Sentry: A Self-evolving Failure Manager for Agentic LLM Systems

Modern LLM agents increasingly operate over long interactive trajectories, where they often suffer from failures such as hallucinations, thinking loops, and invalid tool calls. Self-evolving context-adaptation methods can accumulate useful experience, but adding failure-specific advice to the task agent's context can turn local fixes into misleading general guidance. Runtime intervention methods avoid persistent context modification, but they typically rely on fixed intervention rules and do not learn from successful rescues. We propose \textbf{\textit{Sentry}}, a self-evolving external runtime failure manager for agentic LLM systems. Sentry detects local execution failures, repairs invalid actions or injects playbook-guided recovery advice, verifies whether the failure is resolved, and updates its external playbook with verified rescues. Across WebShop, AppWorld, SWE-bench Lite, and Mind2Web, Sentry improves task performance over runtime intervention and self-evolving context-adaptation baselines. Notably, the strategies and insights Sentry learns from earlier tasks transfer to later tasks, showing that Sentry learns reusable failure-recovery knowledge rather than task-specific fixes. An independent human audit further confirms that Sentry's intervention, recovery, and helpfulness judgments are generally aligned with human assessment. These results suggest that long-horizon agents benefit from a self-evolving failure-management layer that detects failures, guides recovery, and reuses successful repairs to improve reliability.