Retry when agent fails

[kanlanc]

Beyond developers writing custom code for an agent to retry, is there currently a way to retry when agent fails both in getting the correct output or when the agent just fails?

[kavinkumar807]

@boyangsvl Is there any way I can help on this? I really want to contribute to ADK.

[kanlanc]

I'm imagining changing something deep in Runner or using something like Temporal or at least it's concepts or just simply a wrapper that retries if the agent fails

[boyangsvl]

@kavinkumar807 thanks, feel free to work on this. I think the retry should happen on the llm call level. we should not retry the entire agent or tool calls, because those may have side effects.

[kavinkumar807]

Hi @AlankritVerma01 saw your detailed comments explaining the timeout and retry implementation in the discussion #475. Since you have a laid out plan, thought we could work together by splitting this in equal parts to deliver it quickly. Our collaborative work by working on this feature in two parts, retry (myself) and timeout (you) could avoid potential conflicts that might arise when we work in silo. what are your thoughts?

cc: @boyangsvl @Jacksunwei

[AlankritVerma01]

Hey @kavinkumar807 ; thanks for the mention!

Yep, I’m currently working on the retry and timeout logic. I’ve mostly got it planned out (some of it’s already mentioned in #475, just waiting on Jack’s input there). I do think both features are pretty closely related, and there’s definitely potential for overlap/conflict. In my current draft, I’m using a single config object for both (as outlined in #475).

That said, this is shaping up to be a fairly large piece of work, and I might need some help along the way.

My suggestion: I can start implementing both retry and timeout in #475, that’ll involve updating the config structure, discussing default values, input flow, etc. Once the initial pieces are in place, it’d be awesome if you could jump in and build on it so we can tackle the rest together. (we can then follow up about that in 475 itself?)

Let me know what you think!
CC: @Jacksunwei (would also appreciate if you could guide as to which issue should this come under?) (i am also thinking of just creating a new one which clearly outlines both timeout of retry; or we can do them separately but then the order of implementation has to be sequential not parallel).

Thanksss

[kavinkumar807]

@AlankritVerma01

Thanks for the detailed update — understood!

cc: @Jacksunwei @boyangsvl

[kavinkumar807]

@AlankritVerma01 Shall I start working on this ??

Cc: @boyangsvl

[AlankritVerma01]

@kavinkumar807
Ah.. tough.. I am also waiting for Jack to review my suggestions.

So let's just wait until then?

Otherwise if we start on something and then realize that the team wants something completely different... that wouldnt be the best...

[kavinkumar807]

@AlankritVerma01 sure 👍