Microsoft Foundry - Part 1
- Nathan
- Aug 14, 2025
- 5 min read
Updated: Mar 21
Microsoft Foundry has everything you need to design, customize, and manage AI applications and agents.
Microsoft Foundry has gone through quite a lot of changes and transformations over a short period of time. Let's go over a quick timeline:
Nov. 2023: Azure AI Studio is released in Public Preview (link)
Nov. 2024: Azure AI Studio is re-branded as Azure AI Foundry (link)
May 2025: Azure AI Foundry is re-launched with new features under a new Azure resource type (link)
It also incorporates / is now the new name for Azure AI Services (formerly known as Azure Cognitive Services)
Nov. 2025: Azure AI Foundry is re-branded as Microsoft Foundry
The "new" Foundry Portal experience is launched
Agent Service v2 is launched with new features
Azure AI Services is re-branded as Foundry Prebuilt Tools
Here is Microsoft's info-graphic on the new Microsoft Foundry:
Part 1 of this series will focus on some of the initial topics that I found confusing to learn and/or understand while I was working with Microsoft Foundry:
Foundry has 2 different resource types. For now, both types co-exist (the new MS Foundry and the old AI Foundry Hub). I will explain the current state of affairs and which features are available, or not available, on each type.
I'll cover what I found confusing about:
Foundry Models
Foundry Agent Service
Foundry Prebuilt Tools
Microsoft Foundry Resource Types - Old vs. New
The new Microsoft Foundry resource type is based on the Azure Resource Provider for "Microsoft.CognitiveServices/accounts". You create a Microsoft Foundry resource, and from there you create 1 or more "Foundry Projects". Going forward, new investments and capabilities for Agents and Models will be focused on the new Foundry Projects.
The old AI Foundry Hub resource type is based on the Azure Resource Provider for "Microsoft.MachineLearningServices/workspaces". With this old type you create an AI Foundry "Hub", and from there you create 1 or more "Hub-based Projects". This old type still exists for now. However, it is clear that Microsoft will not be focusing much of their time or energy on this old resource type. These old Projects still remain accessible for any Generative AI capabilities that are not yet supported by the new resource type.
Some quick, rapid-fire observations:
Agents v1 (aka Classic Agents) are now deprecated. They will be fully retired on March 31, 2027.
Agents v2 are only available in the new MS Foundry. Microsoft made a big announcement saying that Agent Service v2 is now GA, however that's a little disingenuous. Only 1 of the 3 agents types in v2 is GA, the other 2 are still in preview at the time of this writing.
Confusingly, if you use the Azure Portal to deploy the old AI Foundry Hub resource type, then it will also create the new MS Foundry resource type as well, and they will live side by side. The old resource type can connect to, and use features from, the new resource type (more on that below).
Prompt Flows are only available in the old AI Foundry Hub.
The "new" Foundry Portal experience is only available for new MS Foundry projects.
Foundry Models
Microsoft likes to say that MS Foundry supports over 11,000 AI models. However, that is a little misleading ... let me explain. A model can be deployed to Foundry in multiple ways, and the deployment methods vary between AI Foundry Hub and MS Foundry. The vast majority of those 11,000+ models are only available via 1 deployment method: managed compute. Under the hood, these types of models are running on the compute infrastructure provided by the Azure Machine Learning platform. And, as we learned earlier, the old AI Foundry Hub is built on top of that platform (Azure Machine Learning). So, to summarize, the models that require managed compute can only be deployed to the old AI Foundry Hub resource type.
Below, I compare the various deployment methods that are available:
Note: If you want, you can connect an old AI Foundry Hub to a new MS Foundry resource. By doing so, this will let you do Standard deployments on your old AI Foundry Hub resource. But, in reality this is deploying the model to the new MS Foundry, and then just making a connection to it inside the old AI Foundry Hub.
Here is another info-graphic from Microsoft which shows the various Model deployment methods. This graphic focuses specifically on the new MS Foundry.
Some models give you the option of using Provisioned Throughput Units (PTU). PTUs are just generic units of processing capacity. They help by giving you predictable performance, and by providing stable max latency and throughput for uniform workloads. The allocated throughput is available to you whether you use it or not. You can even purchase Azure Reservations for your PTUs to save some cost.
Deployments have 3 different options for data processing, depending on your chosen model:
Global: this will dynamically route traffic to any region across the globe that has the best availability for each request. Global provides the highest default quota and eliminates the need to load balance across multiple resources.
Data-Zone: this will dynamically route traffic to any region within your data-zone that has the best availability for each request. For example, if your Foundry resource is deployed to the US, then traffic could be dynamically routed to any region within the US.
Regional: traffic will stay within the region of your Foundry. Model availability in each region as well as throughput may be limited.
Foundry Agent Service
In my opinion, the Agent Service is a very big and important feature of Microsoft Foundry. This service has not been immune to Microsoft's rapid advancement and updates and has undergone a lot of changes.
Agent Service v1 (aka Classic Agents)
This is now deprecated, with a retirement date of March 31, 2027.
This is the first version of the Agent Service.
The endpoint for v1 follows this pattern https://<accountName>.services.ai.azure.com/api/projects/<projectName>/assistants
In the old AI Foundry Hub, this was only released in Preview form.
In the new MS Foundry, this was released as Generally Available (GA).
Agent Service v2
This is the latest version of the Agent Service announced at Microsoft Ignite 25. This supports the newest features like Hosted Agents and Multi-Agent Workflows.
The endpoint for v2 follows this pattern https://<accountName>.services.ai.azure.com/api/projects/<projectName>/agents
Agents v2 can only be used in the new MS Foundry, they are not available in the old AI Foundry Hub
Current feature status:
Prompt Agents (aka Declarative Agents) are GA.
Hosted Agents are Preview.
Workflows are Preview.
Foundry Prebuilt Tools
Foundry Prebuilt Tools have now been absorbed by the new Microsoft Foundry resource type. Foundry Prebuilt Tools were previously known as Azure AI Services. Prior to that, they were previously known as Azure Cognitive Services.
When you create a new MS Foundry resource you get an endpoint for your Foundry in the form of https://<foundryName>.cognitiveservices.azure.com. This single endpoint can be used to access many different Foundry Prebuilt Tools:
Azure AI Speech
Azure AI Vision
Azure AI Document Intelligence (formerly known as Azure AI Form Recognizer)
Azure AI Language
Azure AI Content Safety
Azure AI Content Understanding
Document Translation
These services are not built into the older AI Foundry Hub resource type. However, you can take an old AI Foundry Hub and connect it to a new Microsoft Foundry resource. By doing so, this will give you access to all of the Foundry Prebuilt Tools in your old AI Foundry Hub.
Conclusion
That covers everything that I was confused about (so far) while I'm going through my journey of learning and using Microsoft Foundry.
Future parts of this series will cover more technical details of Microsoft Foundry's many features and capabilities.