Do you cap retries at 2 or give it a third chance?

LLMs getting stuck in execution loops is such a huge headache when building DB tools.

We had a similar issue with API retries where the agent kept tweaking minor parameters forever. Capping it at two attempts and routing the failure directly to Slack is a solid circuit breaker.

In our setup, we usually allow two retries with an exponential backoff before killing the thread completely.

How do you handle the Slack notification does a developer jump in and manually correct it or just kill the process?