Microsoft has unveiled a Windows "hybrid intelligence" plan that would let AI agents carry out tasks directly on a computer, provided the user grants permission. Under the setup, Copilot would be able to read local files and recent activity, then perform actions such as organizing files, troubleshooting issues, and writing code. In Microsoft’s demo, the system gathered tax documents, renamed files, bundled attachments, and drafted an email to an accountant.
The company said the architecture relies on local and cloud AI working together. Copilot handles the request, while Windows provides access to files and execution capabilities. Some jobs can run on-device, which cuts cloud token usage and keeps sensitive data local, while tasks that need stronger reasoning can still be sent to cloud models.
To support this approach, Microsoft said it compressed its 137-billion-parameter coding model, MAI Code 1.1 Flash, to 53GB so it can run on high-performance PCs equipped with RTX Spark. GitHub Copilot will also use HydraFusion to choose between local and cloud models automatically. Microsoft has also introduced the MXC security sandbox, which is now generally available, while the new Copilot abilities are due to roll out to Copilot+ PCs in the coming months.
Microsoft has introduced a Windows "hybrid intelligence" plan aimed at letting AI agents carry out tasks directly on a PC.
With user permission, Copilot would be able to read local files and recent activity, while also handling jobs such as organizing files, troubleshooting problems, and writing code. In Microsoft’s demonstration, the assistant pulled together tax materials, renamed files, bundled attachments, and drafted an email to an accountant.
Local and cloud AI split the workload
Microsoft said the system is built around coordination between on-device AI and cloud AI. Copilot receives the task, while Windows provides the ability to access files and execute actions. Some work can be handled by models running on the computer itself, reducing cloud token consumption and keeping sensitive data on the device. When a task needs stronger reasoning, the system can call a cloud model instead.
Model compression, coding support, and sandbox controls
To make local execution possible, Microsoft said it compressed its 137-billion-parameter coding model, MAI Code 1.1 Flash, to 53GB, allowing it to run on high-performance computers equipped with RTX Spark. GitHub Copilot will also use HydraFusion to automatically pick either a local or cloud model, so some coding tasks will no longer depend on paid cloud inference.
Microsoft also released the MXC security sandbox to restrict which files and network resources an agent can access, with the goal of preventing overreach. MXC is already officially available. The new Copilot capabilities, including reading local files and taking actions on the computer, are set to arrive on Copilot+ PCs over the next few months.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
