Microsoft Build 2025: The Biggest News In AI, Agents, Windows

These are some of the biggest reveals to come out of Microsoft’s annual Build conference, which runs through Thursday in Seattle.

Support for Anthropic’s Model Context Protocol (MCP) and Google’s Agent2Agent (A2A) protocol to foster adoption of artificial intelligence agents. An open-source project for building the AI era’s version of HTML. And a host of AI capabilities in preview across the Microsoft productivity portfolio.

These are some of the biggest reveals to come out of Microsoft’s annual Build conference, which runs through Thursday in Seattle.

[RELATED: Human Security CEO: Agentic AI Will Require ‘Trusted Agents’]

Microsoft Build 2025

The innovations come as increased global uncertainty around tariffs appears to not be hampering enterprise investment in cutting-edge new AI technology. Microsoft reported in April during its latest quarterly earnings call growth across a variety of AI products, with GitHub Copilot now reaching more than 15 million users, up over fourfold year over year and Microsoft 365 Copilot serving hundreds of thousands of customers worldwide, with use up threefold year over year.

Redmond, Wash.-based Microsoft, which has a worldwide ecosystem of more than 500,000 partners, echoed what other technology vendors including Google and IBM have said in recent weeks about AI investments withstanding concerns around price increases as the U.S., China and other nations negotiate tariffs.

Read on for the biggest news to come out of this year’s Microsoft Build.

MCP Support

Microsoft is the latest vendor to invest in widespread support for Anthropic’s Model Context Protocol (MCP), with first-party support across GitHub, Copilot Studio, Dynamics 365, Azure AI Foundry, Semantic Kernel and Windows 11.

Windows developers will receive an early MCP platform capability preview “in the coming months” for feedback. MCP Servers need a package identity, code signing, exposed interface security testing and other security requirements before they appear in the Windows 11 MCP server registry to prevent tool poisoning and other attack types.

Microsoft and GitHub are also now MCP Steering Committee members, according to the vendor.

Earlier this month, Microsoft detailed a deeper partnership with Google around the latter’s Agent2Agent protocol.

GitHub Copilot Updates

GitHub has a preview for an asynchronous coding agent integrated in its platform. The agent can autonomously refactor code, improve test coverage to fix defects and even collaborate with other agents on more complex tasks.

However, it is worth noting that starting June 4, all GitHub Copilot plans will start billing for what Microsoft calls “premium requests.” The coding agent will use one premium request per model request the agent makes.

Premium requests are based on model multipliers and features. An example of a premium request is using the AI model GPT-4.5, which has a 50x multiplier, to ask Copilot Chat a question–which would count as 50 premium requests. Paid plan users can use base AI models without the interaction counting against their monthly premium request allowance.

The agent is available to Copilot Enterprise and Copilot Pro+ users. Users choose the repositories where the agent is enabled. Administrators need to turn on the policy for Copilot Enterprise users.

Users can activate agent mode in Visual Studio, Jetbrains, Xcode and other integrated development environments (IDEs). Agents push commits to a draft pull request, allowing users to track their steps through session logs. Branch protections and other existing security policies apply to the agents.

GitHub Models

Microsoft has also launched a preview of the GitHub Models models-as-a-service (MaaS) offer for embedding models from Azure AI Foundry into the GitHub user experience and into repositories.

Users can leverage these models for pull requests, commits, code reviews and ontinuous integration and continuous delivery (CI/CD).

Developers can automate model evaluations for finding the best option based on cost and performance. Users can evaluate models from Microsoft, Microsoft-backed OpenAI, Meta, Mistral and more.

Microsoft also revealed at Build 2025 that “over the next few months,” it will make AI-powered GitHub Copilot extensions capabilities part of the Visual Studio (VS) Code open-source repository.

The vendor is doing this in part because of the improvement of large language models (LLMs) over the last few months, the proliferation of common user experience (UX) AI interaction treatments across multiple editors and to create more transparency in what data is collected by AI editors, according to Microsoft.

Local Windows AI Development

Microsoft has previews available for a variety of capabilities meant to allow local AI development with Windows.

A new iteration of Windows Copilot Runtime called Windows AI Foundry promises a unified platform to take AI developers from model selection to optimization then fine-tuning and then deployment across client and cloud. It includes a new iteration of DirectML called Windows ML.

Previews within Windows AI Foundry include a simplified deployment experience for shipping production apps without the need to package ML runtimes, hardware execution providers and drivers. Another preview is for AI Toolkit for Visual Studio (VS) Code tools that support model and app preparation with proprietary or open-source models.

A private developer preview available “in the coming months” with select partners is native support for Anthropic’s Model Context Protocol. MCP should allow applications to easily participate in agentic interactions, with developers having the opportunity to augment agents’ skills and capabilities local on Windows PCs.

A public preview is now available for low-rank-adaption (LoRA) for Phi Silica to fine-tune in-built small language models (SLMs) with custom data. The preview applies to Qualcomm Snapdragon X Series neural processing unit (NPUs), with access to Intel and Advanced Micro Devices (AMD) Copilot+ PC users “in the coming months.”

A preview is available for semantic search application programming interfaces (APIs), which should aid developers with making search experiences with their own app data, according to Microsoft. The APIs run locally on all device types and support retrieval-augmented-generation (RAG).

Coming to preview for Windows Insider Program members “in the coming months” is a command-line text editor called Edit on Windows. Already open source, this editor is meant to allow developers direct file editing in command lines.

Azure AI Foundry Changes

Microsoft has made more models available for its Azure AI Foundry catalog and has more coming. Grok 3 by Elon Musk’s xAI is now available, with Black Forest Labs’ Flux Pro 1.1 “coming soon” and Sora “coming soon in preview.”

A public preview is available for agentic retrieval in Azure AI Search. Retrieval is a multi-turn query engine meant for complex questions. Retrieval uses conversation context and an embedded LLM to break user queries into sub-queries. It then runs multiple searches in parallel and compiles an answer with citations. Agent retrieval allows agents to connect to enterprise data in a better manner for more accurate, grounded answers.

A preview for Foundry Observability provides end-to-end monitoring and diagnostics to make sure agents are functioning correctly in production. Users can see latency, throughput, usage, quality and other metrics. Users also can view trace logs for each agent’s reasoning steps and tool calls.

Now generally available is Azure AI Foundry Agent Service. The fully managed service offers templates, actions and connectors to Microsoft SharePoint, Microsoft Fabric and more than 1,400 other enterprise data sources, including third-party ones.

Azure AI Foundry Agent Service users can orchestrate multiple specialized agents to bring Semantic Kernel and AutoGen into a single software development kit (SDK) and other complex tasks.

Microsoft 365 Copilot News

Starting in June, Microsoft will roll out M365 Copilot Tuning as part of a Copilot Tuning Early Adopter Program for customers with at least 5,000 Copilot licenses.

Users can leverage Copilot Tuning to train models that perform domain-specific tasks with company data. The new capability follows organization permissions for underlying training data.

Also coming to public preview in June is Microsoft Purview Information Protection for Copilot Studio agents that use Microsoft Dataverse. The extension is meant to enhance data security, prevent leaks and allow automatic data classification at scale.

Now in public preview is multi-agent orchestration in Copilot Studio, allowing agents to exchange data, collaborate on tasks and divide up work based on an agent’s expertise.

Another public preview is for users to bring their own models to Copilot Studio through an Azure AI Foundry models integration, giving users access to 1,900-plus models. A public preview for Microsoft Entra Agent ID allows for Copilot Studio and Azure AI Foundry created agents to have an identity automatically, with no additional work by developers and visibility and control for security administrators. Entra Agent ID aims to avoid agent sprawl, according to Microsoft.

Microsoft moved the agent feed hub into public preview for users looking to oversee agent teams in Power Apps. They can see pending tasks across a team of agents. The hub also flags exceptions where agents get stuck for real-time resolutions.

The vendor revealed a series of generally available capabilities coming to Copilot in May, including the ability to create a Copilot Page with a click on mobile devices and then share or edit that page on the device. In May, the ability to turn a page into a Word document with one click will also become available.

And now generally available is a Microsoft 365 Agents Software Development Kit (SDK) for testing and debugging agents in Copilot, Teams, on the web and in the Agent Playground before deployment.

Also GA– Solution Workspace in Power Apps for developers to use natural language to generate pages, data tables and more.

NLWeb, Edge Improvements

Microsoft used Build 2025 to introduce its NLWeb open project for creating natural language interfaces for websites, with the vendor comparing NLWeb’s potential in the agentic internet to that of HyperText Markup Language (HTML).

NLWeb allows users to query site contents with plain English as if using an AI assistant or Copilot. Every NLWeb instance is an MCP server so that website content can get discovered and accessed by agents if desired. NLWeb supports all major operating systems, major models and vector databases.

Early participants in NLWeb include Chicago Public Media, Eventbrite, Snowflake, Tripadvisor and Shopify.

As for Microsoft’s Edge browser, the vendor revealed that in June, PDF content translation into 70-plus languages will become generally available (GA).

Microsoft also introduced a web content filter preview for Edge for Business for no additional cost to schools and small businesses exclusively standardized on the browser, gaining the option to block millions of inappropriate websites.

Microsoft Teams Agent Capabilities

Microsoft revealed a series of previews coming in May to Teams users interested in more AI capabilities.

Developers will have the ability to build Teams agents with Google’s A2A protocol. Those agents will be able to exchange messages, data and credentials without centralized intermediaries.

Agentic memory on Teams will allow agents to recall user interactions from the communication platform for better performance. And developers will have the chance to build agents for private and group use from the meeting surface.

More Microsoft Security News

Among the security updates Microsoft revealed during Build 2025 is a preview for a spotlighting capability in Azure AI Foundry. This allows for users to identify adversarial prompts embedded in external data sources to reduce cross-domain injection risks.

A preview for a Microsoft Defender for Cloud integration with Azure AI Foundry promises real-time security recommendations and runtime alert monitoring in the AI development workflow. Defender for Cloud security insights will enter the Azure AI Foundry portal by June.

And a preview for personally identifiable information (PII) detection content filters aims to automatically detect and redact sensitive information to meet compliance and data security. Users can even tailor content filers with custom categories in Foundry Model.

Already generally available are prompt shields protect against jailbreaks and injection attacks.

A Purview SDK is now in preview for embedding the data security and compliance tool’s capabilities directly into AI apps. And previews are available for a host of Purview attributes for AI interactions within Azure AI workloads. Those attributes include insider risk management and data security posture management (DSPM).

And for developers already looking past the AI era and thinking about the age of quantum computers, Microsoft said that post-quantum cryptography (PQC) capabilities will come to Windows Insiders Build 27852 and higher and SymCrypt-OpenSSL version 1.9.0 and higher.

Upcoming Database Product Previews

Microsoft’s database products updates from Build 2025 include a preview for Cosmos DB (NoSQL) in Microsoft Fabric, expanding access to enterprise-grade dynamic scalability, reliability and low latency serving for semi-structured data.

Coming to preview over “the next few weeks” is the ability to add Microsoft Fabric data to any custom agent built in Microsoft Copilot Studio before deploying them in Teams, Copilot and other channels. The Fabric data agents will follow data access permissions from OneLake.

Another upcoming preview is digital twin builder in Real-Time Intelligence in Fabric. Users can leverage digital twin builder to manage digital representations of real-world environments with a no-code, low-code interface. Users will have the ability to map data from physical assets and systems.

In the next few months, users will have the ability to chat with data in Power BI through M365 Copilot, allowing them to find reports and answer questions without needing to find the report.

SQL Server 2025 is now in preview, bringing advanced semantic search to users as well as increased workload uptime and other benefits over prior iterations.