Azure AI Foundry's 10 most important updates

2025 is widely regarded as the "Year One of AI Agents". Microsoft has released numerous innovative technologies in the past six months, building a complete technology matrix from the infrastructure layer, development tool layer to the scenario application layer, accelerating the implementation of many "super assistant" agents with autonomous decision-making capabilities, forming a complete AI empowerment ecosystem, and helping enterprises to fully carry out AI transformation in four aspects: enriching employee experience, reshaping customer engagement, optimizing business processes, and promoting innovation curves.

At the just concluded Microsoft Build 2025 Microsoft Developer Conference, Microsoft released a number of new products and services, with the goal of empowering every developer to create the future with AI technology . Next, let's take stock of the top ten important updates in Azure AI Foundry (International Edition).

We are continuously expanding the model catalog and introducing cutting-edge models to provide developers with more choices. xAI's Grok 3 is now officially available in Azure AI Foundry (International Edition), and Flux Pro 1.1 in Black Forest Labs and Sora are also about to be released. Currently, the platform has integrated more than 10,000 open source models from Hugging Face. We also support full fine-tuning (including LoRA/QLoRA and DPO technology) to help developers customize models on demand. In addition, we have launched a new developer-level fine-tuning service , which allows developers to test and evaluate various fine-tuning methods with zero threshold without hosting fees. 

Now it's easier to choose the right model for the task. The newly launched model router can automatically match the most suitable model based on your prompt, and it can reduce costs while improving output quality. Starting next month, we will also expand reserved capacity to more models, including Microsoft Azure OpenAI (International Edition) and selected Foundry Models (including models from Black Forest Labs, DeepSeek, Mistral, Meta, and xAI). All models can be accessed through a unified API and MCP server, enabling seamless transition from prototype development to official launch. 

Microsoft Azure AI Foundry Agent Service is officially released, which can help developers easily design, deploy and expand production-level intelligent agents. More than 10,000 companies , including industry leaders such as Heineken, Carvana and Fujitsu, have used the platform to leverage their own data and knowledge to automate complex business processes. This fully managed service provides out-of-the-box templates and operation components, and supports connecting to more than 1,400 enterprise data sources (including third-party systems) such as SharePoint and Microsoft Fabric (International Edition), thereby accelerating the development of context-aware intelligent agents. With just a few clicks, developers can deploy intelligent agents to Microsoft 365 (Teams and Office applications) or platforms such as Slack and Twilio, allowing intelligent agents to be directly integrated into employees' daily work environment.

In real business scenarios, multiple agents are often required to work together. Microsoft Azure AI Foundry (International Edition) now supports multi-agent collaborative orchestration across cloud environments : agents can call each other as connected tools to solve complex problems through professional division of labor and collaboration. The newly added multi-agent workflow introduces a state management layer to achieve integrated control of context management, error handling, and long-cycle process maintenance, which is particularly suitable for scenarios such as financial approval and supply chain management. We have also implemented open interoperability standards : the inter-agent communication protocol (A2A) supports information exchange and task coordination between different agents, and the model context protocol (MCP) ensures consistent sharing and parsing of context data between agents. These standards enable agents to collaborate seamlessly between Microsoft Azure, AWS, Google Cloud, and local environments . In terms of underlying architecture, we are integrating the Semantic Kernel and AutoGen frameworks to support agent orchestration. 

Due to the deep integration between Microsoft Azure AI Foundry (International Edition) and Microsoft Copilot Studio  , developers can achieve a complete closed loop from model selection, fine-tuning to pre-configured agent deployment. For example, Stanford School of Medicine is using Microsoft Azure AI Foundry (International Edition)'s medical agent orchestrator and Copilot Studio to optimize the efficiency of tumor board meetings with customized clinical workflows. 

We are innovating information retrieval technology to support high-level agents. Agent-based search in Microsoft Azure AI Search is a multi-round query engine designed for complex questions. It breaks down user queries into subqueries through conversation context and embedded large language models, performs multiple searches in parallel, and finally generates comprehensive answers with traceable references. Early tests show that this method improves the relevance of answers to complex multi-dimensional questions by 40% . Agent-based search is now available in public preview, allowing your agents to more accurately connect to enterprise data through advanced search technology and provide reliable answers. 

Intelligent agents in production environments require transparent supervision. We are previewing the new observation feature (Foundry Observability) of Microsoft Azure AI Foundry (International Edition), which provides end-to-end monitoring and diagnosis. You will get built-in indicators for latency, throughput, usage, and quality, as well as detailed tracking logs for each agent reasoning step and tool call . In the development stage, the agent sandbox environment now supports evaluation benchmarks and tracking visualization to help you optimize prompt logic. After entering the CI/CD process, we provide deep integration with GitHub and Microsoft Azure DevOps to automate testing and security protection. In the production environment, the unified dashboard linked to Microsoft Azure Monitor can provide you with real-time monitoring and early warning of models and agents.

Microsoft Entra Agent ID  pioneers identity management solutions for enterprise agents, giving you full control over the behavioral permissions of AI agents. The service assigns a unique identity to each agent created through Microsoft Azure AI Foundry (International Edition) or Microsoft Copilot Studio , giving it the same level of management as manual accounts. Agents will appear in the Microsoft Entra directory, making it easier to set up fine-grained access controls. Upcoming features will also support security administrators to configure conditional access policies, multi-factor authentication, and least-privilege roles for agents, and monitor their login behavior. If an agent attempts to access unauthorized resources, it will be blocked like a normal user.

Responsible AI is our bottom line commitment with our customers. We have added a number of new capabilities to ensure the security and controllability of AI systems from the source: Agent Evaluators  can automatically detect the execution process to determine whether it complies with user intent and tool usage specifications; AI red team agents continuously detect system vulnerabilities and biases to help you fix defects before deployment. Our content filtering system has been intelligently upgraded through the "Spotlighting" technology , which is an enhanced improvement to Prompt Shields, which can more accurately detect and block malicious prompt injection attacks , whether the attack comes from user input or external data streams. We have also strengthened the protection mechanism to prevent agents from leaking sensitive information (PII) or performing unauthorized tasks. In terms of security, the platform is deeply integrated with  Microsoft Defender for Cloud  to provide real-time warnings of security threats. In terms of compliance, we provide out-of-the-box integration with  governance tools such as Credo AI, Saidot and Microsoft Purview  to help you track model performance, fairness and regulatory compliance. Microsoft Azure AI Foundry (International Edition) was designed with the security protection and governance system required to build trusted AI.

Not all AI technologies need to run in the cloud, and edge scenarios often have more advantages. Microsoft Foundry Local  is suitable for  localized operation of AI models and agents for Windows and Mac systems . With this solution, you can build cross-platform AI applications that run offline, localize sensitive data processing and reduce bandwidth costs. This creates new possibilities for industrial scenarios with unstable networks, field services in remote areas, etc. We are advancing integration with Azure Arc, and in the future, through the linkage between Microsoft Azure Arc and Foundry, you will be able to centrally manage and update AI deployments on terminal devices. In short, Microsoft Azure AI Foundry (International Edition) will become your AI technology factory , which can efficiently deliver generative AI technology capabilities regardless of whether the data is in the cloud or locally.

We are exploring the next frontier with Microsoft Research in  Microsoft Foundry Labs  . One exciting invention is  Project Amelie , powered by Microsoft Research’s RD Agent, an autonomous agent capable of building a complete machine learning pipeline from a single prompt. Give it a task like “predict customer churn based on our dataset” and Amelie will ingest data, train a model, and generate a deployable solution — an experiment in AI developing AI.