Microsoft Build 2026: all the news on agents, MAI models, security, developer Windows, HorizonDB, and quantum
Build 2026
Microsoft has outlined a very clear strategy: transforming agent-based artificial intelligence into an infrastructure usable by developers and governable by businesses. The innovations cover the entire stack: business context, proprietary models, security, local development, cloud, databases, scientific research, and quantum computing.
There are numerous updates at all levels: solutions for the development and orchestration of agents, personal agents in the style of OpenClaw, but also hardware, specifically the Majorana-2 chip for quantum computing and the Surface RTX Spark Dev Box, a portable workstation with an NVIDIA GPU designed for the most complex AI workloads.
Microsoft Agent Platform and Microsoft IQ: context becomes the focus for agents
The first area concerns the Microsoft Agent Platform, designed to create agents in GitHub, distribute them on Microsoft Foundry, and make them available in Microsoft Teams, Microsoft 365, or other work environments. Microsoft has focused not so much on the AI model that governs the agent but on how it accesses data and the specific context. Microsoft IQ becomes generally available in GitHub Copilot, Microsoft Foundry, and Copilot Studio. It is the layer that allows agents to use both general knowledge and enterprise knowledge.
Inside Microsoft IQ are several components:
- Work IQ helps connect agents to how work is performed in Microsoft 365, enterprise systems, and external sources: people, emails, documents, meetings, and relationships between these elements. The Work IQ APIs will be generally available from June 16.
- Fabric IQ provides a shared semantic foundation on the company’s structured data. Foundry IQ connects these layers and enables information retrieval in both enterprise knowledge and the web in real-time.
Web IQ: grounding web for AI agents
One of the most interesting announcements is about Web IQ, a web research stack designed for AI agents. It is agnostic to the model, supports MCP, and is designed to retrieve information from the web as well. According to Microsoft, Web IQ delivers grounding almost 2.5 times faster than the best available alternative. This data is interesting because retrieving current information remains one of the operational limits of agents: without a robust connection to external sources, automation risks being confined to internal knowledge or the context provided at execution time.
Microsoft Scout: the always-on personal agent for work
Microsoft introduced Scout, a personal agent for work initially targeted at Frontier companies. It is based on OpenClaw and Work IQ and is designed to manage routine tasks without waiting for explicit instructions: preparing meetings, calendar conflicts, recurring tasks, and using tools like Teams and Outlook. Microsoft describes it as an always-on agent, capable of understanding how the user works and moving within the already used tools in the company. Essentially, it is a type of OpenClaw but with greater control.
The arrival of MAI models: reasoning, images, voice, transcription, and code
At Build 2026, Microsoft unveiled a new family of proprietary AI models. The first is MAI-Thinking-1, a reasoning model developed by the AI Superintelligence Team. MAI-Thinking-1 is a mid-range model with 35 billion active parameters and a context window of 256K tokens. It has been trained from scratch, without distillation, on enterprise-grade, clean, and commercially licensed data. Microsoft positions it for complex instructions, multi-step reasoning, long contexts, and code generation. It is available in private preview on Foundry.
The MAI family also includes MAI-Image-2.5 and MAI-Image-2.5 Flash, the first Microsoft models for text-to-image and image-to-image flows. They are already integrated into PowerPoint, will arrive on OneDrive, and are available on Foundry.
On the audio front, there are MAI Transcribe 1.5, supporting 43 languages with streaming coming, and MAI-Voice-2 with a Flash variant, available in over 15 additional languages with new voice options.
For software development, there is MAI-Code-1, a model optimized for inference and tuned for GitHub. It is already available in Copilot and Visual Studio Code.
The MAI models will not remain confined to the Microsoft ecosystem. They will also be available on Fireworks AI, Baseten, and OpenRouter. Fireworks AI is also generally available on Foundry, with enterprise governance and data residency on Azure.
Frontier Tuning: training agents on the real way the company works
Microsoft has also announced Frontier Tuning, available in private preview. This feature applies reinforcement learning within the organization’s compliance perimeter. The idea is to allow agents to improve by using data, domain knowledge, and business operational flows. This is not simple prompt customization: the goal is to refine the behavior of agents based on the concrete way the company works.
Agent 365, ASSERT, Agent Control Specification, and MDASH: security enters the agent lifecycle
The security aspect is one of the most substantial. Agent 365 for local agents extends Entra, Defender, and Purview into a single control plane. It is designed to observe, govern, and protect agents hosted in different environments built with different frameworks.
Microsoft also introduces an end-to-end open stack for AI agents on any framework. It is based on two open-source projects: ASSERT, which stands for Adaptive Spec-driven Scoring for Evaluation and Regression Testing, for security assessment driven by policies; and Agent Control Specification, which aims to standardize where and how to apply controls in the agent lifecycle. There is also Codename MDASH, a multi-model agent security system. It uses over 100 agents to identify exploitable bugs by analyzing data flow, business logic, and exploit chains. Contextual fixes are applied directly to the Defender Portal.
Surface RTX Spark Dev Box: a local machine for heavy AI workloads
Microsoft brings AI to local development with the Surface RTX Spark Dev Box, designed for extended training, agent pipelines, and local fine-tuning. The machine is based on NVIDIA RTX Spark, offering up to a petaflop of AI computing power and 128 GB of unified memory. It can run LLMs locally with up to 120 billion parameters and contexts up to 1 million tokens, without relying on cloud GPU instances.
It comes with Windows Services for Linux 2, native GPU passthrough, CUDA support, Visual Studio Code, GitHub Copilot, and other pre-installed tools. Availability is expected towards the end of the year in the US on Microsoft.com.
Windows becomes a runtime for agents
At the operating system level, Microsoft introduces Microsoft Execution Containers, or MXC, now in preview. The technology allows the creation of enterprise sandbox environments for agents, with containment managed directly by the operating system. The developer or administrator defines the requirements once, and Windows applies them wherever the agent is executed. OpenClaw on Windows already uses this technology for multi-step workflows. NVIDIA's secure runtime OpenShell for autonomous agents also uses MXC and adds policy management, inference routing, and personal data obfuscation.
Hosted agents in Foundry Agent Service: cloud sandbox for agents
When agents move to the cloud, Microsoft introduces hosted agents in Foundry Agent Service, now in preview. The service offers instant session sandboxes, isolated execution, persistent memory, and elastic scalability. Microsoft presents them as the foundational block for AI agents, similar to what containers represent for cloud-native applications.
GitHub Copilot becomes a desktop app
The new GitHub Copilot app, now in preview, brings agent-based development into a native desktop experience. The developer can start with an idea, a request, or a pull request, launch multiple agent sessions in parallel, and follow changes through to review, continuous integration, and merge. Each session uses git worktree, allowing the work to remain separate. Copilot performs tasks, but the developer retains control of the process.
Project Rayfin and Replit: from prototype to production
Microsoft also introduces Project Rayfin, now in preview. It is a managed backend-as-a-service within Microsoft Fabric, defined through workflows based on GitHub. The practical problem it addresses is that generating an application with AI is becoming faster, but bringing it into production still requires databases, APIs, authentication, and infrastructure. Rayfin aims to reduce this transition.
The integration with Replit creates a pathway from prototype to enterprise implementation, with governance from the start.
Azure HorizonDB: managed PostgreSQL for agent applications
For databases, Azure HorizonDB arrives, a fully managed PostgreSQL service on Azure. It is designed for agent applications with high performance and reliability requirements. In Microsoft's internal tests, HorizonDB delivers over three times the throughput compared to comparable self-managed configurations. This data should be understood for what it is: an internal benchmark, but it signifies Microsoft's desire to enhance the data tier for more complex AI applications.
Microsoft Discovery: AI agents for scientific research
Microsoft Discovery is an agent-based AI platform built on Azure to support scientific workflows. The platform is already being used in various fields. BHP, for example, employs it to accelerate research on copper leaching. Syensqo uses it for semiconductor research. GSK is adopting it in the development of new drugs. Microsoft has also announced a free local Discovery app for the scientific community. It is available in preview and requires a GitHub Copilot account.
Majorana 2: Microsoft’s new quantum chip
The final announcement pertains to quantum computing. Microsoft unveiled Majorana 2, a next-generation quantum chip. According to Microsoft, Majorana 2 achieves an average qubit lifespan of 20 seconds, with peaks of up to a minute, and reliability 1,000 times higher than the previous generation. The stated goal is to reach a million qubits on a chip the size of a palm. Microsoft connects this development to agent-based AI and indicates 2029 as the horizon for a scalable quantum machine.
A comprehensive AI platform, but all to be verified in the field
Build 2026 does not bring a single central announcement but a sequence of tiles that move in the same direction: more contextualized AI agents, proprietary models, local and cloud development tools, governance, security, and data infrastructure. The most interesting part is the integration between GitHub, Foundry, Microsoft 365, Azure, Windows, and Copilot. The part to be verified will be the actual adoption in companies: it is not enough to have more powerful agents; it is necessary to understand how simple it will be to govern them, integrate them with existing data, and bring them into production without increasing complexity, costs, and risks.