loading…
Expected:
{"action":"refund","amount":20}
Observed:
{"decision":"give money back","amount":"$20"}
The same request can also select a different tool. Retrying may repair a malformed sample, but it does not repair a weak contract or ambiguous policy.
Require structured output, validate the schema, validate tool arguments, constrain high-risk tools, and route uncertain or unsafe cases to a human. Track repair attempts and failures as product signals.
Run 50 cases five times in Kintsu Agent. Measure schema-valid rate, task success, tool-selection accuracy, and run-to-run variance. Add validation, bounded repair, fallback, and escalation; then compare the distributions.
Retry when sampling noise can plausibly produce a valid answer. Do not retry when the requested action is forbidden, the tool contract is wrong, or the necessary data is absent.