just saw a post about how we keep blaming the llm when the real issue is just basic network reliability. imagine an agent triggers a $500 refund via stripe and everything looks fine on the model side, but then a random proxy timeout or an OOM error kills the connection before the response hits your logic. if you have a standard retry_decorator running, your system just executes the tool again. suddenly that customer has a $1,000 refund because the first attempt actually succeeded behind the scenes. it is not even an
intelligence issue since the model picked the right tool and parameters every time. we are basically building massive loops of double refunds w/o any idempotency keys.
the model isn't hallucinating, your infrastructure is just leaking money . has anyone actually figured out a scalable way to handle state recovery across distributed agent steps without manually tracking every single transaction id? it feels like we are all just one dropped packet away from bankruptcy
more here:
https://dev.to/arpanghoshal/failed-is-not-unknown-the-retry-bug-hiding-in-every-ai-agent-5721