Runtime Failure Recovery in LLM-Based Agent Systems: A Review of Detection, Attribution and Repair

Agents built on large language models (LLMs) now call tools, edit files and act on external services, yet their evaluation still centres on whether a task succeeded. Much published work classifies why agents fail; far less addresses what a running system should do once a failure occurs, and that work is spread across observability, debugging and orchestration research. This narrative review follows a failure through four stages (detection, attribution, repair and resumption) and organises the evidence in the terminology of dependable computing. Its main argument is that an agent failure is better classified by whether its effects can be undone than by its cause. A failure that left no lasting effect can be retried, one confined to state the system controls can be undone by restoring that state, and one whose effects have reached external systems can only be offset by a compensating action. Published failure taxonomies converge on design, interaction and verification faults, but none records where a failure's effects land. Detection methods now cover silent and slow failures, though their costs are reported in incomparable units. Attribution has improved, yet on the main benchmark the best methods locate the decisive step in fewer than half of hand-crafted cases. Checkpointing systems can restore agent and sandbox state, but each stops at a boundary beyond which nothing is restored. Mechanisms for irreversible effects, adapted from sagas and compensating transactions, have appeared only since 2025, mostly as preprints. Studies of recovery use different fault models and metrics, and none follows deployed agents over time. Four research directions are proposed, of which the handling of irreversible effects is the most pressing.