From Apps to Agents: Nadella’s New AI Computing Layer
Microsoft’s AI agent strategy is a plan to make autonomous, context-aware software agents the primary computing layer that coordinates tasks, data, and devices across work, life, and research, replacing today’s app-centric model with an always-available, multiagent environment. At Microsoft Build, Satya Nadella framed this shift as moving from isolated tools to an “agentic” stack, where infrastructure, models, runtime, and security are tuned for agents rather than traditional applications. Autopilots and agent runtimes based on OpenClaw signal this new direction, as they focus on agents that act on a user’s behalf instead of waiting for explicit prompts. According to Forrester, Microsoft’s updated playbook is “more opinionated, prescriptive, and decidedly full stack,” suggesting the company is no longer experimenting but setting a reference architecture for how Microsoft AI agents should be built, deployed, and governed in the enterprise.
MAI-Thinking-1 and a Bid for AI Independence
At the center of this strategy is the MAI-Thinking-1 model, a 35‑billion‑parameter reasoning system with a 128,000‑token context window that Microsoft describes as designed for “complex multi-step instructions, long context reasoning, and code generation.” Alongside MAI-Thinking-1, the company announced MAI-Image-2.5, MAI-Transcribe-1.5, MAI-Voice-2, and MAI-Code-1, all accessible through Microsoft Foundry and embedded into products from PowerPoint to OneDrive. Commentators have already called this Build “Microsoft’s AI independence day,” as in-house models reduce reliance on its OpenAI partnership and let Microsoft shape a more opinionated AI computing layer. MAI-Thinking-1 is positioned as the brain for many Microsoft AI agents, powering Autopilots and future assistant experiences that run both in the cloud and closer to the edge, including on new hardware like Surface RTX Spark and Solara devices.
Autopilots, Scout, and Agentic Workflows at Scale
Autopilots at Microsoft Build are the clearest sign that the company sees agents as future co-workers rather than chatbots. Built on OpenClaw-based runtimes and tied into tools like Microsoft Scout, these Microsoft AI agents are meant to operate across Teams, Outlook, OneDrive, and SharePoint without constant human prompts. Scout, introduced as an AI assistant that proactively manages calendars, meeting briefs, and project status, reflects a shift toward agents that behave like junior employees embedded in enterprise workflows. Each agent receives a dedicated Entra identity, so IT can specify what data it may see and what actions it may perform, tightening governance as autonomy grows. On Windows, Microsoft Execution Container (MXC) and Windows-native OpenClaw further contain these agents so they cannot freely touch files or system resources, turning Zero Trust ideas into a practical runtime for Autopilots and other multiagent systems.
Surface RTX Spark and Project Solara: Hardware for an Agentic World
Microsoft’s hardware announcements show that its AI computing layer will not live only in the cloud. Surface RTX Spark, described as an AI "data center" for the desk, is aimed at developers and power users who want what Microsoft calls the allure of “unmetered intelligence” for local experimentation and small language models. Project Solara extends this thinking into new device categories, with a chip-to-cloud platform that lets AI agents move across form factors. Microsoft demonstrated a Qualcomm-powered wearable badge and a MediaTek-based desktop companion, both designed so agents can follow users throughout the day instead of being confined to a browser tab. In parallel, Project Soltera, an Android-based agentic OS, targets secure, multiagent environments on non-Windows devices, underscoring Microsoft’s belief that a distributed mesh of agent-ready hardware will define how Surface RTX Spark and Solara-era devices fit into everyday computing.
Developers First: GitHub Copilot App and the Stack Around Agents
Build 2026 reinforced that developers are the main route to spreading Microsoft AI agents. The GitHub Copilot app for Windows aims to make multiagent development less chaotic than juggling many terminals, bringing code generation, orchestration help, and system insight into a single environment. Microsoft’s stack message was tightly structured around layers: infrastructure like Azure Cobalt CPUs, models including the MAI-Thinking-1 model family, an agent runtime secured by MXC and OpenClaw, a context layer built on Fabric IQ, HorizonDB, and Web IQ, and finally tools and observability. Fabric IQ, which combines OneLake, semantics, ontologies, and data agents, is pitched as the data fabric that turns organizational information into an advantage for agentic AI. For developers, the signal is clear: if they adopt Microsoft’s opinionated stack, from Surface RTX Spark dev boxes to GitHub Copilot and Foundry, they get an end-to-end path to building Autopilots that live across cloud and edge.
