Failure Recovery in Tool-Using Large Language Model Agents: A Structured Review of Detection, Correction, and State Consistency

Abstract Large language model agents increasingly act on the world through tools such as code interpreters, application programming interfaces, search engines, perception modules and robotic planners. Every tool call is an opportunity for failure: the model may choose the wrong tool, write malformed input, misread a returned value, trust an erroneous output, or leave external systems in a half-changed condition after a partial sequence of actions. This paper is a structured, qualitative review of how the literature detects such failures, corrects them and keeps the state of the agent and its environment consistent while doing so. It adapts the feedback-loop model and the loud/silent distinction of Sun et al. (2024) to organise a taxonomy of failures, groups detection methods into execution signals, output verification, model self-assessment, external critics and human oversight, correction methods into retry, repair, reflection, replanning, rollback and escalation, and consistency mechanisms into idempotency, compensating actions, checkpointing, action logging, state verification and memory hygiene. A study-by-study comparison of nine representative works, including 2025–2026 studies, records their methods, benchmarks, findings, limitations and treatment of state consistency. Within the reviewed corpus, explicit error messages are comparatively well exploited, silent failures appear to remain a major open problem, self-correction without external feedback is reported to be unreliable on reasoning tasks, and state consistency receives less attention than detection and correction. The paper closes with recommendations for evaluation and design and an agenda for future research. The review is not exhaustive and reports no new experiments.