Microsoft turns Copilot into a local Windows agent

Microsoft is giving Copilot permission-based access to files, local models and Windows actions, while introducing new controls designed to keep autonomous agents in check.

Microsoft has broadened Copilot on Copilot Plus PCs into something closer to an operating layer than a chatbot. At its San Francisco Windows and Surface event on 7 October, the company said the assistant can now draw on local context with permission, carry out actions directly on the device, and route selected work to models running on the PC itself rather than to the cloud. Microsoft is packaging the approach as “hybrid intelligence”, meaning Windows decides task by task whether processing should remain on-device or be sent elsewhere.

The practical effect is a more capable personal agent. With permission, Copilot can inspect files, folders and recent activity stored on the machine, then use that information to answer questions or assemble a task. It can also perform actions across Windows, including reorganising folders, running diagnostics, writing code and executing multi-step workflows. In a demonstration described by the company, an email requesting tax paperwork triggered a chain of work in which Copilot searched local and cloud-synchronised locations, collected relevant documents, sorted them, drafted a reply and completed the task. Microsoft also showed new taskbar search actions that can carry out commands such as switching to dark mode, muting audio, minimising windows or sending a text message without first opening the relevant app.

The same event introduced Copilot Autopilot, Microsoft’s name for a persistent agent that can act on standing goals rather than waiting for one-off prompts. According to Microsoft’s presentation, that agent uses the same local context, local models and local actions framework to complete tasks end to end. The company argues that the economics of AI make this shift necessary: every cloud call consumes metered tokens, and agents can burn through far more of them than a simple chatbot exchange. Microsoft says Copilot Plus PCs are already handling more than 2 trillion local inferences a month across features such as search, imaging and video calls, while more than 40% of new business laptops are now Copilot Plus machines.

Security and governance are central to the update. Microsoft Execution Containers, or MXC, are now generally available on Windows 11 and are designed to define exactly which files and network destinations an agent may access while it is running. The company says the permissions are enforced live, not merely set once at installation, which is intended to reduce the risk created by giving software broad access to a user’s device. Microsoft has also paired the system with Agent 365 and Intune, so each agent has an identity separate from the human user. That separation allows administrators to audit activity, assign policy and monitor token use. Microsoft says MXC already works with agent tools including OpenAI’s Codex, GitHub Copilot, OpenClaw, Replit, LM Studio and Nvidia’s NVOpenShell, with support for Claude Code, Perplexity, Raycast and Meta’s Muse to follow.

The hardware behind the strategy is equally ambitious. Microsoft and Nvidia used the event to introduce the Surface Laptop Ultra, a premium machine built around Nvidia’s RTX Spark chip and aimed at developers and other heavy AI users. The laptop starts at $2,599, with higher-end configurations reaching $5,999, and is due to begin shipping on 16 October, according to reports from Tom’s Hardware, Tom’s Guide, Windows Central, Axios and TechRadar. Those reports said the system includes up to 128GB of memory, a mini-LED display and a design intended to support very large on-device models. Microsoft also announced a Surface RTX Spark Dev Box for November, while Asus, Dell, HP, Lenovo and MSI are also bringing RTX Spark machines to market.

Under the hood, Microsoft is also shrinking its own models so they can run locally without losing too much quality. The clearest example is MAI-Code-1.1 Flash, a coding model with 137 billion total parameters but only 6.8 billion active at any one time. Microsoft says compression reduces the model to roughly 53GB at 3-bit precision while preserving a 256,000-token context window and keeping the benchmark drop modest, with SWE-bench Verified falling from 72.6% to 70.8%. The company is using the same logic in GitHub Copilot through a routing system called Hydrofusion, which automatically decides whether a coding task should run locally at low cost or be sent to a larger cloud model. Microsoft says Windows ML now supports llama.cpp as well, letting developers run open-source models such as Nvidia’s Nemotron and DeepSeek’s V3 Flash on local GPU, NPU and CPU resources.