Agentic RAG systems fail silently in production through three predictable patterns that can bankrupt cloud budgets before anyone notices. Unlike classic RAG's simple retrieve-once architecture, agentic RAG creates a control loop that repeatedly searches, evaluates, and decides whether to retrieve again. This loop's power for complex queries becomes its weakness when agents make bad decisions that compound across iterations.

Three failure modes dominate: retrieval thrash where agents keep searching without converging on answers, tool storms where excessive tool calls cascade until budgets vanish, and context bloat where the context window fills with low-signal content until models stop following instructions. Teams consistently see agents making 200+ LLM calls in 10 minutes, burning $50-$200 before detection. Research shows performance drops 20+ percentage points when critical information sits mid-context, meaning adding retrieved content can actively worsen answers.

Detection requires tracking specific signals from day one: tool calls per task, retrieval iterations per query, context length growth rate, and cost per successful task. Set hard caps like three retrieval iterations maximum, 10-15 tool calls per task, and context token ceilings. When tripwires fire, agents should stop cleanly and return best-effort answers with explicit uncertainty rather than spiraling into more retries. The fix isn't better prompts—it's budgeting, gating, and observability of the agent's decision loop.