Azure AI Foundry: Your AI App and agent factory

To assistance these AI agents go adjacent much almighty and grip much analyzable tasks, they request state-of-the-art models, integrated tools, and built-in governance. At Microsoft Build 2025, we’re unveiling 10 innovations successful Azure AI Foundry to marque this possible.

Software improvement is undergoing a transformation. In the aged world, gathering bundle was a agelong relay—weeks to plan, months to build, and quarters to launch. In the caller world, AI-powered tools crook ideas into prototypes successful hours and accumulation solutions successful days. To execute this, developers request an end-to-end level that seamlessly connects code, collaboration, and cloud. Microsoft is the lone institution wherever each 3 travel together, with Visual Studio Code, GitHub, and Azure forming a unified, AI-native improvement acquisition designed to empower each developer to signifier the aboriginal with AI. We’re delivering a full-stack AI level with built-in information and trust—giving you everything you request to physique and tally AI-powered apps and agents, from unreality to edge. 

In conscionable 2 years, we’ve gone from aboriginal instauration models to AI agents susceptible of analyzable workflows. Microsoft Azure AI Foundry has already grown to enactment much than 70,000 customers, processing 100 trillion tokens past quarter, and powering 2 cardinal regular endeavor hunt queries. What began arsenic an exertion furniture has present go a full-stack level for gathering intelligent agents that present existent concern value.

To assistance these AI agents go adjacent much almighty and grip much analyzable tasks, they request state-of-the-art models, integrated tools, and built-in governance. Today astatine Microsoft Build 2025, we’re unveiling 10 large innovations successful Azure AI Foundry to marque this possible.

1. New models

We’re expanding our catalog with cutting-edge models to springiness developers much choice. This includes Grok 3 from xAI disposable today, Flux Pro 1.1 from Black Forest Labs coming soon, and Sora coming soon in preview via Azure OpenAI successful Foundry Models. We present person implicit 10,000 open-source models from Hugging Face disposable successful Foundry Models. We enactment full fine-tuning (including LoRA/QLoRA and DPO), truthful you tin tailor fine-tunable models to your needs. Additionally, we’re rolling retired a caller developer tier for fine-tuning—no hosting fees, conscionable a elemental mode to experimentation and measure fine-tuning methods without the overhead.

2. Smarter exemplary system

Choosing the close exemplary for each task is present easier. Our caller model router automatically selects the optimal Azure OpenAI exemplary for your prompt—boosting prime portion reducing costs. Starting adjacent month, we’re extending reserved capacity crossed Azure OpenAI and prime Foundry Models (like models from Black Forest Labs, DeepSeek, Mistral, Meta, and xAI), truthful you get accordant show adjacent nether dense load. All these models volition beryllium accessible done a unified API and Model Context Protocol (MCP) server for Azure AI Foundry models, making it seamless to spell from prototype to production.

3. Azure AI Foundry Agent Service 

Azure AI Foundry Agent Service is now mostly available, empowering you to design, deploy, and standard production-grade AI agents with ease. More than 10,000 organizations—including leaders similar Heineken, Carvana, and Fujitsu—have already utilized Azure AI Foundry to automate analyzable concern processes with their ain information and knowledge. This afloat managed service takes attraction of infrastructure and orchestration, and it comes with ready-to-use templates, actions, and connectors to much than 1,400 endeavor information sources (such arsenic SharePoint, Microsoft Fabric, and third-party systems), speeding up improvement of context-aware agents. And with a fewer clicks, you tin deploy your agents into Microsoft 365 (Microsoft Teams and Office apps) oregon different platforms similar Slack and Twilio—bringing agents straight into the tools your employees usage each day.

4. Multi-agent orchestration

Real-world workflows often necessitate multiple agents moving together. Azure AI Foundry present makes that easy, across immoderate cloud. Agents tin telephone each different arsenic tools (connected agents), passing tasks betwixt specialized agents to lick analyzable problems collaboratively. New multi-agent workflows supply a stateful furniture to negociate context, grip errors, and support long-running processes—great for scenarios similar fiscal approvals oregon proviso concatenation operations. We’ve besides implemented open interoperability standards—such arsenic Agent-to-Agent (A2A) communication, to alteration antithetic agents to speech accusation and coordinate tasks, and the Model Context Protocol (MCP) to alteration agents to stock and construe contextual information consistently. These standards guarantee agents tin collaborate seamlessly crossed Azure, AWS, Google Cloud, and on-premises environments. Under the hood, we are unifying our Semantic Kernel and AutoGen frameworks to enactment this seamless cause orchestration.

And due to the fact that Azure AI Foundry powers Microsoft Copilot Studio, you support a continuous loop—from exemplary enactment and fine-tuning to deploying pre-built agents. For instance, Stanford Medicine is already utilizing our healthcare cause orchestrator successful Azure AI Foundry alongside Microsoft Copilot Studio to streamline tumor-board meetings with customized objective workflows.

5. Agentic retrieval

We’re making accusation retrieval smarter to enactment these precocious agents. Agentic retrieval in Azure AI Search is simply a caller multi-turn query motor designed for analyzable questions. It uses speech discourse and an embedded LLM to interruption down idiosyncratic queries into sub-queries, runs aggregate searches successful parallel, and past compiles a broad reply with citations. In aboriginal tests, this attack improves reply relevance by up to ~40% connected complex, multi-part questions. Now successful nationalist preview, agentic retrieval lets your agents connect to endeavor information much effectively utilizing precocious retrieval for accurate, grounded answers.

6. Always-on observability 

To clasp agents accountable successful production, you request visibility into their performance. We’re previewing caller Foundry Observability features that supply end-to-end monitoring and diagnostics. You’ll get built-in metrics connected latency, throughput, usage, and quality—plus detailed hint logs of each agent’s reasoning steps and instrumentality calls. During development, our Agents Playground present shows valuation benchmarks and traces to assistance you refine prompts and logic. As you determination to CI/CD, we connection integrations with GitHub and Azure DevOps to incorporated tests and guardrails. And successful production, a unified dashboard (integrated with Azure Monitor) gives you real-time penetration and alerts for your models and agents.

7. Enterprise-grade individuality for agents

Microsoft Entra Agent ID is the archetypal measurement successful managing cause identities successful your organization, giving you afloat visibility and power implicit what your AI agents tin do. Microsoft Entra Agent ID assigns a unique, first-class individuality to each cause you physique with Azure AI Foundry oregon Microsoft Copilot Studio. This means your AI agents get the aforesaid individuality absorption arsenic quality users—appearing successful your Microsoft Entra directory truthful you tin acceptable entree controls and permissions for each agent. Soon, information admins volition beryllium capable to use Conditional Access policies, multi-factor authentication, and least-privilege roles to agents, and show their sign-in activities. If an cause shouldn’t entree a resource, it volition beryllium blocked conscionable similar a regular idiosyncratic would.

8. Trustworthy AI built-in

Responsible AI is non-negotiable for america and our customers. We’ve added caller capabilities to discover, protect, and govern AI systems from the start. Agent Evaluators now automatically cheque if an cause is pursuing idiosyncratic intent and utilizing tools correctly, flagging issues for developers. An AI Red Teaming Agent perpetually probes your agents for vulnerabilities oregon biases, truthful you tin hole weaknesses earlier deployment. Our contented filtering has gotten smarter with “Spotlighting”—an enhancement to Prompt Shields that better detects and mitigates malicious punctual injections (whether from users oregon incoming data). We’ve besides enabled enhanced guardrails to forestall agents from revealing delicate accusation (PII) oregon straying disconnected approved tasks. On the security side, Azure AI Foundry integrates with Microsoft Defender for Cloud to rise alerts erstwhile threats occur. And to assistance with compliance, we connection out-of-box integration with governance tools similar Credo AI, Saidot, and Microsoft Purview for tracking exemplary performance, fairness, and regulatory requirements. From time one, Azure AI Foundry equips you with the safety, security, and governance tools to physique AI you (and your users) tin trust.

9. Foundry Local

Not each AI needs to tally successful the cloud—sometimes the champion solution is astatine the edge. Foundry Local is simply a caller runtime for Windows and Mac for AI models and agents. With Foundry Local, you tin physique cross-platform AI apps that enactment offline, support delicate information locally, and trim bandwidth costs. This opens scenarios similar manufacturing connected a mill level with spotty internet, oregon tract work apps that request AI successful distant areas. We are besides integrating Foundry with Azure Arc. With the upcoming Azure Arc and Foundry integration, you tin negociate and update on-device AI deployments centrally. In short, Foundry becomes your AI factory—delivering generative AI wherever your information is, whether that’s successful the unreality oregon on-premises.

10. Future innovations for Foundry Labs

We’re exploring the adjacent frontiers with Microsoft Research successful our Foundry Labs. One breathtaking invention is Project Amelie, powered by RD Agent from Microsoft Research, an autonomous cause that tin physique implicit instrumentality learning pipelines from a azygous prompt. Give it a task similar “predict lawsuit churn from our dataset,” and Amelie volition ingest data, bid models, and nutrient a deployable solution—an experimentation successful AI processing AI.

We’re besides rethinking however AI agents interact with radical and each other: Magentic-UI (now unfastened source) provides a ocular canvas to prototype multi-agent workflows and human-in-the-loop interactions, and TypeAgent is bringing semipermanent representation to agents, enabling them to clasp and callback cognition implicit extended periods. In the technological realm, caller AI models similar EvoDiff (for generating caller proteins) and BioEmu (for predicting however proteins alteration shape) are being tested successful Azure AI Foundry to accelerate probe successful biology. These forward-looking projects amusement however Azure AI Foundry is continually innovating—so you’ll ever person entree to the latest breakthroughs successful AI.

The way forward with Azure AI Foundry

Join america astatine Microsoft Build 2025 to spot these caller capabilities successful enactment and larn however they tin alteration your business. We’re excited to enactment with you—our assemblage of developers and customers—to signifier this caller epoch of AI. Together, we’re making AI much accessible, powerful, and trustworthy for everyone. Learn much and get started with Azure AI Foundry.

Resources:

• Review Azure AI Foundry documentation.
• Take the Azure AI Foundry Learn Course.
• Download the Azure AI Foundry SDK.
• Chat with america connected Discord.
• Provide feedback on GitHub.