# Azure AI Foundry Agent Service Is Now GA — Here's What That Means for Your Multi-Agent Strategy
Microsoft Build 2025 just wrapped, and if there was one theme that echoed through every keynote, breakout, and hallway conversation, it was this: agents aren't just the future — they're the present. And the announcement that caught my attention most? Azure AI Foundry Agent Service hit general availability.
This isn't just another "now available" badge on a dashboard. It's a signal that Microsoft is all-in on agentic AI as an enterprise-grade, production-ready paradigm. Let me walk through what this means, what's new, and why it matters if you're building — or thinking about building — AI agents in the Microsoft ecosystem.
The Journey from Preview to GA
First announced as a public preview at Ignite 2024, the Azure AI Foundry Agent Service gave developers a unified platform for building, deploying, and scaling intelligent agents. In the months since, over 10,000 customers put it through its paces. Real-world deployments emerged fast. NTT DATA used it to cut customer service response times dramatically, achieving 50% faster time to market. YoungWilliams built "Priya," an AI assistant that automates complex government program inquiries — handling everything from LIHEAP to SNAP eligibility — with enterprise-grade security and compliance baked in.
Those weren't toy projects. They were production workloads that helped Microsoft harden the platform. And now, GA brings a suite of capabilities that make multi-agent systems genuinely approachable.
The Two Flavors of Agent Collaboration
The headline feature in this GA release is multi-agent orchestration, and it comes in two distinct patterns that serve different needs:
1. Connected Agents (Preview)
Think of this as peer-to-peer agent interaction. One agent can call another as if it were a tool — delegating a specialized subtask and integrating the result. This is perfect for scenarios where you have discrete, modular expertise. Your customer-facing agent needs to verify compliance? It calls a compliance agent. Needs to pull from a knowledge base it doesn't own? It calls the knowledge-management agent.
The beauty of the "agent-as-tool" pattern is that it preserves clean separation of concerns. Each agent owns its domain. You iterate on them independently. You test them in isolation. They compose at runtime.
2. Multi-Agent Workflows (Preview)
If Connected Agents are the conversations, Multi-Agent Workflows are the orchestras. These provide structured, stateful orchestration across long-running, multi-step processes. We're talking about scenarios like customer onboarding — where an identity-verification agent hands off to a provisioning agent, which hands off to a notification agent, with context flowing through every step.
The workflow layer handles the hard stuff: state management, error recovery, and durability. If one step fails, the workflow doesn't just die — it retries, compensates, or escalates based on rules you define. This is the kind of reliability engineering that enterprise workloads demand but that few teams want to build from scratch.
The Semantic Kernel + AutoGen Convergence
Behind the scenes, something smart is happening. Microsoft is converging the runtimes for Semantic Kernel and AutoGen into a single client library that exposes one API surface for both single-agent and multi-agent patterns.
Why does this matter? Semantic Kernel brings production-grade architecture — modular, testable, enterprise-SDK sensibilities. AutoGen brings dynamic orchestration patterns — the kind of fluid, emergent agent collaboration researchers love. The convergence means you get both. Build locally, simulate agent interactions, then deploy to Azure AI Foundry with consistent behavior across environments. No "it worked in dev" surprises.
Agents Don't Live in a Vacuum
GA also brings a raft of integration announcements that make agents practical in the real enterprise:
- •**1,400+ Logic Apps connectors** are now available as agent tools. That means your agent can reach into SAP, Salesforce, ServiceNow, or any of the hundreds of systems Logic Apps talks to — without writing custom integration code.
- •**SharePoint joins the knowledge tool lineup** alongside Microsoft Fabric and Bing Search. If your enterprise knowledge lives in SharePoint document libraries, your agents can now access it natively as grounded context. This is a big deal for organizations that have invested years building their SharePoint knowledge bases.
- •**Agent2Agent (A2A) protocol support** enables cross-cloud orchestration. Your Foundry agent can talk to agents running on SAP Joule or Google Vertex AI through open, standard protocols. Microsoft is betting on an open agentic web — not a walled garden.
- •**Agent Catalog** provides a growing library of reusable agent samples from Microsoft and partners like Auquan, Saifr, and Sight Machine. Don't start from scratch. Clone, customize, and deploy.
What This Means for Developers
Here's my take on the practical implications:
1. Start thinking in agent teams, not just individual agents. The hardest problems in enterprise automation aren't single-task — they're multi-step, cross-system, exception-heavy processes. One agent can't do it all. But a team of specialized agents, coordinated by a durable workflow, can.
2. The tooling is catching up to the ambition. A year ago, building a multi-agent system meant stitching together SDKs, writing your own state management, and hoping for the best. Today, Azure AI Foundry gives you the orchestration layer, the developer experience (including VS Code integration), and the deployment pipeline. The barrier to entry just dropped significantly.
3. The ecosystem is opening up. Support for open protocols like MCP and A2A, partner tools, cross-cloud connectivity — this isn't a lock-in play. Microsoft is building a platform that connects to everything, and that's the right strategy for an agentic future where no single vendor owns all the agents.
4. Windows is becoming a first-class AI development platform. With Windows ML entering public preview, the convergence of ONNX Runtime, DirectML, and hardware execution providers from AMD, Intel, NVIDIA, and Qualcomm means your dev machine can be your inference machine. Build and test agents locally before you deploy to the cloud.
Getting Started
If you want to kick the tires, here's your path:
1. Head to the [Azure AI Foundry portal](https://ai.azure.com) and create an Agent Service project. 2. Explore the Agent Catalog for a sample that's close to your use case. 3. Use the VS Code extension to author your agent locally with the converged Semantic Kernel / AutoGen SDK. 4. Deploy to Foundry and test with real data through the built-in knowledge tools. 5. When you're ready for multi-agent, design a workflow that chains your agents together.
The GA milestone means SLAs, support, and production readiness. This is no longer an experiment — it's infrastructure.
The Bigger Picture
We're living through a shift in how software gets built. Not just AI-assisted coding (though that's part of it), but a fundamental change in what software *is*. Agents that reason, plan, use tools, and collaborate with each other aren't a sci-fi future. They're running in production today at organizations like NTT DATA and YoungWilliams.
Microsoft's bet is that the platform that makes this easiest — with the deepest enterprise integrations, the most mature security model, and the broadest ecosystem — wins. With Azure AI Foundry Agent Service going GA, that bet looks pretty well-placed.
*What are you building with AI agents? I'd love to hear about it. Drop a comment or find me on the socials.*