Retrying at Fifteen Dollars a Try

A subagent within a coding-agent framework attempted to use a file-editing tool, and the call failed due to invalid arguments. Rather than adjusting its approach or reporting the failure, the subagent immediately retried the identical failing call, which failed again, and again, with no limit on the number of attempts and no mechanism to detect that the same error was recurring. Users reported individual invocations running up more than $15 in model API costs before they noticed and manually stopped them, with one report of over $100 lost to the pattern in a single session.

A one-time tool glitch, considered and ruled out. A single failed call is unremarkable, but the system had no limit on how many times an identical failing call could be resubmitted, which is what turned an ordinary error into an expensive, unattended loop.

Apraxia, considered, since the initial edit call itself was malformed. Ruled out as the primary diagnosis because the notable failure wasn't the first bad call (plenty of tool calls fail) it was the complete absence of any adjustment or stopping condition across the many identical retries that followed.

Perseveration, most consistent with the evidence. Once the edit tool call failed, the subagent kept resubmitting the same unmodified request rather than trying a different approach or halting, continuing until a human intervened rather than any internal signal recognizing the pattern.

The financial version of this failure mode is easy to underrate until someone actually gets the bill, a loop that would be a minor annoyance in a chat interface becomes a real cost when each iteration is a billed API call running unattended. A hard cap on consecutive identical tool failures, paired with a circuit breaker that surfaces the problem to a human instead of retrying indefinitely, is a basic guardrail that costs little and prevents exactly this.

Cap consecutive identical tool failures and use a circuit breaker that surfaces the problem to a human.

Retry limits; identical-error detection; per-session spending caps.