With agents becoming the center of the AI narrative, Microsoft wants to be a bigger part of the conversation.
At Microsoft’s annual Build conference in San Francisco on Tuesday, CEO Satya Nadella announced a deluge of new AI products. Unsurprisingly, agents took center stage, along with the models to support them and the defense products to secure them. Taken together, these offerings seek to fill needs at several different layers of the AI stack while bringing agents to every layer of the enterprise.
Along with revealing its Copilot "superapp" that combines coding, chat, and the Copilot Cowork agent to bring "coding to all knowledge work," Nadella revealed what he calls "autopilots," or long-running, enterprise-grade "claws" that can have connectors, context, and memory — as well as a personality and name.
The first of these agents is Microsoft Scout, a new personal agent that’s built on OpenClaw and can proactively handle tasks like meeting prep, scheduling and routine tasks. However, Nadella said, "In the coming months, we will build this out into a complete digital team of autopilots."
Here’s what else is new:
On the agent side, the company made Microsoft IQ, its enterprise knowledge and intelligence system for agents, generally available. It also unveiled WebIQ, a platform that gives agents access to real-time web search.On the model side, Microsoft introduced MAI-Thinking-1, its first reasoning model, with 35 billion parameters and a 128K context window built to lower token costs, something that’s vital as agents drive up enterprise AI bills. Additionally, the company released MAI-Image-2.5, its latest image model; new voice models with MAI-Transcribe-1.5 and MAI-Voice-2; and MAI-Code-1, its "ultra-efficient" coding model tuned for GitHub.On the defense side, the company unveiled ASSERT, an open-source tool to automate AI safety evaluations; Agent Control Specification, an open-source standard to regulate agent controls; and Codename MDASH, an agentic bug-hunting system. Additionally, Microsoft revealed Frontier Tuning, which applies reinforcement learning within your compliance policies to let agents learn your organization.
"We need all of these pieces of the puzzle and many more, and they have to play well together," Sarah Bird, chief product officer of responsible AI at Microsoft, told The Deep View in an interview. "We need to build an ecosystem where agents can function inside of organizations effectively and be safe and secure."
Our Deeper View #
Microsoft is playing catch-up with its agent and model offerings as it tries to meet the needs of the enterprise AI stack. None of these are particularly groundbreaking, given that frontier labs have dominated the model market and other tech giants have cornered the agent market. Still, there are two reasons Microsoft may have a leg to stand on: existing market penetration and security. Microsoft, being a legacy tech brand with 1.5 billion Office users worldwide, has already established the trust of countless enterprises. Additionally, with agents presenting a major security conundrum, the company has plenty of room to do good work in that arena. And the mix of open-source tools announced at Build offers the kind of granular control and defensive capabilities that enterprises will need to make agents safe and reliable.