The world is moving from assistants that answer to agents that act. Systems can now plan, reason, and collaborate across data, tools, and people. The shift creates enormous opportunity and raises the bar for what “production-ready” truly means: agents that operate on your network, respect your security policies, feel natural and intuitive for your end-users, and give you full visibility into what they're doing in production.
AI decision-makers cite security, privacy, and governance as the primary drivers for consolidating onto a unified platform and report difficulty scaling AI across disconnected tools. Microsoft Foundry makes it easier to build enterprise agents by bringing enterprise-grade security, private networking, and observability to every agent you build.
Today we’re announcing:
- General availability (GA) of the next-generation Foundry Agent Service, a modern API, runtime, and developer platform to help teams build agents, move from prototype to production quickly, and operate with confidence.
- Voice Live integration with Agent Service in public preview, enabling seamless real-time speech-to-speech interactions for agents in production.
- General availability of evaluations, tracing, and monitoring in the Foundry Control Plane with out-of-the-box and custom evaluators to continuously and systematically measure quality, groundedness, safety, and risk across the AI development lifecycle.
- General availability of Microsoft Foundry portal at ai.azure.com, a unified, interoperable interface to build, optimize, and govern AI apps and agents with faster iteration and production-grade controls to help teams to ship AI solutions more reliably and at scale.
General Availability: The Next-Generation Foundry Agent Service
Foundry Agent Service is now generally available, with production-ready SDKs across Python, JavaScript, Java, and .NET and a modern API for building durable, tool-using agents designed to streamline development while scaling reliably for enterprise workloads. Built on the Responses API — OpenAI's agentic API — agents are wire-compatible, so teams already using the Responses API can run on Foundry with minimal code changes, immediately gaining enterprise security, private networking, and observability. Developers can publish agents built with Foundry via REST endpoints or directly to Microsoft 365 and Microsoft Teams.
Prompt (declarative) agents are now generally available, while hosted agents and workflow agents remaining in public preview with expanded capabilities and regional coverage.
For teams deploying code-based agents using frameworks like Microsoft Agent Framework or LangGraph, hosted agents provide a path to operationalize agent logic without owning the infrastructure glue. With the regional expansion of hosted agents to East US, North Central US, Sweden Central, Southeast Asia, Japan East, and more, customers globally can meet data residency requirements and reduce latency for global deployments.
Teams building multi-agent workflows in LangGraph can now connect directly to the new Foundry Agent Service — composing complex graphs while keeping individual agents running in Foundry. Developers stay in their existing LangGraph environment and get access to tools like Foundry IQ, code interpreter, MCP, and file search, with all agents visible and governed in Foundry Control Plane and end-to-end observability across the full graph.
These capabilities are already making a difference in production. Corvus Energy, which powers battery systems across 1,500+ maritime vessels, is one of the teams building on Foundry Agent Service. They're using Foundry and Microsoft Fabric to move from manual inspection workflows to agent-driven operational intelligence across their global fleet.
"Microsoft Foundry has delivered real value for our team. The updated interface makes our services and workflows faster, more capable, and easier to use — and we're now able to connect things that simply weren't possible before."
Andreas Hunderi, VP of Global IT at Corvus Energy
Public Preview: Voice Live Integration with Foundry Agent Service
We're introducing Voice Live API integration with the new Agent Service, making voice a native, production-ready channel for agents. Voice Live API is a unified, real-time API that provides premium speech capabilities with foundation models for natural speech-to-speech interactions across voice agents, assistants, and chatbots.
With this integration, agents built with Foundry can leverage seamless, secure voice interactions for scenarios like customer service and field support. Voice Live handles all audio input and output, and inherits Foundry’s security and monitoring features for reliable, hands-free dialogue.
“Gulf Air’s Falcon Eye is a real-time operational intelligence platform that brings together live flight, passenger, crew, revenue, weather, and disruption data into a single command view. Integrated with Voice Live in Microsoft Foundry, executives—from the CEO to operations leaders—can speak naturally and receive immediate, accurate spoken answers grounded in live data. Voice-driven navigation can also open dashboards and analyses, turning airline data into a conversational mission‑control co‑pilot for faster, more confident real-time decision‑making."
Ahmed Naeemi, Chief Information Officer, Technology and Digital Services
To learn more about Voice Live and how it integrates with Foundry Agent service, read this blog.
General Availability: Observability
As organizations move agentic AI systems into production, quality and safety can no longer rely on one-time testing. They require continuous measurement and operational oversight.
Observability in Foundry Control Plane is generally available, enabling teams to systematically assess performance across the AI lifecycle:
- Out-of-the-box evaluators: Evaluate coherence, relevance, groundedness, retrieval quality, and safety using built-in evaluators across generation and retrieval augmented generation (RAG) scenarios.
- Custom evaluators (preview): Define LLM-as-a-judge and code-based custom evaluators aligned to business logic, internal standards, and regulatory requirements.
- Continuous evaluation with monitoring and alerting: Automatically assess sampled production traffic using built‑in and custom evaluators, and surface quality, safety, availability, and performance signals through Azure Monitor to support dashboards and alerts and detect regressions over time.
- Tracing: Trace agent and model execution paths end to end, capturing requests, tool calls, and responses to understand system behavior, diagnose failures, and debug issues across agents, models, and tools in production. Tracing will become generally available later this month.
Today prompt optimizer is also in preview, applying prompt-engineering best practices to restructure, clarify, and enhance your instructions. It provides transparent, paragraph-level reasoning for every change. You can iteratively refine results by adding suggestions and re-optimizing until satisfied, then apply the final output with a single click.
Together, these capabilities allow organizations to move from reactive troubleshooting to proactive, measurable AI quality management in production.
End-to-End Private Networking and MCP Security
Organizations want agents that can reason over private data and take real actions, without opening public network paths. With this release, Foundry Agent Service supports standard setup with private networking so you can bring your own private virtual network (BYO VNet) for an isolated environment. This means no public egress, container/subnet injection into your network, and access to private resources via the platform network. Private networking can extend to tool connectivity, including VNet support for Model Context Protocol (MCP), Foundry IQ by Azure AI Search, and the Fabric data agent.
Private networking secures the transport layer, but agent tools adoption also depends on authentication and least-privilege controls. Foundry supports a full spectrum of authentication methods for MCP servers, including:
- Key-based authentication
- Microsoft Entra - Agent ID
- Microsoft Entra – Foundry project managed identity
- OAuth identity passthrough
This flexibility lets teams choose shared authentication for organization-wide access, or per-user authentication to preserve individual context and enforce least-privilege access. For organizations that require tighter guardrails, teams can disable preview features entirely and standardize on GA-only capabilities for production development.
Partner Security Integrations: Palo Alto Networks and Zenity
Microsoft Foundry is built with a strong commitment to security, grounded in deep integration with Microsoft Security across identity, data protection, and threat detection. This foundation ensures customers can build and operate agents with enterprise‑grade protections and consistent governance across their AI estate.
As part of our commitment to an open ecosystem, we are expanding customer choice through partner‑native security integrations that are directly accessible in Foundry.
Palo Alto Networks' Prisma AIRS now integrates with Foundry (preview). This enables policy-driven runtime inspection of prompts and model responses to help teams detect risks such as prompt injection and toxic content, all within the governance framework of the Foundry platform.
We are also introducing a Zenity integration (preview). Zenity extends the built-in guardrails and controls in Foundry Control Plane, adding runtime security capabilities for AI agents. By analyzing agent intent and execution paths as agents interact with tools and systems, Zenity helps identify and stop unexpected actions, unsafe tool usage, or unintended data access before they impact production systems. Bringing partner capabilities directly into Control Plane gives enterprises model-agnostic security controls across the agent lifecycle, while allowing them the flexibility to use the security stack that best fits their organization.
General availability: Microsoft Foundry Portal
The new Microsoft Foundry portal — available at ai.azure.com — is now generally available, providing unified UI experience for teams to build, optimize, and govern AI apps and agents in one place. The portal brings together Foundry Agent Service, Foundry Models, Foundry Tools, Foundry IQ, and Foundry Control Plane with integrated governance, helping teams move from experimentation to production with the right controls in view.
Watch the Microsoft Mechanics episode on Microsoft Foundry to get familiar with the new portal and services.
General availability is a milestone, not the finish line.
The goal has always been simple: any model, any tool, any framework — governed, observable, and ready for production. With Foundry Agent Service, Observability in Control Plane, and the Foundry portal now generally available on a production-ready backend, our focus shifts to making that even easier — deeper workflow authoring, richer guidance, extended governance, and flexible integration into existing enterprise ecosystems, guided by customer feedback and real-world production usage.
We can’t wait to see what you build with Microsoft Foundry.
Get Started
- Ready to get hands-on? Start building your own intelligent agents today with Microsoft Foundry and download the Foundry SDK
- Learn more about Foundry Agent Service in this Dev Blog and see the demos in action in the latest Model Mondays episode
- To continue your learning journey, visit Microsoft Learn and explore resources including the AI Agents for Beginners, Microsoft Agent Framework, and course materials that help you build and operate agents responsibly
- Dive deeper with the AI Builder’s Guide to Agent Development and Management whitepaper
- Keep the conversation going in GitHub and Discord!
Microsoft