AI Agents is one of the use cases where it is graspable why ๐๐๐ฟ๐ฎ๐ฏ๐น๐ฒ ๐๐๐ป๐ฐ๐๐ถ๐ผ๐ป๐ are a better abstraction than ๐๐ผ๐ฟ๐ธ๐ณ๐น๐ผ๐๐.
With Durable Functions, you wrap steps you want to be durable (e.g., inference calls, tool calls) and don't need to refactor the code into workflows and activities. The Agent stays a long running process (except for suspensions and failover), so caches or expensive transient state keeps working naturally.
If you still need to move steps to a separate physical process, make the step its own Durable Function that is called from the agent function - voilร . You show this to an app developer, and it will look familiar to them, because it is just service handlers (functions) that call/rpc each other.