To assist these AI brokers change into much more highly effective and deal with extra complicated duties, they want state-of-the-art fashions, built-in instruments, and built-in governance. At Microsoft Construct 2025, we’re unveiling 10 improvements in Azure AI Foundry to make this attainable.
Software program improvement is present process a change. Within the previous world, constructing software program was an extended relay—weeks to plan, months to construct, and quarters to launch. Within the new world, AI-powered instruments flip concepts into prototypes in hours and manufacturing options in days. To attain this, builders want an end-to-end platform that seamlessly connects code, collaboration, and cloud. Microsoft is the one firm the place all three come collectively, with Visible Studio Code, GitHub, and Azure forming a unified, AI-native improvement expertise designed to empower each developer to form the longer term with AI. We’re delivering a full-stack AI platform with built-in safety and belief—supplying you with every little thing you have to construct and run AI-powered apps and brokers, from cloud to edge.
In simply two years, we’ve gone from early basis fashions to AI brokers able to complicated workflows. Microsoft Azure AI Foundry has already grown to help greater than 70,000 clients, processing 100 trillion tokens final quarter, and powering 2 billion every day enterprise search queries. What started as an software layer has now change into a full-stack platform for constructing clever brokers that ship actual enterprise worth.
To assist these AI brokers change into much more highly effective and deal with extra complicated duties, they want state-of-the-art fashions, built-in instruments, and built-in governance. At present at Microsoft Construct 2025, we’re unveiling 10 main improvements in Azure AI Foundry to make this attainable.
1. New fashions
We’re increasing our catalog with cutting-edge fashions to provide builders extra alternative. This contains Grok 3 from xAI out there at present, Flux Professional 1.1 from Black Forest Labs coming quickly, and Sora coming quickly in preview through Azure OpenAI in Foundry Fashions. We now have over 10,000 open-source fashions from Hugging Face out there in Foundry Fashions. We help full fine-tuning (together with LoRA/QLoRA and DPO), so you’ll be able to tailor fine-tunable fashions to your wants. Moreover, we’re rolling out a brand new developer tier for fine-tuning—no internet hosting charges, only a easy option to experiment and consider fine-tuning strategies with out the overhead.
2. Smarter mannequin system
Choosing the proper mannequin for every activity is now simpler. Our new mannequin router routinely selects the optimum Azure OpenAI mannequin on your immediate—boosting high quality whereas lowering prices. Beginning subsequent month, we’re extending reserved capability throughout Azure OpenAI and choose Foundry Fashions (like fashions from Black Forest Labs, DeepSeek, Mistral, Meta, and xAI), so that you get constant efficiency even underneath heavy load. All these fashions will likely be accessible by means of a unified API and Mannequin Context Protocol (MCP) server for Azure AI Foundry fashions, making it seamless to go from prototype to manufacturing.
3. Azure AI Foundry Agent Service
Azure AI Foundry Agent Service is now usually out there, empowering you to design, deploy, and scale production-grade AI brokers with ease. Greater than 10,000 organizations—together with leaders like Heineken, Carvana, and Fujitsu—have already used Azure AI Foundry to automate complicated enterprise processes with their very own knowledge and data. This absolutely managed service takes care of infrastructure and orchestration, and it comes with ready-to-use templates, actions, and connectors to greater than 1,400 enterprise knowledge sources (reminiscent of SharePoint, Microsoft Material, and third-party techniques), dashing up improvement of context-aware brokers. And with a couple of clicks, you’ll be able to deploy your brokers into Microsoft 365 (Microsoft Groups and Workplace apps) or different platforms like Slack and Twilio—bringing brokers instantly into the instruments your workers use day-after-day.
4. Multi-agent orchestration
Actual-world workflows usually require a number of brokers working collectively. Azure AI Foundry now makes that simple, throughout any cloud. Brokers can name one another as instruments (linked brokers), passing duties between specialised brokers to resolve complicated issues collaboratively. New multi-agent workflows present a stateful layer to handle context, deal with errors, and preserve long-running processes—nice for eventualities like monetary approvals or provide chain operations. We’ve additionally carried out open interoperability requirements—reminiscent of Agent-to-Agent (A2A) communication, to allow completely different brokers to change info and coordinate duties, and the Mannequin Context Protocol (MCP) to allow brokers to share and interpret contextual knowledge constantly. These requirements guarantee brokers can collaborate seamlessly throughout Azure, AWS, Google Cloud, and on-premises environments. Below the hood, we’re unifying our Semantic Kernel and AutoGen frameworks to help this seamless agent orchestration.
And since Azure AI Foundry powers Microsoft Copilot Studio, you preserve a steady loop—from mannequin choice and fine-tuning to deploying pre-built brokers. As an example, Stanford Medication is already utilizing our healthcare agent orchestrator in Azure AI Foundry alongside Microsoft Copilot Studio to streamline tumor-board conferences with customized medical workflows.
5. Agentic retrieval
We’re making info retrieval smarter to help these superior brokers. Agentic retrieval in Azure AI Search is a brand new multi-turn question engine designed for complicated questions. It makes use of dialog context and an embedded LLM to interrupt down consumer queries into sub-queries, runs a number of searches in parallel, after which compiles a complete reply with citations. In early checks, this strategy improves reply relevance by as much as ~40% on complicated, multi-part questions. Now in public preview, agentic retrieval lets your brokers hook up with enterprise knowledge extra successfully utilizing superior retrieval for correct, grounded solutions.
6. At all times-on observability
To carry brokers accountable in manufacturing, you want visibility into their efficiency. We’re previewing new Foundry Observability options that present end-to-end monitoring and diagnostics. You’ll get built-in metrics on latency, throughput, utilization, and high quality—plus detailed hint logs of every agent’s reasoning steps and gear calls. Throughout improvement, our Brokers Playground now exhibits analysis benchmarks and traces that will help you refine prompts and logic. As you progress to CI/CD, we provide integrations with GitHub and Azure DevOps to include checks and guardrails. And in manufacturing, a unified dashboard (built-in with Azure Monitor) offers you real-time perception and alerts on your fashions and brokers.
7. Enterprise-grade id for brokers
Microsoft Entra Agent ID is step one in managing agent identities in your group, supplying you with full visibility and management over what your AI brokers can do. Microsoft Entra Agent ID assigns a singular, first-class id to each agent you construct with Azure AI Foundry or Microsoft Copilot Studio. This implies your AI brokers get the identical id administration as human customers—showing in your Microsoft Entra listing so you’ll be able to set entry controls and permissions for every agent. Quickly, safety admins will be capable to apply Conditional Entry insurance policies, multi-factor authentication, and least-privilege roles to brokers, and monitor their sign-in actions. If an agent shouldn’t entry a useful resource, will probably be blocked identical to a daily consumer would.
8. Reliable AI built-in
Accountable AI is non-negotiable for us and our clients. We’ve added new capabilities to uncover, shield, and govern AI techniques from the beginning. Agent Evaluators now routinely examine if an agent is following consumer intent and utilizing instruments accurately, flagging points for builders. An AI Pink Teaming Agent always probes your brokers for vulnerabilities or biases, so you’ll be able to repair weaknesses earlier than deployment. Our content material filtering has gotten smarter with “Spotlighting”—an enhancement to Immediate Shields that higher detects and mitigates malicious immediate injections (whether or not from customers or incoming knowledge). We’ve additionally enabled enhanced guardrails to stop brokers from revealing delicate info (PII) or straying off authorised duties. On the safety facet, Azure AI Foundry integrates with Microsoft Defender for Cloud to lift alerts when threats happen. And to assist with compliance, we provide out-of-box integration with governance instruments like Credo AI, Saidot, and Microsoft Purview for monitoring mannequin efficiency, equity, and regulatory necessities. From day one, Azure AI Foundry equips you with the security, safety, and governance instruments to construct AI you (and your customers) can belief.
9. Foundry Native
Not all AI must run within the cloud—generally the perfect resolution is on the edge. Foundry Native is a brand new runtime for Home windows and Mac for AI fashions and brokers. With Foundry Native, you’ll be able to construct cross-platform AI apps that work offline, preserve delicate knowledge regionally, and cut back bandwidth prices. This opens eventualities like manufacturing on a manufacturing unit ground with spotty web, or area service apps that want AI in distant areas. We’re additionally integrating Foundry with Azure Arc. With the upcoming Azure Arc and Foundry integration, you’ll be able to handle and replace on-device AI deployments centrally. Briefly, Foundry turns into your AI manufacturing unit—delivering generative AI the place your knowledge is, whether or not that’s within the cloud or on-premises.
10. Future improvements for Foundry Labs
We’re exploring the following frontiers with Microsoft Analysis in our Foundry Labs. One thrilling invention is Undertaking Amelie, powered by RD Agent from Microsoft Analysis, an autonomous agent that may construct full machine studying pipelines from a single immediate. Give it a activity like “predict buyer churn from our dataset,” and Amelie will ingest knowledge, practice fashions, and produce a deployable resolution—an experiment in AI creating AI.
We’re additionally rethinking how AI brokers work together with individuals and one another: Magentic-UI, now open-source, is an experimental human-centered agent that completes web-based duties, and TypeAgent is bringing long-term reminiscence to brokers, enabling them to retain and recall data over prolonged intervals. Within the scientific realm, new AI fashions like EvoDiff (for producing novel proteins) and BioEmu (for predicting how proteins change form) are being examined in Azure AI Foundry to speed up analysis in biology. These forward-looking tasks present how Azure AI Foundry is regularly innovating—so that you’ll all the time have entry to the most recent breakthroughs in AI.
The trail ahead with Azure AI Foundry
Be a part of us at Microsoft Construct 2025 to see these new capabilities in motion and learn the way they’ll remodel your corporation. We’re excited to work with you—our group of builders and clients—to form this new period of AI. Collectively, we’re making AI extra accessible, highly effective, and reliable for everybody. Study extra and get began with Azure AI Foundry.
Sources:
