What causes a retry storm in an agent?

A retry storm happens when the agent’s retry policy treats a permanent error as if it were transient, so it keeps retrying a call that can never succeed. The signature is a tool call repeated with near-identical arguments, each attempt failing the same way: an unknown tool name, an argument that fails schema validation, a 400 rather than a 429 or a timeout. Retry counts above one on that kind of non-transient error are the tell.

The most common trigger is a hallucinated tool name. In a 200-task ReAct benchmark measured by Towards Data Science, 466 of 513 retries, 90.8 percent of the retry budget, hit a tool the model had invented. A retry policy built for flaky networks and rate limits will run that call forever, because from its point of view every attempt looks like an ordinary failure worth trying again.