What is a Retry Loop?
A retry loop is an agent repeatedly retrying a failed operation in a tight cycle — the pattern behind the $12,400 night and several incident-database entries.
How it happens
An agent hits a failure (rate limit, timeout, error) and retries.
Each retry re-bills the operation: tokens, API call, or purchase.
The cost shape
Cost = single-operation cost × retries. 40 retries = 40× the base cost.
Loops compound across subagents and overnight runs.
The rule that stops it
A velocity limit caps transactions per window — the loop is BLOCKED at the cap, usually within seconds.
FAQ
Why don't agents stop retrying?
Retrying is what they're built to do; nothing tells them the attempt is hopeless or costly.
What's the fastest fix?
A velocity limit — it kills the loop at the source, not after the bill.
Related
Stop the next $12,400 night.
One API call (or MCP tool) in front of every agent transaction — APPROVED, BLOCKED, or FLAGGED, deterministic, ~5 ms, fully logged.
See plans — from $99/mo Try a live check