Microsoft made Microsoft Execution Containers (MXC) generally available as a cross-platform, policy-driven boundary for agent workloads. Developers can scope file, network, process, and UI access independently of the model; Windows offers process/session/WSL isolation, with experimental microVMs on Windows/Linux. Microsoft says GitHub Copilot, OpenAI Codex, OpenClaw, Replit, LM Studio, and Unsloth already support MXC. This is a notable shift from per-agent sandboxing toward a shared OS-enforced control layer; Entra identity attribution and Intune policy management are still forthcoming. Sources: Microsoft announcement ; MXC repository ; HN discussion
Microsoft says GitHub Copilot will add local-model inference and route work between on-device and cloud models across CLI, app, and VS Code, expected by month-end. Its first local coding option is a quantized MAI Code 1.1 Flash MoE (137B total/6.8B active parameters); Microsoft reports 70.8% SWE-Bench Verified and 66.29% Terminal-Bench 2.1 on-device, versus its BF16 cloud model’s 72.6% and 62.9%. These are vendor-run results on an RTX Spark/Surface Laptop Ultra setup, not independent measurements; local inference does not imply offline execution. The agentic significance is model placement and tool execution being coordinated across compute environments, with MXC sandboxing shell tools and local MCP/language servers. Sources: Microsoft technical announcement ; HN discussion