Retry storms begin when callers respond to errors by trying again too quickly or without a shared limit. The added traffic delays recovery, creates more timeouts and can spread failure to other services or agent tasks.
Prevention combines exponential backoff, jitter, circuit breakers, concurrency limits and a maximum retry budget. Agent systems should also distinguish uncertain completion from confirmed failure so retries do not duplicate side effects.
ELI5
A retry storm happens when many programs respond to a slow or failing service by sending repeated requests too quickly. The extra traffic makes the service even less able to recover, causing more timeouts and still more retries.
For example, hundreds of agent tasks may all retry the same unavailable API every second. Increasing waits, adding random timing, limiting concurrent requests, stopping after a fixed budget, and using a circuit breaker can prevent the feedback loop.
How can an agent cause a retry storm?
It can repeatedly call a failing tool across many turns or parallel tasks without adequate delays, limits or circuit breaking.
How is a retry storm prevented?
Use bounded retries, exponential backoff with jitter, circuit breakers, concurrency limits and idempotent operations.
