Not All Repair Helps: Which Interventions Recover a Failing Tool-Using Agent, and Which Cannot Reach It

This preprint studies when interventions can recover a failing tool-using language-model agent. In a deterministic tool-use sandbox the paper examines two failure mechanisms: skipped-lookup, where the agent fabricates an available but unfetched fact, and tool-false-validation where a non-error tool result falsely validates an invalid computation. Recovery is measured by counterfactual replay against a no-op control. The results show that skipped-lookup can be recovered by repairs that license the corrective action while tool-false-validation resists the text-level repairs tested. The paper also reports recovery decay with detection lateness and discusses cross-model limits caused by tool-calling reliability.