16 Billion Windows Users Enter the Agent Era Overnight

MXC sandbox and OpenClaw integration

Microsoft introduced MXC (Microsoft Execution Container), an OS‑level sandbox that enforces isolation policies in the Windows kernel. Four isolation levels are available:

Process‑level for lightweight tasks

Session‑level for user separation

Virtual‑machine/WSL level for stronger boundaries

Windows 365 cloud level for fully isolated cloud environments

During the Build demo, OpenClaw (the "lobster" project) attempted to delete all files on the desktop. MXC’s read‑only policy blocked the operation, leaving 94 image files untouched. The demo highlighted that OpenClaw can now run natively on Windows under MXC control. Peter Steinberger also announced the formation of the OpenClaw Foundation to keep the project open and neutral.

Autopilot agents and Scout

Microsoft announced a new class of enterprise‑grade agents called Autopilots. The first Autopilot, named Scout, runs 24/7 with its own identity and productivity license. Scout can join Teams chats, monitor Outlook mail, and act as a long‑lived digital employee. Copilot Frontier users can enable Scout immediately, and Microsoft plans to add more Autopilot agents to form a “digital team”.

Copilot Super‑App

Copilot will consolidate four capabilities into a single “super app” with one entry point, unified identity, and shared context:

Chat – conversational AI

Cowork – multi‑step task delegation

Code – GitHub Copilot programming assistance

Autopilot – autonomous enterprise agents

This integration removes the fragmented admin experience across separate back‑ends.

MAI model family

Seven self‑trained multimodal AI (MAI) models were released, covering inference, programming, image generation/editing, transcription, and voice synthesis. Key specifications and benchmark results:

MAI Thinking 1 : MoE architecture, ~1 trillion total parameters, 35 billion active parameters, 256 K context window. Achieves 53 % on SWE‑bench Pro, matching Anthropic Opus 4.6.

MAI Code 1 Flash : 5 billion parameters, tuned for VS Code and GitHub Copilot CLI. Scores 51 % on SWE‑bench Pro, comparable to Haiku‑class models.

MAI Image 2.5 and MAI Image 2.5 Flash : First Microsoft models covering both text‑to‑image and image‑editing. Rank 3 on the Arena text‑to‑image leaderboard and rank 2 on the image‑editing leaderboard, surpassing Google’s Nano Banana Pro. The standard variant targets high fidelity; the Flash variant optimizes for efficiency.

MAI Transcribe 1.5 : Supports 43 languages, claimed as the world’s best transcription model. Beats Gemini and OpenAI flagship models on accuracy and runs five times faster. Integrated into GitHub, Teams, Copilot, and Dynamics 365 Contact Center.

MAI Voice 2 and MAI Voice 2 Flash : Support 15 languages with fine‑grained emotional control (e.g., anger, confusion, embarrassment). The Flash version is designed for ultra‑low latency voice agents.

All models are available on the Foundry platform and can be accessed via OpenRouter, Fireworks AI, and Base 10, allowing developers to call them from their preferred ecosystems.

RTX Spark “super chip” hardware

Microsoft unveiled the RTX Spark chip on a Surface laptop, delivering 1 PFLOP of AI compute, 20 CPU cores, and 128 GB of unified memory. Engineer Kayla demonstrated three local large models running concurrently, including a 1 200 billion‑parameter model that consumed 90 GB of GPU memory and processed 3.4 million tokens.

Local model execution on RTX Spark

On the RTX Spark‑equipped Surface, the GPU panel showed 90 GB of memory usage while processing 3.4 million tokens locally, with no external dependencies or latency spikes. This demonstrates fully offline, high‑throughput AI workloads on a consumer‑grade PC.

Overall architecture

Windows is being repositioned from a static desktop OS to an “Agent operating system”. The stack consists of:

MXC sandbox for secure, kernel‑enforced isolation of agents.

MAI models providing multimodal intelligence.

RTX Spark (Maia) silicon delivering petaflop‑scale compute.

Combined, these layers enable a unified Copilot super‑app that serves as the primary entry point for 1.6 billion Windows users, turning the desktop into a platform for autonomous, enterprise‑grade agents.

Code example

往
期
推
荐
1、
2026 年编程 Agent 实测排行榜：Cursor、Claude Code、Codex 看看谁更强？
2、
别再死磕MySQL了，AI时代这个数据库更香！
3、
忘掉Win 12吧，曝微软要把Win 11打造成本地“龙虾”
4、
十万个Why：两个字段都建了单列索引，为什么加了 OR，执行计划还是全表扫描？
5、
Linux 系统误将 chmod 权限改成 了 000，如何恢复?
6、
面试官皱眉：“Claude Code 你用了半年，CLAUDE.md 平时怎么维护？”我：“呃...没用过”，他：你牛
点
分
享
点
收
藏
点
点
赞
点在看