Automatically stopping calls to a failing dependency for a while, instead of hammering it with retries. It protects the failing service from being kept down, and it protects your system from spending its whole capacity waiting for timeouts.
Technical view
Trips from closed to open when the failure rate over a rolling window exceeds a threshold, rejecting calls immediately, then moves to half-open after a cooldown to test recovery with limited traffic. Combined with bounded retries and backoff with jitter to avoid synchronised retry storms. Particularly relevant with model providers, where timeouts are long and retry amplification can consume the entire rate limit.
Related
Part of AI Terminology Index, section: Running It in Production.