Microsoft Foundry

1 · What it is

Microsoft Foundry is where companies build AI apps and agents on Microsoft’s Azure cloud. An agent here means a model given instructions plus tools, so it can think through a request and then take an action. Foundry collects the pieces an agent needs: the models, the tools, the place it runs, and the checks that keep it honest and safe. It used to be called Azure AI Studio and then Azure AI Foundry.

Start with the models. Foundry’s catalogue sorts them into two groups. Models sold by Azure include every Azure OpenAI model plus selected models from other makers. Azure bills them through your Azure subscription (your paid Azure account). They come with service-level agreements, which are written promises about how well the service will run, and Microsoft supports them. Models from partners and community mostly come from outside organisations, such as partner companies and research labs. If Azure does not sell one, Microsoft’s terms call it a non-Microsoft product. That matters when you ask who helps if a model breaks.

Next come the agents. The simplest kind is a prompt agent: you write instructions, pick a model and tick the tools it may use, for example file search or searching the web. Foundry then runs it for you, so you write no server code. Teams that already have agent code, for example built with LangGraph, can package it as a container, a sealed bundle of code and settings. Foundry runs that as a hosted agent. It gets its own endpoint (a web address other programs can call), grows to handle more traffic, and has its own identity. Either way, you can swap the model without rewriting the agent.

Here is an everyday example. A bank wants a help agent that answers from its policy documents. The team picks a model, adds file search over those documents and writes instructions such as “only answer from policy”. Before a customer’s question reaches the model, a guardrail can scan it. A guardrail is a set of safety checks. The agent then searches the files and drafts an answer, and another check scans the reply before the customer sees it.

Finally, Foundry watches what happens. Tracing records each model call and tool call, so an engineer can see why a request was slow or wrong. Evaluators are automatic graders. They score answers for quality, safety and whether the agent used its tools correctly. Dashboards track token use (tokens are the small chunks of text a model reads and writes), speed, errors and quality scores over time. Builders can do all this in the Foundry portal, a website. Or they can write code with software kits for Python, C#, JavaScript or Java. Finished agents can be shared into Microsoft Teams and Microsoft Copilot.

Foundry does not make an AI system correct by itself. Its guardrails rely on classifier models, which are models that sort text and flag likely harm. A flag is a best guess, not a promise. Some agent checks are also still in preview, which means not finished yet. It also needs an Azure subscription, because it is a place to build AI apps rather than a chat app you open to ask questions.