Python: Agent Framewrok Issues with composit workflow and variable state management. #4490
Description
Query 1:
As part of implementing the composite workflows (workflows with multiple orchestration patterns), we have tried two different options such as:
- Compose by wrapping orchestrations/sub workflows as agents.
- Compose by wrapping orchestrations/sub workflows in custom executors.
From our experience the 2nd options seems better and it natively support HITL, checkpointing, custom state management etc. First option uses intrenal/private methods with less code, but HITL is not efficient as the 2nd option and also need manual chaining of workflows.
can you please let us know which method will be the better choice for seamless implementation and working and also future proof? Also please let us know if there is any other better approach available.
Query 2:
If we use the workflow context state management option in our application, and we enable the horizontal scaling in our azure web application, does it cost any session or context unavailability issue (for the existing context) in the new auto scaled instances?
Query 3:
Based on the Agent framework documentation, it shows two different set of conditional flows/operation in both programmatic workflows and declarative workflows. We are bit confused why thse two scenarios the framework have two different set conditional operators. Can you please help us to understand?
