Start with your goal
Find the row closest to what you’re trying to do.| I want to | Start here | Why |
|---|---|---|
| Send a prompt to a model, with no agent or tools | Build with models and agents | The shortest path to a working call. Add an agent later only if you need tools or orchestration. |
| Build my first agent | Choose how to build | Decides prompt agent versus hosted agent before you invest in either. That choice shapes everything after it. |
| Host an agent I already wrote | Deploy a hosted agent | Runs your existing code or container with a managed endpoint, scaling, and identity. |
| Use an open-source framework with Foundry | Frameworks for hosted agents | Shows how LangGraph, Microsoft Agent Framework, and Semantic Kernel map onto Foundry hosting. |
| Evaluate agent quality | Evaluate generative AI apps | Establishes a baseline score first, so later changes are measurable rather than anecdotal. |
| Add observability | Foundry Observability | Tracing, monitoring, and evaluation share one data model, so set them up together. |
| Deploy and operate an application | Set up CI/CD for agents | Gets a repeatable release path before you scale. For fleet-wide governance, see the control plane. |
Major products and capabilities
Each row names one decision-level entry point. The When to use it column states the fit first, then the boundary where another option is a better match.| Product or capability | What it is | When to use it | Who it’s for |
|---|---|---|---|
| Foundry portal | Web experience for models, agents, evaluations, and resources. | Use to explore and prototype without writing code. If you need repeatable builds, move to the SDKs or CLI. | Anyone evaluating Foundry |
| Prompt agents | Declarative agents defined by instructions, a model, and tools. | Use when you want an agent with no runtime to manage. If you need custom code or a container, use hosted agents instead. | Agent developer |
| Hosted agents | Your own agent code or framework, run by Foundry. | Use when you need full control of agent logic and dependencies. If instructions and tools are enough, prompt agents are simpler. | Agent developer |
| Foundry Models | Catalog of models from Microsoft, OpenAI, Anthropic, Meta, and others. | Use to choose and deploy the model behind your app or agent. If a model needs your domain data, consider fine-tuning. | Application developer |
| Fine-tuning | Customization of a model on your own data. | Use when prompting and retrieval can’t reach the quality you need. If the gap is missing context, use retrieval instead. | ML engineer |
| Toolbox | Managed endpoint that packages the tools an agent can call. | Use to give an agent actions and to govern tools centrally. If the agent only needs your own data, start with Foundry IQ. | Agent developer |
| Foundry IQ | Agentic retrieval over a knowledge base. | Use to ground answers in your content. If you only need a few uploaded files, file search is lighter weight. | Agent developer |
| Memory (preview) | Persistent agent recall across sessions. | Use when an agent must remember prior turns or user context. If each request stands alone, skip it. | Agent developer |
| Workflows | Orchestration of multiple agents and steps. | Use when one task spans several agents or stages. If a single agent with tools can finish the job, stay simpler. | Agent developer |
| Microsoft Foundry SDKs | Client libraries for Python, C#, JavaScript, and Java. | Use to build Foundry into an application. If you’re scaffolding and deploying agent projects, add the Azure Developer CLI. | Application developer |
| Azure Developer CLI (azd) | Command-line scaffolding, deployment, and environment management. | Use to create, run, and ship agent projects repeatably. If you only call models from an app, the SDKs are enough. | Platform engineer |
| Visual Studio Code extension | Foundry projects, models, and agents inside the editor. | Use to build and debug without leaving VS Code. If you want an AI assistant to drive Foundry, use coding agents. | Application developer |
| Coding agents and MCP | Model Context Protocol access to Foundry for coding agents. | Use to let GitHub Copilot or Claude Code work against Foundry. If you prefer direct authoring, use the VS Code extension. | Application developer |
| LangChain and LangGraph | Integration of Foundry models, tools, memory, and tracing. | Use when your team already builds on LangChain or LangGraph. If you’re starting fresh, native agents need less glue code. | Agent developer |
| Foundry Observability | Tracing, monitoring, and dashboards for agents and models. | Use to see what agents actually did in production. If you’re comparing versions before release, start with evaluations. | AI quality engineer |
| Evaluations | Scoring for generative AI apps and agents. | Use to measure quality and catch regressions before release. If you need adversarial coverage, add AI red teaming. | AI quality engineer |
| AI red teaming | Automated adversarial scans against a model or agent. | Use to probe for unsafe or exploitable behavior. If you’re measuring everyday quality, evaluations fit better. | Security administrator |
| Guardrails and controls | Content filters and safety controls for models and agents. | Use to enforce safety at run time. If you’re testing safety before release, pair it with AI red teaming. | Security administrator |
| Foundry control plane | Governance for agents, models, and tools across your estate. | Use to set limits, policy, and oversight across many projects. If you’re governing a single project, project settings are enough. | Platform engineer |