The Foundry Mental Model
Level Up Team
This post summarizes The Foundry Mental Model, module 5.1 of Level Up’s Azure AI Fundamentals series. The lesson is on YouTube as Azure AI Fundamentals AI-901 Exam: 5.1: Foundry Mental Model. Module 4.3 is Responsible AI Scenario Drill.
Foundry is the application platform
Microsoft Foundry is one environment for building, evaluating, deploying, and operating AI applications and agents. It brings together models, agents, tools and knowledge, evaluations and observability, and Azure security and governance. The path is the business application, then Microsoft Foundry, then the Azure platform and infrastructure.
A Foundry resource is the governance boundary
A Foundry resource is the top-level Azure resource for the shared AI environment. It controls security and networking, model deployments, connections, access control, and shared governance. In the Azure hierarchy, the tenant and subscription are the organization and billing boundary, a resource group groups resources for lifecycle, and the Foundry resource plus its projects separate environment governance from solution work. One resource can contain multiple projects. A Foundry resource governs the environment. A project organizes development work.
A project isolates one solution
A project is the working boundary for a specific AI solution. It organizes agents, files and assets, evaluations, development activity, and project-level access. One Foundry resource might hold a customer-support assistant, a document-processing workflow, and an internal employee assistant. Shared controls stay on the resource. Each project isolates one solution.
A model is not yet a deployment
Selecting a model from the Foundry Models catalog does not make it available to an application. A deployment is a callable instance of that model: an endpoint, capacity and quota, regional configuration, content-safety controls, and usage-based or provisioned billing.
Model selection answers which capability you need. Deployment answers how the application will access that capability. The path is catalog, then deployment endpoint, then application.
Tools act. Agents decide. Models generate.
Tools give an application access to external capabilities, data, or actions: search and retrieval, functions and APIs, code execution, and connected business systems. An agent combines a model with instructions, tools, and knowledge to complete a task.
In the support assistant, the model interprets the customer’s request, a retrieval tool finds relevant support information, and the agent decides which information or action the request requires. A model generates or interprets content. A tool retrieves or acts. An agent decides how to use both.
The portal is a stage, not the finished application
Portal experiments inform the application. The playground is not the customer solution.
| Foundry portal | Application development |
|---|---|
| Explore available models | Build repeatable application logic |
| Test prompts in playgrounds | Call deployments through SDKs or REST APIs |
| Compare model behavior | Integrate identity, data, and tools |
| Configure prompt agents | Add testing, monitoring, and deployment workflows |
| Run quick evaluations | Operate the solution at production scale |
Experiment with prompts and compare models in a playground in the Microsoft Foundry portal before writing application code.
Trace the solution in one sequence
- Azure provides the cloud foundation.
- A Foundry resource provides shared governance.
- A project organizes one AI solution.
- A model deployment provides an accessible endpoint.
- Tools connect the solution to information and actions.
- An agent coordinates the model and tools.
- The application delivers the user experience.
You made it this far
If these lessons have been useful, you already know how we teach AI-901. Module 2 is the natural place to keep going with your own team: the workload families, and how to choose the right one when a scenario is trying to distract you.
Level Up runs that material as a workshop or as custom training, alongside Module 1, delivered by Azure-certified experts.