How do you decide when an AI agent should retry instead of stopping?

Retries seem simple at first, but deciding when an AI agent should retry can be surprisingly difficult.

Too many retries waste resources and can create loops, while too few retries may cause agents to give up on tasks they could have completed successfully.

I'm curious how different teams approach retry strategies in production.

Do you limit retries by task type, confidence score, error category, or something else?

I'd love to hear how you've found the right balance.