Sentry: Learning to Recover from LLM Agent Failures at Test Time

2026-10-02

AI Summary

This study addresses task failures in LLM agents caused by invalid tool invocations or reasoning biases, as well as the interference introduced by existing methods that globally inject failure knowledge. To mitigate these issues, this work proposes the Sentry layer, which formulates failure knowledge as conditional knowledge. Specifically, it employs test-time fault detection to trigger conditionalized RAG retrieval from an external manual, exposing lessons on demand to guide recovery. Following reward-free verification, the knowledge base is dynamically updated to enable continuous learning at runtime. Experiments demonstrate that Sentry outperforms the strongest baseline by 37% on average across multiple benchmarks. Furthermore, the acquired lessons transfer effectively across tasks, substantially enhancing agent reliability.

Abstract

LLM agents often fail mid-task due to invalid tool calls, repeated actions, or poorly grounded reasoning, and learning from these failures is a path to reliability. We find that how failure knowledge reaches the agent matters as much as what it contains. Failure lessons are conditional: kept in the agent's context, they misfire when their failure is absent, and removing them from an evolving playbook improves performance. Runtime interventions, in contrast, act only when a failure occurs but do not learn from their repairs. We argue that failure knowledge is conditional knowledge and should be conditionally exposed, and instantiate this principle in Sentry, a failure-management layer that runs alongside the agent. When Sentry detects a failure, it retrieves matching lessons from an external playbook to guide recovery, verifies without access to task rewards whether the agent recovered, and stores a new lesson only if it did; the full playbook never enters the agent's context. Across multiple agentic benchmarks, Sentry outperforms the strongest runtime-intervention baseline on every benchmark, by 37\% on average, and the strongest context-evolution baseline by 39\% on the two benchmarks where both are evaluated; combining Sentry with context evolution yields further gains. Learned lessons transfer to held-out tasks, and controlled experiments show that exposing the full playbook to the agent lowers performance even when relevant lessons remain available on demand.

Problem

• LLM agents
• failure recovery
• conditional knowledge
• tool call errors
• test-time learning

Innovation

• failure recovery
• test-time learning
• conditional knowledge
• LLM agents
• external playbook