(Photo by JASON REDMOND/AFP via Getty Images)
SEATTLE—Microsoft is now focused on "building the open, agentic web at scale," CEO Satya Nadella said in an opening keynote at the Build conference this week, an ambitious vision of giving users a stable of Copilot agents who do things for them, from handling customer service calls to checking company policies.
Developers will be tasked with creating individual agents and secure experiences that work together. It’s a new approach that means lots of new tools for developers, Build’s target market.
Nadella announced all sorts of software for developing AI applications, including enhancements to Microsoft’s Azure AI Foundry and Copilot Studio developer tools, as well as Visual Studio and GitHub. I was particularly glad to see a focus on open-source and industry-standard protocols; the tech won’t reach its full potential without them.
According to Nadella, we’re in "the middle innings of another platform shift," equivalent to 1991 when Win32 developer tools were rolling out, or 1996, when a variety of companies built new development tools designed for the internet.
M365 Copilot Gains Researcher and Analyst Agents
As more of a corporate user and manager, I was very interested in the changes to Microsoft 365 Copilot—which Nadella described as the biggest update since Teams launched. Just how big? Nadella said that the overhaul turns Copilot into "the UI for AI."
Copilot now provides five distinct capabilities. It starts with the Chat interface common to AI assistants. Then it adds Search, which can look through your personal documents and anything else you have access to, either in Microsoft applications or connected third-party apps. Notebook lets you gather results from Chat or Search and organize them, then allows you to share the results in the company’s Pages collaboration view. Create lets you generate images and now does things such as turn PowerPoint presentations into videos.
Now Copilot is adding Agents; the first two are Researcher, which uses deep reasoning to search your documents and the web to create a research report, and Analyst, which can look through lots of source information, including spreadsheets, to do things like find specific data or compare results. For now, these agents are only available to organizations in Microsoft’s Frontier program (typically larger enterprises), but the plan is to roll these out more broadly and eventually allow for other software companies and individual organizations to create their own agents.
To make all this work, Microsoft is setting up an Agent Store, where organizations will be able to publish and get new agents, and making big improvements to the Copilot Studio where individuals or organizations can build simple agents. There will be support for multi-agent orchestration, so agents can invoke other agents and they can work together. For example, you may have an agent for onboarding a new hire, which in turn would call on specific agents from HR, finance, or IT to ensure all the appropriate forms are filled out for each department.
Another new feature is Copilot Tuning, which lets organizations fine-tune models based on their own data.
Nadella said the agent concept is already gaining traction, with over 1 million agents created in the past year. Together, the setup provides "expertise at your fingertips," Nadella said. Sound familiar? It’s a play on a phrase Bill Gates used decades ago: "information at your fingertips."
AI Foundry Adds Grok, Multi-Agent Orchestration
(Photo by JASON REDMOND/AFP via Getty Images)Nadella announced new versions of Visual Studio and Visual Studio Code, and said the latter will be released as open source.
However, the GitHub Copilot Coding Agent seems to be the most important reveal. Instead of acting as a "pair programmer" that checks your work, the agent can now take on specific development tasks, acting more like a "peer programmer."
"All the apps will have to evolve to become these full, stateful applications that are multimodal and multi-agent," Nadella said, as opposed to an app that just uses an application programming interface (API) or a chatbot to call a single model, return an answer, and move on to the next
AI Foundry already has 70,000 organizations using it, and Microsoft is now processing five times the number of tokens than it did a year ago.
In one portion of the presentation, which included video talks with OpenAI's Sam Altman and xAI's Elon Musk, Nadella pointed out the huge number of models—1,900—that AI Foundry now supports. For OpenAI, Nadella said, Microsoft has added 15 new models this year; support for the Sora video tool is expected next week. Meanwhile, it is also announcing a new model router for Azure AI Foundry.
When it comes to xAi, Microsoft is now supporting the latest Grok 3.5 models. In addition, Nadella said, the platform supports models from Mistral, Meta's Llama, Black Forest Labs, and DeepSeek. Developers can also choose one of the 11,000 models on Hugging Face.
Much of the effort here seems to be focused on ways to improve agents' interactions with other agents and services. This includes Foundry Agent Services, which can be used to orchestrate multi-agent workflows. It’s also supporting the Model Context Protocol (MCP), first introduced by Anthropic, which is rapidly becoming a standard for governing agents’ access to external data, and for Agent2Agent, a protocol first introduced by Google for connecting the agents.
Other changes include making it easier to switch among models, conduct better searches, and take a model from AI foundry and use it within Copilot Studio.
Building Local AI Applications on Windows
Nadella also talked a lot about building AI applications that run "on the edge," or on local PCs. To that end, he introduced Foundry Local, a way of running models on Windows or macOS machines.
For instance, the company announced Windows AI Foundry, an enhanced version of Windows Runtime, for developers creating applications that run on the Windows client. This includes the Windows Copilot and Windows ML runtime environments, as well as a LoRA (Low-Rank Adaption) feature for fine-tuning Microsoft’s Phi Silica model, available for preview on Snapdragon-based Copilot+ PCs and coming to AMD and Intel systems later.
It also includes support for semantic search (i.e., search by meaning, including image search) and lexical search (i.e., search by exact words), as well as RAG (Retrieval Augmented Generation), so developers can base the results on a specific set of data, such as customer support documents or company policies.
Nadella also announced native support for MCP within Windows itself, meaning that agents can query things like the file system or system settings.
Opening the Web and Other Sources to AI Agents
Microsoft CTO Kevin Scott laid out a quick vision for the tools needed to build an agentic web and talked a bit about what’s missing from the AI ecosystem, including a better way of handling memory.
Microsoft "wants to solve everyone's problems," Scott said. One big caveat, though: "For this vision of really powerful agents to come into existence, they have to do some things they can't do right now."
Open standards such as MCP and Agent2Agent, are crucial. Scott introduced NLWeb, an interesting concept designed to make it easy for websites to provide a conversational interface for users or agents, using the model of their choice and their own data. This tool, which Microsoft says builds on MCP, is similar to the role HTML plays on the web today.
Later, Nadella said that while the web democratized the creation of intelligence, NLWeb "democratizes the aggregation of this intelligence" by letting agents gather data from all sorts of sources. I agree that the concept of openness—whether in MCP, A2A, or NLWeb—is likely to be very important to how AI tools develop in the near future.
Data and Infrastructure Improvements
Nadella announced a number of additions to Microsoft’s data products, including SQL Server 2025, which integrates Cosmos DB and Azure Databricks directly into Foundry, GenAI-powered reasoning in PostgreSQL, and a new Digital Twin Builder.
I was particularly interested in Copilot for PowerBI, which should let you "chat" with your data, letting you do things like ask, "How do sales in Chicago look this quarter compared to last year?"
There was relatively little infrastructure discussion during the keynote, but Nadella reiterated Microsoft's aim "to offer the lowest-cost, highest-scale infrastructure." (In other words, deliver the most tokens per watt per dollar.) He said that there are multiple S-curves driving prices down now, including classic Moore's Law reductions in computing costs plus system-software and model optimization.
Nadella did say Microsoft is rolling out Nvidia GB200 systems at scale, with much higher throughput, and showed a video conversation with Nvidia CEO Jensen Huang that focused on scaling. In passing, Nadella noted that Microsoft is installing huge amounts of optical fiber for a very fast backbone between data centers, as well as its own Maia and Cobalt chips. (Teams and Defender are already running on Cobalt, he said, and many third parties are looking at it as well.)
Nadella concluded with an introduction of Microsoft Discovery, an agent platform for science. A demo showed a researcher using the tool to find a chemical with properties they needed and run simulations on it. According to Nadella, researchers created this chemical, tested it, and it works.
All in all, there were an almost unbelievable number of AI-related product announcements (more than could fit into the keynote) and some very engaging demos. Still, there's a big difference between a great demo and a set of applications that are ready for production use at massive scale. That will require some of the less flashy tools for things like observability, security, and governance.