[MS] What’s new in Microsoft Foundry | October and November 2025 - devamazonaws.blogspot.com
TL;DR
- Move from prototype to production in hours, not weeks: The new Microsoft Agent Framework and Hosted Agents let you build, test, and deploy multi-agent AI systems with enterprise-grade security—no Kubernetes or container headaches.
- Orchestrate any model, anywhere: Model Router and BYO Model Gateway let you mix and match thousands of models (including Claude, GPT, and your own) with unified governance and compliance—no code changes required.
- Ship agents to Teams and M365 with one click: New low-code/no-code tools, templates, and deployment channels make it easy to launch and scale AI agents for your users.
- Build smarter, more reliable workflows: Multi-agent orchestration, persistent memory, and deep Microsoft 365 integration enable robust, context-aware solutions for complex enterprise scenarios.
- Fine-tune and innovate faster: Redesigned UI, support for reinforcement fine-tuning (RFT) on GPT-5, and parity for non-OpenAI models like Mistral accelerate custom model development.
- Access the best models in one place: Azure is now the only cloud with both Anthropic’s Claude and OpenAI’s GPT models—choose the right tool for every job.
- Build with confidence: Foundry Control Plane, new guardrails, and granular security controls give you enterprise-grade observability, compliance, and peace of mind.
Agents
Microsoft Agent Framework + Hosted Agents
Microsoft Agent Framework is an open-source SDK and runtime for building, deploying, and managing multi-agent AI systems. It combines the enterprise-ready stability of Semantic Kernel with the innovative orchestration patterns from AutoGen, creating a unified foundation for both experimentation and production. Current agent frameworks are fragmented and lack enterprise features like observability, compliance, and durability. Microsoft Agent Framework addresses these gaps by providing security, governance, and scalability for agentic AI applications.- Rapid prototyping: Build and test agentic workflows locally, then deploy seamlessly to production—no need for complex containerization or infrastructure setup.
- Cross-cloud flexibility: Connectors for Azure, AWS, and GCP let you leverage best-in-class models and tools, regardless of cloud provider.
- Enterprise-grade features: Built-in observability, identity, governance, and autoscaling mean you can focus on agent logic, not ops.
- Unified build-and-run story: Move from prototype to production with transparent consumption-based pricing and robust compliance.
Foundry Agent Service
At Microsoft Ignite 2025, we introduced the next evolution of Foundry Agent Service, managed runtime for agentic apps, supporting multi-agent workflows, persistent memory, and enterprise governance. It delivers expanded enterprise capabilities and unified developer experience across the Foundry portal and SDK, enabling seamless orchestration, deployment, and governance for agents built with any framework or model. The release adds new interoperability layers, intelligent context systems, managed hosting, and deeper Microsoft 365 integration. Build, host, and scale intelligent agents securely and at enterprise scale.- Multi-agent orchestration: Coordinate complex workflows with persistent state, error recovery, and context sharing.
- Enterprise governance: Secure, compliant deployments with role-based access, auditability, and lifecycle management.
- Telemetry and monitoring: Built-in observability lets you debug, optimize, and scale agentic solutions confidently.
"Bring your own models"
The "bring your own model" feature allows enterprises to connect Foundry-hosted models through any AI gateway services like Azure API Management, Mulesoft, and Kong, to Foundry Agent Service. The Agent Service honors pre/post LLM hooks, policy-based model selection, and multi-region/multi-provider load-balancing with automatic failover, so teams can optimize for latency, cost, compliance, and availability without changing application code. This enables organizations to build agents using their own managed endpoints while maintaining enterprise-grade governance, security, and compliance.- Flexibility & Choice: Enterprises often have specialized models tuned for their domain. "Bring your own models" ensures these models can coexist with Foundry’s 11,000+ available models, giving developers ultimate choice.
- Preserve Investments: Many organizations have invested heavily in custom models. "Bring your own models" lets them leverage those investments without rebuilding everything for a new platform.
- Unified Governance & Security: Even when using external models, the gateway enforces Foundry’s security, identity, and policy controls (via Control Plane), so compliance and observability remain consistent.
- Accelerates Innovation: Developers can combine internal models with Foundry’s hosted agents, multi-agent workflows, and tools—unlocking richer, more tailored AI solutions without sacrificing speed or trust.
- Future-Proofing: "Bring your own models" supports hybrid and multi-cloud strategies, making it easier to adapt as model ecosystems evolve.
Enterprise MCP enhancements in Foundry agents
Microsoft Foundry has enhanced integration to Enterprise MCP to enable secure, authenticated connections to MCP servers, allowing credentials to be stored and passed for public environments. These updates enable enhanced flexibility, security, and compliance when integrating MCP with Microsoft Foundry.Built-in memory in Foundry Agent Service
The new long-term memory in Foundry Agent Service gives developers a robust, scalable foundation for agents that can remember, adapt, and act coherently across sessions and workflows. Integrated directly into the Foundry runtime, memory provides a persistent layer where agents can store and retrieve chat summaries, user preferences, and key task outcomes. It enables more natural, consistent, and personalized agent behavior.Multi-Agent Workflows
Multi-agent workflows are an orchestration capability in Foundry Agent Service that lets developers design and run complex, multi-step processes involving multiple specialized agents. Instead of building a single monolithic agent, you can coordinate a network of agents—each focused on a specific task—into a structured workflow.- Fine-grained control: Design workflows visually or in code, with robust state management and human-in-the-loop controls.
- Enterprise extensibility: Add specialized agents (e.g., payment, escalation) and maintain clean separation of concerns.
- Long-term reliability: Persistent state and error recovery make workflows robust for mission-critical scenarios.
Agent mitigations and guardrail customization
At Ignite, Microsoft introduced guardrails for agents in Microsoft Foundry Control Plane. Formerly called content filters, guardrails can now be applied at the agent level rather than only to model deployments. They include existing controls to help mitigate prompt injection attacks, harms and risks, and to support groundedness. Our newest risk that can be detected, task adherence, helps keep agents on task, detecting when an agent’s next proposed action starts to drift off task. Microsoft also introduced a new control model that gives customers more flexibility in how they configure their guardrails and steer agent behavior. Customers can choose the risk they want to monitor, pick where the system should intervene, and decide what action to take. Intervention points include prompts and outputs, and for Ignite now also includes tool calls and tool responses. Available actions include annotation or blocking.Models
Anthropic Models now available on Microsoft Foundry
Microsoft Foundry now offers Anthropic’s Claude models (Haiku 4.5, Sonnet 4.5, Opus 4.1) alongside GPT models, making Azure the only cloud platform with access to both frontier model families in one place.- Model Strengths & Use Cases
- Claude Haiku 4.5: Fast, cost-efficient for real-time experiences and lightweight agents.
- Claude Sonnet 4.5: Best for complex agents, coding, cybersecurity, and research.
- Claude Opus 4.1: Advanced reasoning for long-horizon tasks, AI agents, and specialized workflows.
- Flexibility & Choice: Access both Claude and GPT models in one platform for diverse use cases.
- Enterprise Integration: Built-in governance, observability, and compliance for production-ready AI.
- Agentic Workflows: Move beyond prototypes to scalable, automated workflows with Claude as the reasoning core.
- Reusable Skills: Standardize and scale automation across projects using modular Skills.
- Advanced Capabilities: Deep Research and MCP integration enable richer, context-aware AI systems.
Sora 2 API
OpenAI’s Sora 2 is now available in public preview within Microsoft Foundry. Microsoft Foundry offers a unified environment with multiple generative models (e.g., Sora, GPT-image-1, Flux 1.1), enabling developers to access diverse capabilities in one secure platform.- Generative media: Create rich, interactive video and audio experiences with advanced simulation and dialogue.
- Unified workflow: One API for both video and audio—streamline development and integration.
- Enterprise-grade performance: High fidelity, scalability, and compliance for production scenarios.
- Safety and Governance: Built-in content filters for inputs and outputs, plus robust compliance and privacy controls to ensure responsible AI use.
Model Router GA
Model router in Microsoft Foundry is an AI orchestration layer that automatically selects the best LLM for each prompt in real time. Instead of developers manually choosing from thousands of models, model router evaluates factors like:- Prompt complexity
- Performance needs
- Cost efficiency
- Simplifies Model Selection: No need to manually pick from thousands of models—Model Router handles it.
- Optimizes Cost & Performance: Uses smaller models when possible, saving compute costs while maintaining quality.
- Scales Easily: One deployment covers multiple underlying models, reducing operational overhead.
- Future-Proof: Auto-updates to include new models and capabilities without extra work.
- Supports Multi-Agent Architectures: Can route prompts for different agents intelligently.
Fine-Tuning updates
Microsoft Foundry Fine-Tuning is updated to reflect a complete redesign of Foundry’s UI, delivering an agent-first experience for developers and data scientists. The new interface streamlines model creation, evaluation, and deployment, integrates with VS Code, and modernizes navigation. With unified workflows and enhanced usability, enterprises can accelerate AI development, reduce friction, and future-proof their agent ecosystem. Public preview launches at Ignite, with full migration planned after July 2025. As a part of this update, Microsoft Foundry now supports Reinforcement Fine-Tuning (RFT) for GPT-5, enabling developers to adapt large language models for specialized, context-aware applications. This update empowers teams to optimize model performance for unique business scenarios, driving innovation and improved outcomes. Foundry also now supports fine-tuning for non-OpenAI models like Mistral using the same APIs, UI, and workflows as GPT-4.1. Developers can fine-tune, deploy, and manage models with full parity across training and deployment, benefiting from quota management and developer tier hosting. This unified experience lowers costs, accelerates iteration, and simplifies production deployment for custom AI solutions.Tools
Foundry Tools
Foundry Tools (rebranded from Azure AI Services) provide a unified suite of prebuilt, production-ready AI capabilities for audio, video, images, documents, and text. Seamlessly integrated into the Microsoft Foundry platform, these tools empower developers to rapidly build intelligent agents and applications with plug-and-play functionality, robust governance, and enterprise-grade security. The rebranding simplifies user experience and accelerates onboarding for agentic AI development. We released major innovations across this Azure tools portfolio:- Azure Content Understanding is now generally available with a new bring-your-own model option starting with GPT family of models, support for VNETs, managed identities, CMKs, expanded regions, and simplified pricing.
- Live Interpreter is now generally available, delivering real-time multilingual interpretation for inclusive global collaboration. This is the same underlying technology that powers the Interpreter agent in Microsoft Teams.
- LLM Speech brings advanced speech-to-text transcription and translation with prompting and diarization, available in public preview.
- Photo Avatar, powered by VASA-1, enables lifelike avatars from a single photo, also available in preview.
- Faster Time to Production: Prebuilt AI skills (vision, speech, language, etc.) and unified APIs reduce complexity and accelerate development.
- Flexibility & Extensibility: Bring your own MCP servers or register custom tools to tailor solutions for unique business needs.
- Real Business Impact: Agents can act on real-time data from 1,400+ business systems (SAP, Salesforce, Dynamics) via Logic Apps connectors.
- Secure & Governed: Built-in observability, identity, and cost controls give confidence to deploy at scale.
- Future-Proof: Designed for agentic workflows and multimodal AI scenarios, making it easier to innovate as AI evolves.
Foundry IQ by Azure AI Search
Foundry IQ, powered by Azure AI Search, is a knowledge system that delivers a smarter way to ground agents in enterprise data. Agents only need to connect to a single knowledge base to access multiple sources, removing the need to manage separate APIs and function calls. Knowledge bases perform knowledge retrieval over supported sources including Microsoft SharePoint, Fabric OneLake, and web. Unlike traditional RAG and vector databases, knowledge bases use an agentic retrieval engine holistically across sources and conversations. It also has broader access to data, because it can connect to data remotely or in a synced index. This enables better response quality out of the box, where single-shot retrieval can fall short. Foundry IQ is enterprise-ready, with Purview governance and Entra-based access control, so developers can build with trust and confidence. Create and manage knowledge bases from the Foundry portal, or directly in Azure AI Search. [embed]https://www.youtube.com/watch?v=bHL1jbWjJUc[/embed]Platform
Foundry Control Plane
[embed]https://www.youtube.com/watch?v=_ge7U48JPcI[/embed] Foundry introduces a unified platform for enterprise AI agent observability, security, and governance. The Control Plane integrates identity, monitoring, and compliance, providing centralized control, deep observability, and compliance for agent fleets. Entra Agent ID anchors unified discovery and governance, ensuring agents adhere to enterprise security and Zero Trust standards.- Fleetwide visibility: Monitor health, cost, and performance across all agents in real time.
- Unified governance: Enforce policies, guardrails, and compliance from a single dashboard.
- Lifecycle management: Pause, update, or retire agents with one click—no manual ops required.
Observability in Foundry Control Plane
Observability in Foundry Control Plane is a comprehensive suite of capabilities that enable developers to evaluate, monitor, and optimize the quality, performance, and safety of multi-agent AI systems. These capabilities are integrated throughout the agent development lifecycle—spanning both build and operate experiences—empowering teams to deliver production-grade AI with confidence and speed. This feature empowers teams to build and manage reliable agents with integrated evaluations, synthetic datasets, and visual cluster analysis for rapid improvement. It provides actionable insights through customizable dashboards, end-to-end OpenTelemetry based tracing, and model upgrade recommendations to optimize cost and performance. Unified observability across agents offers a single view of cost, performance, and security for agents built on Foundry or third-party platforms. Additionally, AI Red Teaming enables continuous safety testing with automated adversarial simulations to uncover vulnerabilities such as data leaks and jailbreaks, ensuring robust and secure AI deployments. Observability capabilities—including tracing, a wide range of quality, risk, and safety evaluators, the ability to run evaluations in the cloud, and the AI Red Teaming Agent—are now generally available.New, granular controls for network and integration security in Microsoft Foundry
Microsoft Foundry streamlines AI development and management for the enterprise. At Ignite, we’re introducing more granular controls for network and integration security and resource management that empower IT admins to deploy AI solutions while meeting enterprise-grade security, compliance, and governance requirements. These capabilities are in public preview:- Configure Managed VNET for Foundry resources: Configure network isolation, private endpoint routing, and fine-grained outbound rules using RBAC at the resource level. Through built-in defaults and automatic provisioning, managed VNET reduces operational overhead and speeds up deployment times while enabling granular control to meet enterprise security requirements. This coming soon to Public Preview.
- Manage access to models and tools with AI Gateway in Foundry: Manage and govern AI model and MCP access with Azure APIM AI gateway, now available directly in Foundry portal. This setup offers granular controls for authentication, token consumption, rate limiting, cost tracking, and more. This feature is coming soon.
- BYO Storage for Speech and Language in Foundry: Link your own Azure Storage account to Speech and Language services in Foundry Tools, for complete control over where and how data is stored, including access policies and retention. This feature is available now.
Streamline IT governance, security, and cost management experiences with Microsoft Foundry
Microsoft Foundry streamlines AI development and management for the enterprise. These features are now generally available, empowering IT admins to deploy AI solutions while meeting enterprise-grade security, compliance, and governance requirements:- Manage Foundry costs with Azure Cost Management: Attribute and track costs at the Foundry resource and model deployment level with tags, set budgets with alerts and auto-pause, and provide developers with visibility into costs at the deployment, agent, and run level for AI apps using Azure Cost Management.
- Manage Foundry resources with a unified dashboard in Azure Portal: View and manage all resources, projects, connections, users, quota, and API Management integrations for Foundry from a unified dashboard within Azure Portal.
- NSP Support for Foundry and Azure OpenAI: Isolate resources within secure, private boundaries using VNETs and private endpoints. This feature ensures data never traverses public networks, simplifies compliance, and strengthens protection against data exfiltration, making it ideal for regulated industries and sensitive workloads.
- BYO Key Vault connection for Foundry resources: Securely store and manage secrets, credentials, and encryption keys using your own vaults to ensure compliance. See: Set up an Azure Key Vault Connection.
Developer Training tier: low-cost training SKU for FT (idle capacity)
Microsoft Foundry introduces Developer Training tier, offering ultra-low-cost model training by leveraging spot capacity. It’s designed to match the affordability of Developer Tier for hosting but applied to the training phase of fine-tuning workflows. This tier is ideal for developers testing multiple model variants, MVPs, and teams with low-throughput workloads, enabling rapid experimentation and affordable iteration before scaling production.One-click deploy channels: Teams, Microsoft 365 and non-Microsoft
Microsoft Foundry now enables one-click, no-code publishing of custom agents to Microsoft Teams and Microsoft 365 Copilot. This feature removes the complexity of pro-code deployment, allowing organizations to instantly reach millions of users while maintaining compliance, security, and governance. The streamlined process accelerates adoption and unlocks new enterprise scenarios for AI agents.New AI Templates in Microsoft Foundry
Microsoft Foundry introduces new ready-to-use, customizable AI templates for popular use cases such as live voice agents, release management, data unification, and SharePoint integration applications. These templates eliminate repetitive setup tasks, reduce deployment friction, and accelerate the journey from concept to production. With customizable code samples, pre-integrated Azure services and GitHub-hosted quick-start guides, development teams can launch faster, customize easily, and scale securely on Microsoft’s trusted AI infrastructure.Local to edge
Foundry Local Updates
Foundry Local now extends its capabilities, supporting the Whisper model and advanced on-device AI for smartphones, tablets, and IoT devices. Whisper model integration delivers robust, privacy-preserving speech and audio processing. The redesigned Foundry Local SDK let’s developers quickly create and launch AI-driven applications. It features dynamic hardware accelerator compatibility and built-in chat and audio interfaces, all operating independently without needing outside dependencies. Foundry Local Android Support is now in private preview, extending its capabilities to Android, in addition to Windows and Mac supporting advanced on-device AI for smartphones, tablets, and IoT devices, while Whisper model integration delivers robust, privacy-preserving speech and audio processing. The redesigned Foundry Local SDK lets developers quickly create and launch AI-driven Android applications. It features dynamic hardware accelerator compatibility and built-in chat and audio interfaces, all operating independently without needing outside dependencies.Join the community
It's been a busy few months—Azure AI Foundry is now Microsoft Foundry; we have a refreshed portal experience and shiny new icon; but Foundry still helps builders everywhere turn vision into reality with a modular, interoperable, and secure agent stack. We hope you enjoyed catching some of our sessions at Microsoft Ignite 2025. And if you ship something with these updates and tell us about it! Join the discussion on Discord, open a thread in GitHub Discussions forum, or catch the next live deep‑dive (Model Mondays + Agent AMA). We'll be back soon with more updates. Until then, happy coding!Post Updated on December 19, 2025 at 04:55AM
Thanks for reading
from devamazonaws.blogspot.com
Comments
Post a Comment