Microsoft is expanding Copilot's reach inside Windows, giving it access to files across the PC for search and productivity tasks. The company frames this as "hybrid intelligence": a HydraFusion model router decides when to use local models for simple or privacy-sensitive work and when to send complex requests to cloud models like OpenAI's GPT-6 or Anthropic's Claude. Local agents will run in Microsoft Execution Containers (MXCs), which apply fine-grained permissions at runtime.
To support local inference, Microsoft is extending Windows ML, its ONNX-based inference framework, to work with Llama.cpp. That should let Llama.cpp use a wider range of accelerators. The Register notes the security claims are optimistic, since models and routers are not infallible.
The new Copilot features will not have access to the entire file system, but they do open the door to advanced search and Office integrations. In a demo, Copilot found and organized tax documents, then drafted an email with those files attached. Windows is also getting AI-enhanced search for settings and a conversational interface for tasks like sending a text message.
Keeping sensitive tasks on-device may ease privacy concerns that have followed Copilot+ PCs since 2024, and it should cut server load. But the Register cautions that local model execution could hurt battery life, increase fan noise, and affect output quality. GitHub Copilot gets the hybrid functionality next week, with the broader Copilot app rollout following over the next few months. The source is a single Register report, so no independent confirmation is available.{...}? Wait I need output strict JSON only. I accidentally included