AI Apps
Microsoft Foundry
The platform an AI application or agent is built on, with the model catalogue, evaluation, and tracing in one place rather than assembled per project.
Product
Microsoft FoundryMicrosoft Foundry is the platform AI applications and agents are built, deployed, and observed on. It carries a catalogue of models from several providers so a workload can be pointed at the model that suits it rather than the one it started with, an agent service for tool calling and orchestration, a prompt and dataset workspace, and an evaluation harness that scores groundedness, relevance, and safety against a fixed question set. Tracing records what was retrieved, what was passed to the model, and what came back, so a wrong answer can be diagnosed rather than argued about. Content safety filters sit in front of and behind the model with thresholds set per deployment.
How it works
- A project is created and the models it may call are chosen from the catalogue, with the deployment, region, and quota agreed before the design assumes them.
- Prompts, datasets, and the retrieval configuration are held in the project rather than in application code, so a change is a change to the workload and not to a release.
- The agent service registers the tools an agent may call and handles the planning and the calling, so the application surface stays thin.
- An evaluation set of real questions is run against the workload, scoring groundedness, relevance, and safety, and the same set is rerun after each model or prompt change.
- Tracing captures the retrieval, the grounded prompt, and the response for each request, so quality problems are inspected rather than reproduced by guesswork.
- Content safety filters are configured per deployment, with the thresholds and the blocked categories recorded rather than left at whatever the default was.
- The workload is deployed behind the customer's own identity and network controls, and the operating team is handed the evaluation set along with it.
A problem it solves
An assistant demonstrates well in a workshop and then nobody can say whether it is getting better or worse, because there is no question set it is scored against and no record of what it retrieved. Microsoft Foundry puts the models, the prompts, the evaluation set, and the traces in one project, so a change to the model or the prompt is measured against the same questions each time instead of being judged by whoever tried it last.