Microsoft Execution Containers (MXC): An OS-Level Sandbox for AI

To support the security demands of this new agentic world, Microsoft introduced the Microsoft Execution Containers (MXC) SDK, now in early preview. MXC is a policy-driven execution layer baked into the Windows kernel that lets developers and administrators define precisely what an agent can access—files, networks, tools—with containment enforced by the operating system itself, not by hopeful application-level restrictions .

Each agent runs in a lightweight sandbox with its own identity, bringing enterprise-grade manageability to locally running AI. Microsoft also announced that Agent 365, its endpoint management and security suite, will natively integrate with MXC to discover and govern agents like OpenClaw, Claude Code, and GitHub Copilot CLI through policy-based controls . It represents a shift from monitoring agents to actively containing them.

Microsoft Scout: The First “Autopilot” Agent

The headline product reveal was Microsoft Scout, the first agent in a new category Microsoft calls “Autopilots.” These are always-on, autonomous agents with their own identity that understand workflows across applications and can act proactively without being prompted each time .

Scout is a desktop application for Windows 11+ and macOS 12+, built on the OpenClaw multi-agent framework and deeply integrated with Microsoft 365. It can handle meeting preparation, schedule triage, and cross-app task coordination autonomously. It launched into an immediate experimental preview through Microsoft’s Frontier program . This sits within a broader Copilot “super app” strategy that now unifies Copilot Chat, Cowork, GitHub Copilot, and Autopilots under a single multi-model experience.

Project Solara: A Chip-to-Cloud Platform for Agent-First Devices

Perhaps the most forward-looking reveal was Project Solara, a chip-to-cloud platform designed for a new class of hardware that runs AI agents instead of traditional apps . Unlike conventional computers, these “agent-first devices” are task-oriented and built from silicon up to run a lightweight, secure operating system based on the Android Open Source Project rather than Windows .

Microsoft showed two working concept devices running on Qualcomm and MediaTek chips:

• A desk companion resembling a smart speaker or smart clock that responds to voice commands, recognizes users through facial identification, and surfaces critical daily information. With a monitor attached, it can serve as a full Cloud PC terminal .
• A wearable lanyard or work ID badge, roughly the size of a small phone, featuring a fingerprint sensor and the ability to record video and take voice commands for hands-free information retrieval .

Pilot programs are already underway with Best Buy, CVS Health, Levi’s, and Target . Microsoft describes the OS layer as “liminal”—minimal on the edge device and fluidly moving between the device and cloud via agents .

MAI-Thinking-1: Microsoft’s In-House Reasoning Model

On Day 2, Mustafa Suleyman unveiled MAI-Thinking-1, Microsoft AI’s first reasoning model, as the flagship of a new family of seven in-house MAI models covering reasoning, coding, image generation, transcription, and voice .

The spec sheet marks a significant engineering achievement:

• Architecture: 35 billion active parameters out of approximately 1 trillion total parameters, using a sparse Mixture of Experts design .
• Context window: 256,000 tokens, sufficient to process roughly a 600-page document .
• Training: Built from scratch on clean, enterprise-grade, commercially licensed data with no distillation from third-party models like OpenAI’s GPT series .

Benchmark results position it competitively:

• 52.8% on SWE-Bench Pro (software engineering)
• 97.0% on AIME 2025 (mathematical reasoning)
• 87.7% on LiveCodeBench v6 (competitive coding)

Microsoft reports that independent human raters on the Surge platform preferred MAI-Thinking-1 to Anthropic’s Claude Sonnet 4.6 in blind side-by-side evaluations for overall quality . The model enters private preview through Microsoft Foundry, signaling a serious post-OpenAI-exclusivity strategy for Microsoft’s AI stack .

Azure Cobalt 200 VMs and Developer Hardware

On the infrastructure side, Microsoft announced the early access preview of Azure Cobalt 200 Arm-based Virtual Machines. Built on Microsoft’s second-generation custom Arm processor using the Arm Neoverse V3 Compute Subsystems platform and fabricated on a 3nm process node, these VMs deliver up to 50% better generational performance over Cobalt 100. They are designed from the ground up for scale-out, cloud-native, and Linux-based agentic AI workloads .

The company also unveiled the Surface RTX Spark Dev Box, created in partnership with Nvidia. With 1 petaflop of AI compute, it is purpose-built for developing and testing AI agents locally .

Supporting Announcements and the Bigger Picture

Several other launches reinforce the agent-first direction:

• Microsoft IQ reached general availability, providing a context layer that grounds agents in both world knowledge and enterprise data across GitHub Copilot, Microsoft Foundry, and Copilot Studio .
• Web IQ debuted as a set of APIs that connect AI agents to fresh web information for real-time grounding .
• Azure AI Foundry now supports multi-model orchestration, automatically selecting the best model for each task from a pool that includes OpenAI, Anthropic, and open-source options .
• Azure HorizonDB entered preview as a fully managed PostgreSQL-based database built for the demands of agentic applications .

None of these announcements exist in isolation. Together, they paint a picture of a Microsoft that is betting the company’s future on agents replacing apps as the primary unit of computing—and building the security models, processing power, and foundational AI to make that shift real.

Microsoft Execution Containers (MXC): An OS-Level Sandbox for AI

To support the security demands of this new agentic world, Microsoft introduced the Microsoft Execution Containers (MXC) SDK, now in early preview. MXC is a policy-driven execution layer baked into the Windows kernel that lets developers and administrators define precisely what an agent can access—files, networks, tools—with containment enforced by the operating system itself, not by hopeful application-level restrictions .

Each agent runs in a lightweight sandbox with its own identity, bringing enterprise-grade manageability to locally running AI. Microsoft also announced that Agent 365, its endpoint management and security suite, will natively integrate with MXC to discover and govern agents like OpenClaw, Claude Code, and GitHub Copilot CLI through policy-based controls . It represents a shift from monitoring agents to actively containing them.

Microsoft Scout: The First “Autopilot” Agent

The headline product reveal was Microsoft Scout, the first agent in a new category Microsoft calls “Autopilots.” These are always-on, autonomous agents with their own identity that understand workflows across applications and can act proactively without being prompted each time .

Scout is a desktop application for Windows 11+ and macOS 12+, built on the OpenClaw multi-agent framework and deeply integrated with Microsoft 365. It can handle meeting preparation, schedule triage, and cross-app task coordination autonomously. It launched into an immediate experimental preview through Microsoft’s Frontier program . This sits within a broader Copilot “super app” strategy that now unifies Copilot Chat, Cowork, GitHub Copilot, and Autopilots under a single multi-model experience.

Project Solara: A Chip-to-Cloud Platform for Agent-First Devices

Perhaps the most forward-looking reveal was Project Solara, a chip-to-cloud platform designed for a new class of hardware that runs AI agents instead of traditional apps . Unlike conventional computers, these “agent-first devices” are task-oriented and built from silicon up to run a lightweight, secure operating system based on the Android Open Source Project rather than Windows .

Microsoft showed two working concept devices running on Qualcomm and MediaTek chips:

• A desk companion resembling a smart speaker or smart clock that responds to voice commands, recognizes users through facial identification, and surfaces critical daily information. With a monitor attached, it can serve as a full Cloud PC terminal .
• A wearable lanyard or work ID badge, roughly the size of a small phone, featuring a fingerprint sensor and the ability to record video and take voice commands for hands-free information retrieval .

Pilot programs are already underway with Best Buy, CVS Health, Levi’s, and Target . Microsoft describes the OS layer as “liminal”—minimal on the edge device and fluidly moving between the device and cloud via agents .

MAI-Thinking-1: Microsoft’s In-House Reasoning Model

On Day 2, Mustafa Suleyman unveiled MAI-Thinking-1, Microsoft AI’s first reasoning model, as the flagship of a new family of seven in-house MAI models covering reasoning, coding, image generation, transcription, and voice .

The spec sheet marks a significant engineering achievement:

• Architecture: 35 billion active parameters out of approximately 1 trillion total parameters, using a sparse Mixture of Experts design .
• Context window: 256,000 tokens, sufficient to process roughly a 600-page document .
• Training: Built from scratch on clean, enterprise-grade, commercially licensed data with no distillation from third-party models like OpenAI’s GPT series .

Benchmark results position it competitively:

• 52.8% on SWE-Bench Pro (software engineering)
• 97.0% on AIME 2025 (mathematical reasoning)
• 87.7% on LiveCodeBench v6 (competitive coding)

Microsoft reports that independent human raters on the Surge platform preferred MAI-Thinking-1 to Anthropic’s Claude Sonnet 4.6 in blind side-by-side evaluations for overall quality . The model enters private preview through Microsoft Foundry, signaling a serious post-OpenAI-exclusivity strategy for Microsoft’s AI stack .

Azure Cobalt 200 VMs and Developer Hardware

On the infrastructure side, Microsoft announced the early access preview of Azure Cobalt 200 Arm-based Virtual Machines. Built on Microsoft’s second-generation custom Arm processor using the Arm Neoverse V3 Compute Subsystems platform and fabricated on a 3nm process node, these VMs deliver up to 50% better generational performance over Cobalt 100. They are designed from the ground up for scale-out, cloud-native, and Linux-based agentic AI workloads .

The company also unveiled the Surface RTX Spark Dev Box, created in partnership with Nvidia. With 1 petaflop of AI compute, it is purpose-built for developing and testing AI agents locally .

Supporting Announcements and the Bigger Picture

Several other launches reinforce the agent-first direction:

• Microsoft IQ reached general availability, providing a context layer that grounds agents in both world knowledge and enterprise data across GitHub Copilot, Microsoft Foundry, and Copilot Studio .
• Web IQ debuted as a set of APIs that connect AI agents to fresh web information for real-time grounding .
• Azure AI Foundry now supports multi-model orchestration, automatically selecting the best model for each task from a pool that includes OpenAI, Anthropic, and open-source options .
• Azure HorizonDB entered preview as a fully managed PostgreSQL-based database built for the demands of agentic applications .

None of these announcements exist in isolation. Together, they paint a picture of a Microsoft that is betting the company’s future on agents replacing apps as the primary unit of computing—and building the security models, processing power, and foundational AI to make that shift real.