What is Microsoft Foundry and why it was built

Microsoft Foundry reuses the same vocabulary across the portal, SDK, and docs. Learn these terms first, because later concepts in this course assume you know them. Older material may still say Azure AI Studio or Azure AI Foundry; Microsoft documents those names as earlier brands for the same platform, now called Microsoft Foundry.

Microsoft Foundry is Azure's unified place to build and run AI applications that combine models, agents, and tools. In plain terms, it is one management home for those pieces, with tracing, monitoring, and evaluations built in. You set who has access, how networking works, and which policies apply in one place, instead of opening a separate admin screen for every AI service.

Day to day, the Foundry resource sets the rules: which models are deployed, how networking works, and who has permission. Foundry projects are where builders work: agents, evaluation runs, uploaded files, and other work scoped to that project live there. Connected Azure services such as Storage, Key Vault, and Azure AI Search remain separate resources with their own rules; Foundry links to them through connections, and you manage their policies alongside the Foundry resource.

The Microsoft Foundry portal is the main website for exploring the model catalog, using playgrounds, building prompt agents, running evaluations, and administering resources. You can also use SDKs, the Azure Developer CLI, and Visual Studio Code extensions, but the portal is where most beginners find their way around before they write code.

A production AI feature rarely needs only a model API. A support agent for an online shop needs instructions that stay consistent across turns, a tool that looks up orders, a way to test answers before customers see them, logs when tool calls fail, and access control so only the right people can change production settings. Without a platform, a team wires together model hosting, secret storage, search, monitoring, and identity by hand.

• Model access and deployments, including quotas and region availability.
• Agent hosting, conversation state, and tool execution through Foundry Agent Service.
• Knowledge retrieval through file search, vector stores, or Azure AI Search indexes.
• Observability through tracing, metrics, and evaluations tied to the same project.
• Governance through role-based access, guardrails, and (at scale) Foundry Control Plane fleet views.

Foundry uses layers on purpose. IT and security teams apply centralized controls on the Foundry resource, while product teams experiment inside projects without setting up shared model capacity again each time. Microsoft recommends a Foundry resource with one project as the default starting point for a single developer exploring the platform.

The agent development lifecycle in Foundry runs from creation through monitoring in production: choose an agent type, create and test in the Agents playground or in code, add tools, save versions, trace runs, evaluate quality, publish a stable endpoint, then monitor and improve. You do not need every step on day one, but the order explains why the portal groups models, agents, and evaluations together.

• Sign in to the Microsoft Foundry portal. For instant access models during preview, turn on the New Foundry toggle and work in Foundry (new), not the classic portal.
• Create or open a Foundry resource and a Foundry project. Concept 2 in this course walks through the screens; for a guided setup, use the quickstart linked under Go deeper.
• Explore models in the Foundry Models catalog or a model playground. When you deploy a model in the portal, you land on that model's playground to tune prompts and compare outputs.
• Choose how to call models: create a model deployment with a deployment name, or during preview use instant access models in a West US 3 project to call supported model names such as gpt-5-mini without creating a deployment first.
• Copy the project endpoint from the welcome screen or project details when you move to the SDK or code; sign in with Microsoft Entra instead of embedding long-lived keys in source code.

Imagine a small e-commerce team wants a chat assistant that answers "Where is my order?" using real order data, stays polite and on-topic, and escalates refunds to humans. They are not ready to host their own containers, so they start with a prompt agent in Foundry Agent Service.

• Pick a model from the catalog (or an instant access model name during preview in West US 3) and set instructions that describe tone, what the agent may promise, and when to hand off to staff.
• Attach tools: file search or Azure AI Search for policy documents, plus a function or API tool that looks up order status from the shop's backend.
• Test in the Agents playground, save a version when behavior looks right, and run an evaluation dataset before exposing the agent to customers.
• Use tracing to confirm the agent calls the order tool with the right arguments, then publish when metrics look stable.

The same Foundry project holds the agent definition, evaluation runs, and traces, while the Foundry resource supplies shared model deployments and security settings. That split is the practical idea behind one management home for Foundry: builders improve work inside the project; administrators set rules at the resource level.

The Foundry platform itself is free to use and explore. Each product you consume inside Foundry, such as model inference, Agent Service usage, or Azure AI Search, bills at that product's normal rate. Microsoft's pricing FAQ states that you do not need an Azure account to explore Foundry, but you need an Azure subscription before you start building agents.

• Estimate spend with the Microsoft Foundry pricing pages and the Azure pricing calculator before you run large evaluations or high-traffic deployments.
• Instant access models (preview) let you call supported models by model name with no deployment. The same SDKs and clients work; you pass a name such as gpt-5-mini instead of a deployment name.
• During preview, instant access is available only for Foundry projects in West US 3. Deployments remain the right choice when you need reserved throughput, custom content filters, data residency, or other enterprise deployment settings.
• Preview features are offered under supplemental terms and are not recommended for production workloads until you confirm support for your region and scenario.

The catalog grows frequently, so treat the exact count as a moving target. The practical takeaway is that you compare models in the catalog and benchmarks before you commit to a deployment, and you can route requests with features such as model router when you need cost and quality trade-offs at scale (covered in later concepts).

• Assuming free to explore means free inference. Model tokens, agent runtime, and connected services still bill when you use them.
• Creating agents in the playground but never saving versions. Unsaved changes are lost if you leave the Microsoft Foundry portal; saved versions are required for full evaluation and monitoring workflows.
• Calling a model by its catalog name in production code while only a deployment exists, or the reverse during instant access preview in the wrong region.
• Treating Storage, Key Vault, or Azure AI Search as part of the Foundry resource. They are connected resources with separate networking and access rules.
• Giving everyone broad admin rights. Docs recommend the Foundry User role for developers on the resource as a common starting point; creating a deployment is an admin-style action, while running an agent is day-to-day use, and those permissions can be assigned separately.

Concept 2 in this course shows where you work inside the Microsoft Foundry portal: projects, playgrounds, Manage versus Operate, and how that maps to the resource hierarchy described here.

• Official overview of the platform, evolution from Azure AI Studio, and recommended paths for new developers: What is Microsoft Foundry?
• More detail on how resources, projects, roles, deployments, and connected services fit together: Microsoft Foundry architecture
• Step-by-step setup of a resource, project, and first model deployment: Create Foundry resources
• FAQ on free exploration, subscriptions, and how each service bills: Microsoft Foundry pricing