GitHub user pltbkd added a comment to the discussion: [Feature] Sub-agent 
Resource for Flink Agents - Framework part

Thanks for the detailed thoughts. This design aims to keep the user-facing 
interface simple. I agree that these concerns belong to the framework rather 
than every individual subagent implementation. At the same time, I'd also like 
to clarify that most of the mechanisms you mentioned are already part of the 
current design, although they are exposed differently:
- Stable Invocation ID → the deterministic SessionId.
- Invocation lifecycle → all subagent calls are expected to go through 
durableExecute.
- Recovery / reconciliation → handled by durableExecute, not by the subagent 
author.
- Completion semantics → durableExecute provides durable execution for internal 
subagents. Combined with framework-managed Action state, it enables strict 
consistency; without Action state, it naturally degrades to completion-only 
semantics, where intermediate execution state is not observable. External 
subagents are expected to provide the same completion semantics by using 
durableExecute.

Cancellation semantics are indeed missing today. I suppose it won't changes the 
overall architecture, but it is definitely a gap that should be addressed.

For external subagents, the intention is that they follow the same durability 
contract (for example, by wrapping external operations with durableExecute). 
What is currently missing is a shared framework for external subagents that can 
enforce these conventions automatically rather than relying on implementations 
to follow them. If you have concrete requirements or examples in this area, I'd 
be happy to discuss them further.

GitHub link: 
https://github.com/apache/flink-agents/discussions/909#discussioncomment-17694034

----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]

Reply via email to