A conversation and a workflow recover differently
A durable agent keeps conversation state outside the worker process. If a customer starts a support chat and returns later with an order number, a compatible worker can pick up the same conversation using its durable session ID. Durable Task stores the history, but the application still has to retain the session reference. Serializing that reference is not the same as saving it somewhere the next request can find it.
A durable workflow remembers which steps have finished. It can coordinate agent calls, application activities, and external events, whether developers write the orchestration directly or register an Agent Framework workflow graph. After an interruption, Durable Task replays the execution history and reuses recorded results. That means a completed model call whose result was saved does not have to run again just because a worker restarted.
The two capabilities fit together in an approval process. An agent can start a workflow and return an instance ID instead of holding the request open. The workflow can prepare a draft, notify a reviewer, and then wait without keeping a worker active. When the decision arrives, the application sends it to that workflow instance so processing can resume.
What teams still have to own
Azure Functions provides generated HTTP endpoints and managed scaling. Self-hosting gives teams control over where the worker runs, but leaves them responsible for the API, worker lifecycle, networking, authentication, and deployment. Microsoft recommends Durable Task Scheduler as the backend. The hosting choice also affects Python requirements: Microsoft Learn specifies Python 3.13 or later for Azure Functions and Python 3.10 or later for self-hosting.
The most important production caveat is that a checkpoint does not guarantee an external action happens only once. Microsoft Learn warns that a worker can fail after writing to another system but before recording the result. On recovery, the activity may run again. For a workflow that creates support tickets, a useful staging test is to interrupt the worker at that point and verify that retrying the same operation does not create a second ticket.
Teams upgrading an existing deployment have a separate check: whether the new workers can replay histories written by the old ones. Microsoft Learn warns that migrating agent session data does not make incompatible workflow histories replayable. If affected runs still need to finish, keep them on their original workers and task hub. For a simple agent that answers each request independently, the standard framework may be enough; the extension earns its keep when losing a conversation or repeating unfinished work has a real cost.
References
- (Launched) Generally Available: Durable Task extension for Microsoft Agent Framework Azure Updates · 2026-11-17T20:00:32Z
- Durable Task extension for Microsoft Agent Framework is now Generally Available | Microsoft Community Hub techcommunity.microsoft.com
- Durable Task Extension for Microsoft Agent Framework | Microsoft Learn learn.microsoft.com