Towards Data Science is a community-powered publication that showcases work in data science, machine learning and artificial intelligence. Every day newcomers, seasoned researchers and industry practitioners publish tutorials, research notes and real-world case studies that help the field move forward.

Towards Data Science

Agentic RAG systems fail in production through three compounding failure modes: retrieval thrash (the agent loops without converging), tool storms (excessive cascading tool calls that drain budgets), and context bloat (low-signal content filling the context window until instruction-following degrades). Each failure is traced to missing hard caps, weak stopping rules, and lack of observability. Concrete detection signals are provided — tool calls per task, retrieval iterations, context growth rate, p95 latency, and cost per successful task — along with specific mitigations: cap retrieval at 3 iterations, set per-tool budgets, summarise tool outputs before injecting into context, and use intent-based routing. The post also includes a decision framework for when to use agentic RAG versus classic single-pass RAG.

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

What Agentic RAG Is (and What Makes It Fragile)

Failure Mode Taxonomy: What Breaks and Why