Single model call vs prompt agent vs hosted agent vs Toolbox vs Models catalog
Verdict: No tools, no orchestration: a single model call. Instructions, a model and tools set through the portal with no code to maintain: a prompt agent. Your own containerised orchestration code: a hosted agent. Comparing providers first: the Models catalog.
| Criterion | Single model call | Prompt agent | Hosted agent | Toolbox | Models catalog |
|---|---|---|---|---|---|
| What it is | Prompt-in, completion-out, no agent involved | Instructions, a model and tools specified in the portal or SDK | Your own code or container, e.g. LangGraph, packaged as a source package or container image | Curates tools, memory and retrieval for an existing agent | Browsable list of over ten thousand models from providers including Microsoft, OpenAI, Anthropic and Meta |
| Code or containers to maintain | None | None - Foundry hosts and runs it | Yes - the team's own runtime code is what Foundry containerises | None on its own; extends an agent | None; a catalog, not a runtime |
| Identity model | N/A | Has an Entra agent identity, but not one automatically dedicated per agent | Each hosted agent automatically receives its own dedicated Microsoft Entra identity | N/A | N/A |
| Scaling | N/A | Managed by Foundry | Scales by adding container instances per session and request volume | N/A | N/A |
| Choose when | A plain completion with no tools or retrieval | Prototyping quickly with no dedicated platform engineering | Existing custom orchestration code must run as a managed, autoscaling endpoint | An existing agent needs extra tools or retrieval attached | Scoping a generative AI workload and comparing candidate models first |
Rules
- Hosted agents are built for code-based frameworks such as LangGraph, shipped as a container image or source package, and run by Foundry with a managed endpoint, automatic scaling, and a dedicated identity.
- A prompt-in, completion-out call with no tools or orchestration is met with a single model call, not an agent.
- A prompt agent needs only instructions, a model and tools specified in the portal or SDK, with no application code or containers to maintain.
- More than ten thousand models from providers including Microsoft, OpenAI, Anthropic and Meta are browsable in the Foundry Models catalog.
- A hosted agent automatically receives its own dedicated Microsoft Entra identity per agent, whereas a prompt agent has an Entra agent identity that is not automatically dedicated, and Foundry autoscales the hosted agent by container instance per session and request volume.
Traps
- A hosted agent's own agent logic is runtime code the team must maintain - it is not code-free the way a prompt agent is, even though both are fully managed by Foundry.
- The Toolbox extends an agent with tools, memory and retrieval; it does not itself act as an agent type or list models from different providers.
- A prompt agent exists to attach tools and instructions - it is the wrong starting point when a scenario explicitly needs no tools or orchestration.