and 20 petaflops of compute, aimed at running models up to a trillion parameters locally or serving as a shared inference box for teams running many agents at once.
How are the AI models themselves changing? #
Running serious models locally requires shrinking them without gutting their capability. Microsoft’s clearest example is MAI-Code-1.1 Flash, its own coding model: a mixture-of-experts architecture with 137 billion total parameters but only 6.8 billion active at any given time. Compressed to 3-bit precision, it shrinks by roughly 80% to about 53 GB while keeping its full 256,000-token context window. The accuracy cost is small: the compressed version scores 70.8% on SWE-bench Verified versus 72.6% for the uncompressed model, a gap of less than two points.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
That model feeds into GitHub Copilot through a routing system GitHub calls Hydrofusion. In its automatic mode, Copilot decides whether a given coding request is simple enough to run locally for free or complex enough to route to a larger cloud model. This is rolling out as an experimental preview in the GitHub Copilot app, its CLI tool, and VS Code. Microsoft also says Windows ML now supports llama.cpp, letting developers run open-source models like Nvidia’s Nemotron and DeepSeek’s V3 Flash locally across GPU, NPU, and CPU. Other companies are building on the same shift: Perplexity demoed a local Windows agent that handles tasks like resizing photos and building a gallery entirely on-device, requiring an RTX GPU with at least 24 GB of video memory and using no cloud credits.
Frequently Asked Questions #
What is “hybrid intelligence” in Windows?
It’s Microsoft’s framework for letting AI tasks run either locally on the device or in the cloud, with Windows automatically deciding which option fits each task based on complexity, cost, and the hardware available.
Does Copilot need my permission to access my files?
Yes. Microsoft describes local context as requiring user permission before Copilot can read files or recent activity on the device.
What is Microsoft execution containers (MXC)?
It’s a security system, generally available on Windows 11, that lets administrators restrict which files and networks an AI agent can access, with those restrictions enforced dynamically while the agent runs.
Do I need new hardware to use these Copilot features?
The local model and local action features are built around Copilot Plus PCs and new Nvidia RTX Spark hardware like the Surface Laptop Ultra, which is designed to run large models entirely on-device. Older or non-Copilot-Plus machines won’t have the same local compute capacity.
Is this the same as running an open-source model locally with something like Ollama?
It’s related but broader. Microsoft’s update includes support for running open-source models (via Windows ML and llama.cpp) alongside its own compressed models, with Windows itself handling the decision of whether to run a task locally or send it to the cloud.