Foundry agent types: prompt vs hosted vs connected
Verdict: No code or compute means a prompt agent. A managed, callable endpoint for packaged code means a hosted agent. Several instances of one knowledge tool means connected agents behind an orchestrator.
| Criterion | Prompt agent | Hosted agent | Connected agent |
|---|---|---|---|
| How it is built | Configuration only, in the portal or SDK; no code or compute | Package Agent Framework code as a container or source zip | A registered agent orchestrated by another agent |
| Runtime | Foundry fully manages the runtime | Foundry runs the code on autoscaling Micro VMs with a dedicated Entra identity and a callable endpoint | Runs as a specialist called by an orchestrator agent |
| Choose when | No custom orchestration logic is needed | Other apps must call a network-addressable managed endpoint | One agent needs several instances of a knowledge tool, e.g. multiple Azure AI Search indexes |
Rules
- Prompt agents need no application code or compute because Foundry fully manages the runtime.
- Hosted agents run packaged code on autoscaling Micro VMs with a dedicated Entra identity and a managed endpoint other apps can call.
- Use connected agents when a solution needs several Azure AI Search indexes, since only one instance of each knowledge tool type is allowed per agent.
- Coordinate tasks from a hosted agent with Microsoft Agent Framework workflows or another code framework, not the visual designer.
- Preserve multi-turn context by creating a conversation once and passing its id on each responses.create call.
Traps
- Attaching the same Azure AI Search tool multiple times at run level is unsupported; only one instance of each knowledge tool type is allowed per run.
- Hosted agents are not supported in the visual workflow designer.