Microsoft Foundry Agent Service is Now GA: Everything You Need to Know

TL;DR

Foundry Agent Service is GA — Production-ready AI agents on Azure
OpenAI compatible — Responses API works with minimal migration
Full VNet isolation — No public egress for model or tool calls
Voice Live preview — Real-time speech-to-speech in one API
Evals are GA — Built-in monitoring that pipes to Azure Monitor
Agent Framework RC — Microsoft's SDK for building agents is stabilizing
Azure DevOps MCP — Native DevOps integration now in preview

Big news dropped today: Microsoft Foundry Agent Service is now generally available. If you've been waiting to go production with AI agents on Azure, the wait is over.

This isn't just a version bump—it's a fundamental upgrade to how enterprises can build, deploy, and govern agentic AI at scale.

Connect with us on Discord! Join the Code to Cloud community here.

🚀 What's New in This Release

Here's what actually matters for your architecture decisions.

Responses API Runtime — OpenAI Compatibility

Wire-compatible with OpenAI agents. If you're already building on the Responses API, migration is straightforward—your existing patterns and interfaces work with minimal changes. This is a big deal for teams that started on OpenAI and need enterprise controls without a rewrite.

📖 Foundry Agent Service Quickstart →

Full VNet Isolation — Zero Public Egress

No public egress. Period. This covers model calls and tool calls—MCP servers, AI Search, Fabric. Your agents can operate entirely within your network boundary. Data never leaves your perimeter. This is the feature that gets security teams to sign off.

📖 Private networking documentation →

Voice Live — Real-Time Speech (Public Preview)

Real-time speech-to-speech plugged directly into your agents. One API instead of stitching STT + LLM + TTS together. If you've built voice agents before, you know how painful that pipeline is. Voice Live handles the orchestration so you can focus on conversation design.

📖 Voice Live sample code →

Evaluations are GA — Built-In Observability

Built-in evaluators plus custom ones, with continuous monitoring on live traffic. Pipes directly into Azure Monitor so you actually know when things degrade. No more flying blind in production—you get real-time visibility into agent performance and quality.

MCP Auth — Complete Authentication Coverage

All authentication patterns in one service: key-based, agent identity, managed identity, and OAuth passthrough (agent acts as the user). No more cobbling together auth flows—Foundry handles the complexity so your MCP integrations just work.

📖 MCP integration documentation →

6 New Regions — Expanded Availability

Now available in: East US, North Central US, Sweden Central, Southeast Asia, Japan East—with more coming. This expansion addresses data residency requirements for global enterprises.

Open Model Support — Use Any Model

Run GPT, Llama, Mistral, DeepSeek—mix and match across your agent fleet. Use LangGraph, LangChain, or go direct. The runtime handles the orchestration regardless of your framework choice. Pick the right model for each task without changing your architecture.

🔧 Also Announced: Azure DevOps MCP Server & Agent Framework RC

Alongside the Foundry GA, Microsoft dropped two more announcements that complete the picture.

Azure DevOps Remote MCP Server — Public Preview

Your agents can now interact with Azure DevOps natively through MCP. The Azure DevOps Remote MCP Server has entered public preview, enabling seamless integration between your AI agents and your DevOps workflows.

What this unlocks:

Query work items — Agents can read and search issues, user stories, and tasks
Manage repositories — Access code, branches, and pull requests
Trigger pipelines — Kick off builds and deployments programmatically
Access build and release information — Get deployment status and history

Combined with the MCP auth capabilities in Foundry, this opens up serious DevOps automation scenarios—from intelligent code review agents to deployment assistants.

📖 Read the Azure DevOps MCP announcement →

Microsoft Agent Framework — Release Candidate

The SDK for building agentic AI on Foundry has reached release candidate. This is the framework Microsoft recommends for production agent development, and it's stabilizing ahead of 1.0.

What you get:

Production-ready APIs — Stable interfaces you can build on
Deep Foundry integration — Native support for Agent Service features
Multi-agent orchestration — Built-in patterns for agent-to-agent workflows
Tracing and observability — Debug and monitor your agents in production
Open model support — Works with any model Foundry supports

If you've been on Semantic Kernel, this is your migration path for agentic workloads.

📖 Read the Agent Framework RC announcement →

📖 Microsoft Agent Framework on GitHub →

💡 What This Means for Your Architecture

If you've been hesitant to move AI agents to production, this release addresses the main blockers:

Concern	Solution in GA
"Our data can't leave the network"	Full VNet isolation with no public egress
"We're locked into OpenAI's API"	Wire-compatible Responses API runtime
"We don't know when agents fail"	Built-in evals + Azure Monitor integration
"Auth is a nightmare with MCP"	Key, agent, managed identity, OAuth passthrough
"We want to use different models"	Open model support—GPT, Llama, Mistral, DeepSeek
"Voice agents are too complex"	Voice Live handles the STT → LLM → TTS pipeline

📚 Key Resources

Official Announcements

Documentation

Code & Samples

🎯 Next Steps

Review your current agent architecture — If you're on OpenAI Responses API, migration just got easier
Talk to your security team — VNet isolation removes a major blocker
Set up monitoring — Deploy evals before you scale, not after
Explore MCP integrations — Azure DevOps MCP is just the start