Update (October 2026): This guide was rewritten for the current Microsoft Foundry. The April 2026 version described hubs, threads and runs,
azure-ai-inferenceand models such as GPT-4 and Llama 3, all of which Microsoft has since replaced or retired. Every statement below was checked against Microsoft Learn on 20 October 2026.
Microsoft Foundry, formerly Azure AI Foundry and before that Azure AI Studio, is Microsoft's platform for building AI applications and agents on Azure. In plain terms it is one place, and one Azure resource, where you choose a model, give it instructions and tools, check its answers are good and safe, and ship it. Microsoft describes it as unifying "agents, models, and tools under a single management grouping". If you are preparing for AI-103, it is the platform the whole exam is built around.
Microsoft Foundry in plain words
Imagine a workshop for building AI assistants.
- Models are the brains you can hire: more than 10,000 of them, from OpenAI, Microsoft, Anthropic, Meta, DeepSeek, Mistral and others.
- Agents are the assistants you build: a model plus instructions plus the tools it is allowed to use.
- Tools are what an agent can reach: your documents, a search index, the web, your own APIs, and ready-made services for speech, language, documents and safety (Foundry Tools, explained in our Foundry Tools guide).
- Observability is quality control: tests that score the answers, and traces that show what the agent did step by step.
- Guardrails are the safety rules: what gets blocked on the way in and on the way out.
Everything lives under one Azure resource, so access control, networking and policies are set once.
How Foundry is organised: one resource, many projects
This is the part that changed most, and the part older tutorials get wrong.
The Foundry resource
A Foundry resource is a single Azure resource of kind AIServices. It is the governance boundary: networking, security, billing and model deployments sit here. Storage, Key Vault and Azure AI Search are not built in. You attach them as connected resources when you need them, and Microsoft-managed storage and keys are the default.
Projects
A project is a child of the Foundry resource and the place a developer works. Each project keeps its own agents, evaluations, files and connections, and its own access control. Most teams use one resource with a project per application or team.
What happened to hubs
Originally, Foundry was a hub with child projects, plus separate Azure OpenAI and Azure AI Services resources. Hubs still exist, but only in Foundry (classic), and Microsoft documents them as the legacy setup. New work goes into a Foundry resource with projects. If you have an Azure OpenAI resource, you can upgrade it in place to a Foundry resource and keep its endpoint, keys and deployments.
For AI-103, this is what "Design Azure infrastructure for AI apps and agent-based solutions" and "Configure an application to connect to a Foundry project" refer to today: a resource, its projects, and the connections and roles between them.
Models: choosing and deploying
The catalog
The Foundry Models catalog has more than 10,000 models. OpenAI's GPT-5 family sits next to Microsoft's Phi models, Anthropic, Meta's Llama 3.3 and Llama 4, DeepSeek, Mistral, Cohere, xAI and many open models from Hugging Face. Models retire on a schedule, so check the retirement page before you build on one: GPT-4 and Llama 3 are already gone, and gpt-4.1 retires on 14 April 2027.
Deployment options and types
There are two ways to deploy a model:
- Serverless API, which Microsoft calls "the preferred and most capable deployment path". It works for Azure OpenAI and partner models alike. Deployments live on the Foundry resource and every project can use them.
- Managed compute (preview), where a model runs on dedicated virtual machines in your project.
Serverless deployments come in types that decide where data is processed and how you pay:
- Global Standard, the recommended starting point: pay per token, processed anywhere in Azure's global network.
- Data Zone Standard: the same, but processing stays inside a data zone such as the EU or the US.
- Standard: a single region.
- Provisioned (Global, Data Zone or Regional): reserved throughput for steady production load.
- Batch (Global or Data Zone): cheaper, asynchronous jobs.
- Developer: for testing fine-tuned models.
Instant access (preview) even lets you call some models by name with no deployment at all.
Try a real Azure AI lab, free
A pre-provisioned Azure environment in your browser. No Azure account, no credit card, no setup.
Most large enterprises already run on Microsoft infrastructure. Azure AI Foundry is the platform those teams use to build and deploy AI apps, so the skills you practice here map directly to real job demand.
Agents: what Foundry Agent Service builds
An agent is a model plus instructions plus tools that can reason over a request and take action. Foundry offers three kinds, from least to most code:
- Prompt agents: you choose a model, write instructions and attach tools, in the portal or the SDK. Foundry hosts and runs it for you.
- Voice-based prompt agents: the same, but people speak to it in real time, through Voice Live.
- Hosted agents: you bring your own code, for example built with Microsoft Agent Framework or LangGraph, package it as a container, and Foundry runs it with a managed endpoint, scaling, identity and observability.
The new agent API
Agents now run on the Responses API (Agents v2). The old Assistants API, with its Threads, Messages and Runs, was retired on 26 August 2026. The current names are Conversations, Items, Responses and Agent Versions, and an agent is created as a version, so you can roll forward and back. Responses are synchronous by default, so the old polling loop is gone.
Tools
An agent's tools include file search, code interpreter, Azure AI Search, web search and Grounding with Bing, function calling, OpenAPI tools, MCP servers, A2A connections to other agents, browser automation, computer use, image generation, SharePoint and Azure Functions. Microsoft's recommended way to give an agent tools is a Foundry Toolbox: a curated, versioned set of tools exposed through one MCP-compatible endpoint, with credentials and guardrails managed centrally.
Knowledge with Foundry IQ
Foundry IQ is the managed knowledge layer for agents: knowledge bases built from your sources on top of Azure AI Search, with retrieval that respects who is allowed to see what. It is how an agent answers from company documents instead of guessing.
Multi-agent work
For several agents working together, Foundry has workflows (preview) in the portal, and in code Microsoft Agent Framework, which Microsoft describes as the direct successor to both Semantic Kernel and AutoGen. Agents can be published to stable endpoints, Microsoft 365 Copilot and Teams, and each agent gets its own Microsoft Entra identity.
Quality and safety
Observability: evaluation, monitoring and tracing
Foundry treats quality as three linked jobs:
- Evaluation scores answers with built-in evaluators for quality (coherence, fluency, relevance, groundedness), retrieval, safety and agent behaviour, before release and continuously after it.
- Monitoring dashboards track usage, latency and token consumption in production.
- Tracing records every step an agent took, using OpenTelemetry, so you can see why it answered as it did.
Microsoft's improvement loop is to evaluate, look at clusters of failures, adjust the agent (an agent optimizer, in preview, can suggest better instructions, tool descriptions and models), and evaluate again.
Guardrails
What used to be called content filters is now configured as guardrails: named sets of controls applied at four points, the user's input, a tool call, a tool's response and the model's output. They cover harm categories, Prompt Shields against jailbreaks and indirect prompt injection, protected material and personal data. Groundedness detection exists but is in preview and applies to models, not agents. The Foundry control plane (preview) lets an organisation apply guardrail policies across all its agents, with Azure Policy, Defender and Purview.
Access
Roles were renamed with the platform: Foundry User, Foundry Project Manager, Foundry Owner and Foundry Account Owner. Creating resources and deployments is a control-plane action; calling models and agents is a data-plane action, and the two need different roles.
Still reading? Grab the AI-103 exam study guide
One-page reference of every Foundry topic the AI-103 exam tests.
How developers connect
- One project endpoint:
https://<resource>.services.ai.azure.com/api/projects/<project>. - The Foundry SDK:
azure-ai-projectsversion 2.x for the current Foundry (1.x targets classic, and mixing them causes errors), with clients for Python, C#, JavaScript and Java. - The OpenAI SDK: the standard
openaipackage against the project's/openai/v1/routes, with no monthlyapi-versionparameter. The oldazure-ai-inferencepackage was retired on 26 August 2026. - Other surfaces: the Foundry portal, the Azure Developer CLI for hosted agents, a Visual Studio Code extension, and a Foundry skill and MCP server for coding agents.
- Foundry Tools (Speech, Language, Translator, Document Intelligence, Content Understanding, Vision, Content Safety and the rest) keep their own SDKs and are called on the resource endpoint. See the Foundry Tools guide.
Foundry vs Azure OpenAI
Azure OpenAI is one family of models inside Foundry, not a separate layer underneath it. A Foundry resource is a superset of an Azure OpenAI resource: it serves OpenAI models and every other model in the catalog, plus agents, tools, evaluations and guardrails. Agent Service exists only in Foundry. If you already have an Azure OpenAI resource, you can upgrade it in place and keep the same endpoint and keys. The naming history is in Microsoft Foundry vs Azure AI Studio vs Azure OpenAI.
How AI-103 uses Foundry
The AI-103 exam, "Developing AI Apps and Agents on Azure", leads to Microsoft Certified: Azure AI Apps and Agents Developer Associate. Microsoft's skills outline has five domains, and Foundry runs through all of them:
| Domain | Weight | Where Foundry comes in |
|---|---|---|
| Plan and manage an Azure AI solution | 25 to 30% | Foundry resource and project design, deployment options, keyless credentials, private networking, roles, guardrails, monitoring |
| Implement generative AI and agentic solutions | 30 to 35% | Models, prompt and hosted agents, tools, Foundry IQ, multi-agent workflows, evaluation |
| Implement computer vision solutions | 10 to 15% | Image and video generation, multimodal understanding, Content Understanding |
| Implement text analysis solutions | 10 to 15% | Language analysis with models and Foundry Tools, Translator, Speech |
| Implement information extraction solutions | 10 to 15% | Retrieval pipelines with Azure AI Search, and Content Understanding analyzers |
The exam mostly tests generally available features, and preview features can appear when they are widely used. The AI-103 study guide goes domain by domain.
Getting hands-on
Reading the portal is not the same as using it. The AI-103 labs walk through Foundry in the same order as this guide: exploring the portal and deploying a model, choosing between models with evidence, building and grounding an agent, connecting MCP, Foundry IQ, evaluations, multi-agent workflows and guardrails, in real Azure with nothing to set up. The first lab is free. See the AI-103 labs.
Frequently asked questions
Is Azure AI Foundry the same as Microsoft Foundry?
Yes. Azure AI Studio became Azure AI Foundry, which became Microsoft Foundry. The portal is at ai.azure.com, and the older hub-based experience is now called Foundry (classic).
Do I still need a hub to use Foundry?
No. New work uses a single Foundry resource with projects. Hubs and hub-based projects remain only in Foundry (classic) for existing setups.
What replaced threads and runs in Foundry agents?
The Responses API. Since the Assistants API retired on 26 August 2026, agents use Conversations, Items, Responses and Agent Versions instead of Threads, Messages, Runs and Assistants.
Which SDK should I use with Microsoft Foundry?
The Foundry SDK, azure-ai-projects 2.x, together with the standard openai package against your project endpoint. The older azure-ai-inference package was retired in August 2026.
What are Foundry Tools?
The new name for Azure AI Services: ready-made APIs such as Speech, Language, Translator, Document Intelligence and Content Understanding. Our Foundry Tools guide explains each one.