Microsoft opens Build 2026 with a sweeping slate of agent-focused announcements, including an in-house reasoning model, agent hardware, and Windows-native AI capabilities. The moves signal Redmond’s intent to own the infrastructure where autonomous AI workers operate.
Microsoft opened Build 2026 in San Francisco today with the clearest "we are betting everything on agents" announcement slate in the company's history. The message was simple: AI is no longer the assistant. AI is the worker. And Windows is the operating system where AI workers live.
The most strategically loaded reveal of the day was MAI-Thinking-1, Microsoft's first in-house reasoning model. The model has 35 billion active parameters and is optimized for multi-step reasoning, long-context tasks, and code generation. Microsoft says the model was trained on clean data, without distillation from external frontier systems.
Microsoft is publicly claiming MAI-Thinking-1 is not a quiet rip of GPT-5 or Claude — the loudest possible signal that Redmond now treats frontier AI as an in-house product, not a vendor relationship with OpenAI.
It is in early access for select partners and ships on Microsoft Foundry and the MAI Playground. The launch was accompanied by a wave of other first-party models, including MAI-Image-2.5 for text-to-image and image-to-image, a transcription model, voice in 10+ new languages, and a lightweight code model for GitHub Copilot and VS Code.
The enterprise reveal of the day was Microsoft Discovery, an agentic AI platform for scientific R&D that reached general availability. Its core Discovery Engine deploys specialized agents to generate and validate hypotheses across enterprise knowledge bases, running a continuous scientific-method loop with security, transparency, and governance controls built in for regulated sectors like food and energy. Microsoft also previewed a local Microsoft Discovery app that lets researchers and students run core capabilities on their own machines with just a GitHub Copilot account.
https://x.com/Microsoft/status/2061880215614247185
The flagship hardware reveal was Project Solara, a chip-to-cloud platform for agent-first systems. Microsoft showed two concept devices — a small desktop unit resembling a wireless speaker, and a portable intelligent lanyard about the size of a small phone. Project Solara is a reference architecture, not a product you can buy. Hardware partners and enterprise customers will build agent-purposed devices on top of it, the same way they built PCs on top of Windows reference designs in the 90s.
Microsoft is staking out the position that ambient AI agents need their own hardware category, and the company wants to own the substrate.
For developers, Microsoft announced the Surface RTX Spark Dev Box — a compact developer machine powered by NVIDIA RTX Spark silicon with 1 petaflop of AI compute, 128 GB of unified memory, and enough headroom to run 120-billion-parameter models locally. No Azure costs, no rate limits, no data leaving the room. The box isn't yet FCC-authorized for sale, so the reveal is real but the ship date isn't. The Dev Box is part of a broader Microsoft and NVIDIA collaboration on a unified developer stack spanning silicon, OS-level AI APIs, and GPU-accelerated analytics.
The Windows OS bet most people will miss is Aion 1.0 Plan, a 14-billion-parameter reasoning and tool-calling model with a 32K context window that now ships in-box as part of Windows. Any application can call into it to reason over user intent, invoke tools, manage files, and orchestrate sub-agents without making a single network call. Alongside it, Microsoft introduced Windows 365 for Agents, a cloud-PC product that provisions Windows sessions specifically for autonomous agents, and OpenClaw on Windows, the agent runtime that powers Microsoft Scout, now exposed to third-party developers.
Microsoft also unveiled a new product category called Autopilots — long-running autonomous agents — and revealed the first of them, Microsoft Scout. Scout runs always-on across Teams, Outlook, OneDrive, and SharePoint, surfacing tasks you need to complete and starting the work before you ask. It is built on OpenClaw, powered by Work IQ, and gated by Entra identity and policy controls.
Autopilots aren't "ask the assistant." They're "the assistant already started."
For backend developers, Microsoft unveiled Azure HorizonDB, a managed PostgreSQL service designed for the agentic-app era. Available in preview today, Microsoft claims it delivers ultra-low latency and 3x faster transactions than self-managed Postgres. It is a direct shot at AWS Aurora and Google AlloyDB. Frontier Tuning, also in preview for select partners, lets organizations train domain-specific frontier models on their own data, workflows, and policies entirely within a single tenant — an unblock for regulated industries that have been blocked from frontier AI by data-residency rules. And Rayfin, an open-source SDK and CLI, takes a description of the app you want to build and generates the full backend, then ships it to Microsoft Fabric as a managed service.
Pull all of it together and the through-line is unmistakable. Microsoft is reframing Windows as the agent operating system — the way Apple reframed macOS as a creative OS in the 2000s. Aion 1.0 ships in Windows. Windows 365 for Agents gives agents a cloud Windows session. OpenClaw lets any developer wire one up. Project Solara devices boot Windows. The RTX Spark Dev Box runs Windows. Scout runs on Windows. Discovery runs on Windows. And under all of it sits MAI-Thinking-1 — proof that Microsoft no longer needs OpenAI to make a frontier model work.
For the last decade, Windows survived because of inertia. Today, Microsoft made the bet that Windows can grow because of agents — because the company that owns the runtime where AI workers live owns the next platform shift.
The keynote didn't bury OpenAI. It quietly built a moat around what comes next.
A handful of smaller-but-real announcements rounded out the day. Microsoft IQ expanded Fabric IQ into a unified system covering how an organization works, what it knows, and real-time global web signals, with Web IQ piping fresh content into AI agents. GitHub Copilot got a new standalone desktop app, demoed working on multiple projects simultaneously with multimodal capability. The Microsoft Agent Framework 1.0 reached general availability, and Copilot Studio's Computer-Using Agents, agent-to-agent communication, and real-time voice all moved from preview to GA. Windows AI APIs got an expansion: Phi Silica gained GPU support, while video super resolution and live captions gained CPU support, opening capabilities previously locked to Copilot+ PCs. A Speech Recognition API entered preview with real-time on-device speech-to-text. Windows 11 also gained native Linux command-line utilities, WSL containers, an "Intelligent Terminal," and simplified developer setup. Microsoft and the Mayo Clinic announced a partnership to co-develop a healthcare-specific frontier model.
Tap a lens to see what this story means for you.
Reader-supported · Daily Brief
Daily brief at 7 AM ET. Top tech stories, every morning. Sourced and fact-checked.
Reader-supported
The Circuitry is a passion project I've always wanted to build, and I love the work behind it.
Running it costs real money. APIs, hosting, time. To keep improving the site and growing this into something useful for everyone, those costs have to be covered.
Any contribution is appreciated. If not, no pressure. Thanks for reading.
A data breach at DentaQuest exposed sensitive records for 2.6 million accounts after ShinyHunters publicly leaked more than 234 GB of stolen data. The compromise includes names, government IDs, health insurance details and other personal information, raising risks of phishing and social engineering.
OpenAI is making ChatGPT’s memory feature smarter with a new synthesis system and extending it to free users for the first time. The changes address staleness, correctness and scalability for hundreds of millions of users across multi-year time horizons.
Pinterest announced a planned $4 billion commitment to AWS through 2031, its largest infrastructure deal ever. The investment will scale AI models for visual search and discovery used by more than 600 million monthly users.
