Alongside the Surface Dev Box, the unified stack includes Nvidia DGX Station for Windows and deskside DGX systems. These provide a scaling path for AI factories, allowing developers to prototype on a local RTX Spark PC and then scale the same agent workloads seamlessly to more powerful local stations or across to Azure’s AI infrastructure . This is the core of the hardware promise: a continuum of accelerated computing from the edge to the cloud.

The Models: From In-House to Open-Weights

The unified stack is model-diverse by design. The partnership featured major model announcements from both companies.

Microsoft’s AI Superintelligence Team released a family of seven new in-house models, headlined by MAI-Thinking-1, the company’s first reasoning model . In an early sign of performance, independent raters preferred it to Sonnet 4.6 in blind tests, and it matched Opus 4.6 on the SWE Bench Pro coding benchmark . These MAI models are co-designed with Microsoft’s custom Maia 200 chip, a software-hardware optimization that yields a 1.4x efficiency gain . Microsoft also announced Project Polaris, its own AI coding model that is slated to replace GPT-4 in GitHub Copilot by August 2026 .

Nvidia countered with its own open-weight offering. Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model, was announced at the conference, and Nvidia’s entire family of open models is now available directly through Microsoft Foundry for development and deployment .

Rethinking Security for Autonomous Agents

The most critical challenge for personal AI agents is trust. An agent powerful enough to act on your behalf across apps and files needs security controls that are just as powerful. The Build 2026 announcements tackle this at the operating-system level .

Microsoft announced new Windows security primitives specifically for agent containment, identity, and manageability. These are foundational OS-level capabilities, not after-the-fact app guards. The centerpiece is the Microsoft eXecution Containers (MXC) framework, a policy layer that defines and instruments isolation and containment for agent processes .

On top of this, Nvidia introduced OpenShell, a secure Windows runtime built on MXC. OpenShell applies kernel-level controls to isolate agent processes and limit their permissions based on the task context, bringing a zero-trust model to autonomous, always-on agents .

This is further complemented by industry governance tools. Microsoft released the Agent Control Specification (ACS), a new open-source standard that allows security teams to define granular policy files specifying what agents can and absolutely cannot do, including when to require human approval . For enterprises, the Agent 365 SDK reached general availability, integrating observability, access controls, and compliance enforcement directly into the agent development lifecycle .

The Developer Platforms Tying It All Together

The hardware, models, and security layers are unified by a suite of new and updated platforms.

• Windows Agent Framework 1.0 shipped as the native runtime for building on-device personal agents .
• Azure Agent Mesh was announced as a new cloud service for orchestrating and federating agent execution across Azure .
• Microsoft Foundry now serves as the unified portal where developers can access Nvidia’s Nemotron models, Microsoft’s MAI models, and deployment infrastructure, all with enterprise-grade security and lifecycle management .
• The Microsoft Agent Platform, powered by the new Microsoft IQ context engine, enables developers to build an agent in GitHub, deploy it to Foundry, and have it auto-optimize across the best available models .

Altogether, the Build 2026 announcements represent a decisive shift. The unified stack co-designed by Microsoft and Nvidia is not just about offering tools; it is about engineering a secure, scalable, and model-agnostic operating system for the coming "agent era."

Alongside the Surface Dev Box, the unified stack includes Nvidia DGX Station for Windows and deskside DGX systems. These provide a scaling path for AI factories, allowing developers to prototype on a local RTX Spark PC and then scale the same agent workloads seamlessly to more powerful local stations or across to Azure’s AI infrastructure . This is the core of the hardware promise: a continuum of accelerated computing from the edge to the cloud.

The Models: From In-House to Open-Weights

The unified stack is model-diverse by design. The partnership featured major model announcements from both companies.

Microsoft’s AI Superintelligence Team released a family of seven new in-house models, headlined by MAI-Thinking-1, the company’s first reasoning model . In an early sign of performance, independent raters preferred it to Sonnet 4.6 in blind tests, and it matched Opus 4.6 on the SWE Bench Pro coding benchmark . These MAI models are co-designed with Microsoft’s custom Maia 200 chip, a software-hardware optimization that yields a 1.4x efficiency gain . Microsoft also announced Project Polaris, its own AI coding model that is slated to replace GPT-4 in GitHub Copilot by August 2026 .

Nvidia countered with its own open-weight offering. Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model, was announced at the conference, and Nvidia’s entire family of open models is now available directly through Microsoft Foundry for development and deployment .

Rethinking Security for Autonomous Agents

The most critical challenge for personal AI agents is trust. An agent powerful enough to act on your behalf across apps and files needs security controls that are just as powerful. The Build 2026 announcements tackle this at the operating-system level .

Microsoft announced new Windows security primitives specifically for agent containment, identity, and manageability. These are foundational OS-level capabilities, not after-the-fact app guards. The centerpiece is the Microsoft eXecution Containers (MXC) framework, a policy layer that defines and instruments isolation and containment for agent processes .

On top of this, Nvidia introduced OpenShell, a secure Windows runtime built on MXC. OpenShell applies kernel-level controls to isolate agent processes and limit their permissions based on the task context, bringing a zero-trust model to autonomous, always-on agents .

This is further complemented by industry governance tools. Microsoft released the Agent Control Specification (ACS), a new open-source standard that allows security teams to define granular policy files specifying what agents can and absolutely cannot do, including when to require human approval . For enterprises, the Agent 365 SDK reached general availability, integrating observability, access controls, and compliance enforcement directly into the agent development lifecycle .

The Developer Platforms Tying It All Together

The hardware, models, and security layers are unified by a suite of new and updated platforms.

• Windows Agent Framework 1.0 shipped as the native runtime for building on-device personal agents .
• Azure Agent Mesh was announced as a new cloud service for orchestrating and federating agent execution across Azure .
• Microsoft Foundry now serves as the unified portal where developers can access Nvidia’s Nemotron models, Microsoft’s MAI models, and deployment infrastructure, all with enterprise-grade security and lifecycle management .
• The Microsoft Agent Platform, powered by the new Microsoft IQ context engine, enables developers to build an agent in GitHub, deploy it to Foundry, and have it auto-optimize across the best available models .

Altogether, the Build 2026 announcements represent a decisive shift. The unified stack co-designed by Microsoft and Nvidia is not just about offering tools; it is about engineering a secure, scalable, and model-agnostic operating system for the coming "agent era."