A fully managed platform in Microsoft Foundry for hosting, scaling, and securing AI agents built with any supported framework or model
Hi Wade, did you manage to clarify this? If not, the key detail in your second screenshot is “Agent Application endpoints.”
When you reuse a conversation ID in the lab, Foundry stores the conversation history and uses it for subsequent turns. Your application retains the ID to identify that conversation; it doesn’t need to store and resend the full transcript. Passing an ID on each request doesn’t make the history client-managed.
The statement in your screenshot describes the legacy Agent Application endpoint. Microsoft documents that endpoint as supporting only stateless POST /responses, with /conversations unavailable. For that endpoint, the client needs to maintain the history for multi-turn interactions.
That explains the apparent contradiction: the lab uses service-managed conversation history, while the screenshot describes a limitation of a particular published endpoint.
These two Microsoft references explain the distinction:
- Agents, conversations, and responses - how conversations persist history across turns.
- Agent Applications: Responses protocol - the stateless limitation described in your screenshot.
The second article is now marked as the legacy publishing experience, so that restriction should be read in that specific context.