I got up this morning and I was thinking about what these launch events mean to all of us,
said Microsoft CEO Satya Nadella, opening the event in San Francisco on 7 October. Microsoft has demonstrated its vision for the Windows operating system transforming into a foundation for locally executed artificial intelligence (AI) agents. The company announced the general availability of Microsoft Execution Containers (MXC) to govern autonomous software at runtime.
Sandboxing Autonomous Workloads via MXC
The push to transform Windows into a platform for local and hybrid AI execution relies on isolation layers. Microsoft Execution Containers (MXC) are now generally available to give IT administrators runtime controls to enforce policies that determine which files and other enterprise IT resources an agent has access to. This containment layer separates agent activity from user behavior, allowing security teams to evaluate agent actions independently.
These local agents operate within agent sandboxes enforcing fine-grained permissions at runtime, a design Pavan Davuluri, executive vice-president for Windows and devices, cited as making Windows the most secure platform for agents. GitHub Copilot utilizes MXC alongside the BaseContainer tier of the ProcessContainer backend on Windows, Seatbelt on macOS, and bubblewrap on Linux.
When sandboxing is enabled, shell commands and local Model Context Protocol (MCP) servers run within the process boundary, though built-in file tools check requests against effective policies rather than relying entirely on OS-enforced child-process isolation.
Hybrid Intelligence and Intelligent Routing Across Models
Microsoft is rolling out hybrid intelligence infrastructure to balance local hardware constraints with cloud-scale capabilities. The system relies on a model router—referred to as Project HydraFusion—to dynamically route execution between local inference frameworks and cloud models like OpenAI or Anthropic depending on task complexity and privacy needs.
To support this on-device execution, Microsoft is extending Windows ML inference framework support (built on the ONNX runtime) to Llama.cpp, enabling models to leverage a wider array of hardware accelerators. GitHub Copilot will incorporate automatic routing, deciding whether coding tasks execute locally or travel to cloud models based on task context and cache state.
Performance metrics released for MAI Code 1.1 Flash indicate a peak memory usage of 75.5GB at a 256k context length, with quantized models scoring 70.8% on SWE-Bench Verified compared to 72.6% for full-precision variants.
Ecosystem Adoption and Enterprise Governance
A broad array of AI developers and projects are integrating with the new Windows architecture. MXC compatibility is featured in offerings such as Codex from OpenAI, GitHub Copilot, OpenClaw, Replit, LM Studio, OpenShell from Nvidia, and Unsloth AI. Microsoft is also working to onboard Anthropic Claude Code, Box, Egnyte, Heidi Health, Hermes Agent by Nous Research, Manus, Perplexity, Raycas, and Simular, alongside Meta’s Muse as a native Windows app.
Divya Venkataramu, director of product marketing at Microsoft, emphasized that combining Windows security features with Microsoft Agent 365 allows organizations to identify active agents, track specific activities, investigate risky behavior, and apply targeted policies without relying on user-level data or device-level capabilities.
Additional Performance and Context
With containment, identity and manageability built into the platform, Windows helps you control what agents can access, know which agent acted, and govern them at scale,
Pavan Davuluri stated in a blog post.
