Microsoft used its annual Build conference in San Francisco to unveil a wide range of AI products aimed at enterprise customers, with agents, models, and safety tools all getting prominent attention. CEO Satya Nadella framed the announcements as part of a larger effort to make AI more useful across workplace software while also addressing the security concerns that come with agentic systems.
The company highlighted a new Copilot "superapp" that brings together coding, chat, and the Copilot Cowork agent. Microsoft said the goal is to extend coding capabilities beyond software development and into broader knowledge work. Nadella also introduced what he called "autopilots," describing them as long-running, enterprise-ready agents that can combine connectors, context, memory, and even their own names and personalities.
The first of those agents is Microsoft Scout, a personal assistant built on OpenClaw. Microsoft said Scout is designed to help with meeting preparation, scheduling, and other repetitive tasks. Nadella said the plan is to expand that concept into a larger digital team of agents over the coming months.
Microsoft also announced that Microsoft IQ, its enterprise knowledge and intelligence system for agents, is now generally available. Alongside it, the company unveiled WebIQ, a platform that gives agents access to live web search. Together, the tools are intended to help agents gather and use information more effectively inside business environments.
On the model side, Microsoft introduced MAI-Thinking-1, its first reasoning model. The company said it has 35 billion parameters and a 128K context window, and that it is designed to help reduce token costs, which can rise quickly as companies deploy more agent-based workflows. Microsoft also released MAI-Image-2.5, new voice models called MAI-Transcribe-1.5 and MAI-Voice-2, and MAI-Code-1, a coding model tuned for GitHub and described by the company as highly efficient.
The broader message was that Microsoft wants to offer more of the pieces needed to build and run agent systems, not just the front-end experiences. That includes the underlying models, the memory and search layers, and the tools needed to control and monitor how those agents behave.
Security was a major theme in the announcement slate. Microsoft unveiled ASSERT, an open-source tool for automating AI safety evaluations, as well as Agent Control Specification, an open-source standard meant to govern agent controls. It also introduced Codename MDASH, an agentic bug-hunting system, and Frontier Tuning, which uses reinforcement learning within an organization’s compliance policies so agents can learn from enterprise data without stepping outside approved boundaries.
Sarah Bird, Microsoft’s chief product officer of responsible AI, said the company believes the different parts of the stack need to work together if agents are going to be effective in business settings. She told The Deep View that Microsoft wants to build an ecosystem where agents can operate safely and securely inside organizations.
Microsoft is entering a crowded market, where other major technology companies and frontier AI labs already have strong positions in models and agent platforms. But the company has a significant installed base through Microsoft 365 and other enterprise products, which could help it distribute new AI features widely.
The Build announcements suggest Microsoft sees that combination of reach and security as a key advantage. Rather than relying on a single flagship model or agent, the company is trying to assemble a broader toolkit for enterprise AI, one that spans productivity software, model development, and governance.