Microsoft has added local model support and sandboxed agentic sessions to GitHub Copilot, with a new model called MAI Code 1.1 Flash running on-device rather than in the cloud.
Auto-orchestration: local vs. cloud
The headline feature is what Microsoft calls Auto orchestration. GitHub Copilot now decides on its own whether a given task should run on a local model or a cloud-scale model. According to Microsoft, the goal is to remove infrastructure decisions from the developer entirely.
“Rather than forcing developers to manage infrastructure decisions themselves, GitHub Copilot automatically coordinates local and cloud inference.”
Developers who want explicit control can still select a local model directly instead of relying on Auto mode.
️ Sandboxed agents via Microsoft Execution Containers
Agentic coding sessions now run inside Microsoft Execution Containers (MXC), which control access to files and network resources. The sandboxing is designed to let automated tasks run without exposing the broader system to unintended side effects.

Local inference also includes dynamic memory management, which Microsoft says reduces overhead when running models on-device.
Hardware requirement to note
Full local inference performance is tied to NVIDIA RTX Spark Windows PCs, with the Surface Laptop Ultra called out specifically. That hardware requirement is a real constraint for anyone not already on that class of machine.
The operator takeaway
If you run a small dev shop or solo code on Windows, the Auto orchestration feature is worth watching. The ability to automate tasks like reading a local repo, running tests, and generating an HTML report in a single daily script is the practical use case Microsoft points to. The MXC sandboxing makes that less risky than running unconstrained agent sessions. Full details are on the Microsoft Command Line Blog.
