Copilot Hybrid Intelligence Gives Windows Agents Local File Access

What Microsoft announced

At its Windows and Surface event on October 7, 2026, Microsoft showed a Copilot upgrade that can read local files on a PC and carry out actions across Windows.4 The company calls the approach "hybrid intelligence." Under it, apps and agents use a mix of on-device and cloud models and send each task to wherever it can be done most efficiently.45

The upgrade has three parts. Local context lets Copilot, with the user's permission, understand files and recent activity on the machine. Local actions let it organize files, run device diagnostics, troubleshoot, write code and carry out workflows on the device. Local models let it run tasks on AI models stored on the PC and call cloud models when it needs more power.1 All three apply to the redesigned Copilot app Microsoft introduced on September 25, which has three sections: Home, Code and Autopilot.1

Jacob Andreou, Microsoft's EVP of Copilot, showed a tax demo on stage. Autopilot searched several folders for an accountant's requested documents, renamed them, zipped them and drafted an email with the archive attached.4 BGR's live coverage reported that the file handling in that demo stayed on the device.3 Andreou said the agent "has the same level of the ability to control and change things that I do as a user, but always with permission."6

Rollout details depend on the source

The main facts match across outlets. The timing and scope are less consistent. Andreou said the features would reach Copilot over the "next couple months."4 Microsoft's written announcement is narrower: the features will roll out to Copilot+ PCs "in the coming months," with timing that varies by device, market and chip platform.1 One detailed review points out that Microsoft has published no build number, country list or subscription requirement. It also says early reports overstated this as a general Windows feature, when it is limited to Copilot+ hardware.3

The new Windows Search, which can trigger thousands of actions from the taskbar, is described in two ways. The Verge reported Davuluri saying it would reach Windows 11 PCs this fall.4 Other coverage says it began rolling out on October 7 only to Windows Insiders in the experimental channels.13 Both can be true: it starts with Insiders and spreads later. Still, enterprises should plan on the Insider timeline, not the keynote wording.

In practice, this is a roadmap commitment for a specific class of hardware. IT departments are not getting a capability they can roll out this quarter.

The pitch is about cost

Microsoft's main argument is economic, more than technical. It says customer demand for AI is growing faster than cloud budgets, and that hybrid intelligence is meant to stretch those budgets while keeping access to the most capable models.1 Axios quoted Davuluri saying people's appetite for intelligence "is effectively unlimited."3 Microsoft also says more than 2 trillion inferences already run locally each month across Copilot+ PCs, and that more than 40% of laptops built for business are Copilot+ PCs.1

The timing explains why cost leads the message. Two weeks earlier, Microsoft split Microsoft 365 Copilot pricing in two. Everyday chat stays on per-user licenses. Cowork, Code, Autopilot, long-running agent features and frontier models move to usage-based billing.18 Usage-based services are off by default, and administrators set spending limits for the organization, groups or individual users through Agent 365.14 With metered agent work on the cloud side, local models that carry "no inference charge," as Microsoft AI described on-device calls in GitHub Copilot, become a direct way to control spending.10

The developer side follows the same logic. Microsoft is shrinking its MAI Code 1.1 Flash model, which has 137 billion total parameters, to 3-bit precision for on-device use. It says this cuts the model's size by nearly 80% while keeping a 256K context window.1 GitHub's HydraFusion router, which picks a model for each task, will extend to local models in experimental previews for the GitHub Copilot app, the CLI and Visual Studio Code later this month.110

IDC's Tom Mainelli described the enterprise situation directly. Most companies are still working out token usage and where to run models, and the need to run AI locally will grow as they extend it to more employees.6 He added that current business laptops still rely mostly on cloud Copilot, and that this changes only once the right mix of CPU, GPU, NPU and memory is widely installed.6 Analyst Mahapatra noted that the installed base of enterprise PCs is still "highly mixed."6

Governance is ahead of the features

For enterprise buyers, the most important news may not be Copilot itself. Microsoft Execution Containers (MXC) became generally available on Windows 11 the same day. MXC lets organizations define which files and networks an agent can reach, and enforces those rules at runtime.1 The policy sits outside the agent's control, so neither the agent nor the code it writes can grant itself more access.1 Isolation options range from process isolation through session containers, WSLc and virtual machines to Windows 365 for Agents.1

General availability matters. One analysis notes that MXC was an "early preview" at Build in June. A preview gets tested on a lab machine, while a generally available product can go to a whole department with support behind it.13 Several agents already support MXC, including OpenAI's Codex, GitHub Copilot, OpenClaw and Replit. Anthropic's Claude Code, Box, Perplexity and others plan to add support.1

Microsoft is also linking agent containment to identity and management. Windows will soon let Microsoft Entra separate agent activity from user activity in Agent 365. Intune policy for MXC process containers is also coming.1 This follows September's design, in which Autopilot agents run under their own governed Entra identities and can have their own mailboxes and calendars. Each one becomes another identity to license, audit and pay for.

One gap in the coverage deserves attention. Microsoft has not said that every local action Copilot takes, such as renaming tax files, runs inside MXC. One review warns that permission prompts and runtime containment are separate controls and should not be treated as the same thing. Crypto Briefing reports that activation is opt-in, folder access is restricted and Copilot agents run under accounts separate from the user's.2 Microsoft's own blog post, however, does not describe the permission prompts, the default access scope or whether there is an activity log.3 My reading is that Microsoft has built strong containment for developer and third-party agents but has documented much less about its own consumer-facing Copilot. IT teams should ask about that difference before they turn anything on.

Why trust is the bottleneck

Recall is the background to the scrutiny. When Microsoft first showed Recall, researcher Alex Hagenah found its captured data stored in an unencrypted local database. Microsoft then made it off by default, added encryption and moved it back into Insider testing.9 Asking Copilot to read files and act on them is a bigger request than Recall's searchable snapshots, and observers expect it to get closer scrutiny.3 Windows security and platform leads Dana Huang and Logan Iyer wrote earlier this year that agents which read files and modify environments bring "new risk to control and trust."9

The opt-in design helps on privacy but limits adoption, because the feature reaches only users who choose to grant access.2 Crypto Briefing also reports that the Microsoft 365 Copilot user base is still small, with few users engaging with premium features.2 That makes Windows' advantage harder to use. No AI competitor owns the operating system on most of the world's PCs, and the deep integration only pays off if users and administrators allow it.2

Mahapatra's advice is likely to shape how this rolls out. Trust should rest on controls, not on the AI itself. Read-only help, summarization and search will probably be adopted well before autonomous file changes or business-process execution.6 He argued that as agents move from making recommendations to taking actions, organizations will need stronger approvals, monitoring, auditing and human oversight, especially around regulated data and systems of record.6

What enterprises should take from it

This announcement is less about one Copilot feature than about Microsoft lining up its whole enterprise AI stack. Windows provides local context and containment. Entra and Agent 365 handle identity, management and cost control. GitHub handles model routing. Surface and its NVIDIA RTX Spark partners supply hardware that can run large models locally.110 The company is betting that the next phase of enterprise AI will depend on governance and the cost of each task, not on model benchmarks.

For IT leaders, the practical steps come before the features. Inventory the Copilot+ PCs in the fleet, since only those devices are in scope. Check Intune and Agent 365 licensing. Treat agent identity as a new audit category. Spending limits should be in place before metered features go live.18 Microsoft has built the controls for enterprise agents, but the first wrong file Copilot moves on a production machine will test how much they help.