Microsoft Build 2025: New models, interfaces and tools for AI agents

Contents

The topic of AI agents will dominate this year's Microsoft Build 2025 in-house developer fair in Seattle (May 19-22). At the center of the announcements is the Azure AI Foundry platform (formerly AI Studio), which is intended to offer developers a full-stack tool and bridges the gap between VS Code, GitHub, and Azure.

As part of the build, Microsoft is equipping Foundry with new models, functions, and interfaces from – some of which are open source – which are intended to expand capabilities, increase interoperability and coherence and make developers' work easier. A new model router, for example, automatically selects the model that it believes matches the query for a prompt.

Grok, Black Forest and Finetuning

Foundry is introducing Grok 3 from xAI, Flux Pro 1.1 from Black Forest and Sora from OpenAI (for video) as new models. Microsoft is opening o4-mini for reinforcement fine-tuning (RFT), but only as a preview and only in the East US 2 and Sweden Central regions. There is also access to over 10,000 models from Hugging Face.

On the platform side, Microsoft is deepening the orchestration and infrastructure between agents and data, for example by implementing open standards such as Agent-2-Agent (A2A) and Model Context Protocol (MCP), thus simplifying communication with artifacts on GCP and AWS.

Foundry also offers functions to control the multi-data and multi-agent workflow. At its core, Microsoft combines the Semantic Kernel with AutoGen. An agent catalog is also planned, in which developers and architects will be able to find code examples and configurations from other manufacturers. In the future, local platforms can be operated with a model runtime on Windows or Mac.

Agent identity, security, and governance

There are also tools for observability, agent ID management and ensuring trustworthiness. In the future, a Red Teaming Agent will be responsible for this, which not only checks the code for security vulnerabilities, but also monitors governance and bias before the agents are deployed. It is based on the open-source tool PyRIT and tests how a model reacts to potentially malicious requests. Other tools can be used to monitor AI tools while they are working, for example for corrupt prompts or the disclosure of internal company information.

Microsoft now also offers a fully managed agent service that takes care of infrastructure, orchestration, and data resources. It can access SharePoint, Microsoft Fabric and systems from other manufacturers and connects the agents to Office 365, Slack, Twillio and other applications “with just a few clicks”.

Amelie independently builds agents from prompts

Build also takes a look into the future: Microsoft is working on an agent called Amelie that creates and deploys instant agents: Amelie responds to requests and prompts by independently selecting suitable models, training them on the fly and then sending the result into the pipeline. In an open-source project, the company is working on Magnetic UI, an open human-machine interface.

(who)