Three quarters of small-model database agent failures happened after the model had already called a tool.

Eleven days of production data from an open-source SQL client's agent mode, 39 locally-served open-weight models plus one hosted control, 8,199 runs and 110,711 ledger events. Of 2,100 model-attributed losses, 1,590 (75.7%) came from runs that had invoked at least one tool, holding in 99.7% of clustered resamples. Transport failures (tools used, no deliverable) are 36.2%, pure capability failures 17.3%. The operational lesson: production ledgers recorded refusal codes but never the model's arguments, hiding the cause for ten days, and capturing arguments exposed five server defects including one tool demanding a field its sibling forbade. Five server-side changes touching no model moved the numbers.