As it tries to reposition itself in the AI race following the end of its exclusive partnership with OpenAI, Microsoft is appealing to enterprise developers by showing how they can use the cloud giant’s devices, tools and capabilities to develop and deploy AI agents.
At its Microsoft Build conference on Tuesday, the vendor introduced a new desktop machine, the Surface RTX Spark Dev Box, a device that runs personal agents and new AI models and platforms to help enterprise developers deploy agentic workflows.
“Microsoft is trying to communicate to the developer audience in particular that it is the best place for them to come and build their AI agents,” said Ed Andersen, an analyst at Gartner. “Everybody wants to grab that developer audience because in this moment of AI agent development, that’s a pretty key part of the market to win.”
Microsoft’s new direction sharply contrasts with its former position in a market where it started as one of the dominant forces, only to end its exclusive partnership with the ChatGPT creator. After OpenAI and Microsoft split earlier this year, both vendors have moved toward greater independence. OpenAI partnered with other cloud providers, including AWS, Google and Oracle. Microsoft has also invested billions of dollars in OpenAI’s chief rival, Anthropic, and has integrated its Claude models into the Microsoft Copilot ecosystem.
New Models
Although integrating other models is beneficial in a market where many enterprises are model-agnostic, Microsoft also showed at the conference that it can develop leading-edge models, especially with its MAI family.
“It’s a proof point,” said Bradley Shimmin, an analyst at Futurum Group. “They want to show they can innovate all on their own.”
The need to innovate in the model arena was evident in the array of new models Microsoft introduced on Tuesday. They include an AI model for desktop computers and the vendor’s first reasoning model, the MAI-Thinking-1, which is also Microsoft’s first homegrown entrant in the hot AI coding arena. It boasts a 128K context window and 35 billion active parameters — considerably bigger than consumer-grade models but not as big as top foundation models. It is optimized for multi-step instructions, long context reasoning and writing code.
Along with MAI-Thinking-1, Microsoft introduced models for text-to-image, image-to-image and voice generation workloads, all available on PowerPoint and OneDrive.
A Focus on Agents
Microsoft is focusing on providing tools that enterprises can use to build AI agents quickly.
“What they’re trying to achieve is to give customers the path of going through the full lifecycle of getting AI and agents into production,” said Mitch Ashley, a Futurum Group analyst.
Microsoft’s strategy for agent-building is to give enterprises more control over the AI they use and where they use the AI, Ashley said.
For instance, the cloud provider introduced several on-device models: Aion 1.0 Instruct, Aion 1.0 Plan and Phi Silica. Aion 1.0 Instruct is a small language model built for text intelligence and fast processing. The Aion 1.0 Plan is for reasoning and tool-calling and is designed to orchestrate local agentic workflows. Phi Silica is designed for Copilot+ PCs and to manage background Windows tasks.
Unmetered Intelligence
For Microsoft, integrating the Aion models into its Surface RTX Spark Dev Box, a developer-focused machine for maximizing compute and memory, is one way to achieve what CEO Satya Nadella called “unmetered intelligence.” The idea is to move away from deploying AI models in the cloud and instead enable developers to run them locally on their computers.
“If we can deliver unmetered intelligence to every desk and every home … it takes us all the way back to the very beginning,” Nadella said during a conference keynote. “You now have a full local agentic loop, you can give it tools access, and build fully-owned agentic applications, without having to run to the cloud.”
Along with a new computer and a machine for developers to run agents locally, Microsoft said developers and IT administrators can create enterprise-grade sandboxed environments for agents in its new OS layer, called Microsoft Execution Containers.
The system enables users to define any requirements they want on their PCs, and Windows will enforce them when the agents run. This also applies to OpenClaw on Windows, which lets users set boundaries for where OpenClaw — the personal agent framework originally developed for Mac OS — runs. Microsoft also launched Microsoft Scout, a personal AI assistant built on top of the OpenClaw framework. It works with Microsoft 365 apps such as Outlook, Teams and OneDrive.
With these AI moves announcements, Microsoft is seeking to reposition itself as an AI platform provider, Ashley said.
“They’re trying to give customers a path all the way up and down the stack, but still provide integration layers across the stack,” he said. He added that while the vendor showed how developers can build agents, it still has much to prove about how these agents will live in applications, which will be the next step.
