Agentic RAG Failure Modes: Retrieval Thrash, Tool Storms, and Context Bloat (and How to Spot Them Early)

Agentic Retrieval-Augmented Generation (RAG) systems fail differently than classic RAG due to their control loop architecture. Three common failure modes are retrieval thrash, where the agent searches endlessly without converging; tool storms, involving excessive and cascading tool calls; and context bloat, where the context window is filled with low-signal information. These issues arise from weak stopping criteria and a lack of observability, leading to spiraling costs and degraded performance. Mitigations include setting hard caps on iterations and tool calls, compressing tool outputs, and tracking key signals like latency and cost per task to detect failures early.