Your Agent Is Only as Good as What It Retrieves

Most agentic RAG systems have a retrieval problem, not a model problem. When an agent's tool-use loop hands the model the wrong context, the agent doesn't fail loudly — it reasons confidently over the wrong documents and returns a wrong answer with nothing broken in the logs. This talk is for developers and ML engineers building agents who've shipped something that passed every test and then quietly gave someone a wrong answer in production. No retrieval evaluation experience needed. Through demos on a real OpenSearch index, we'll trace exactly what an agent's retrieval step actually returns, identify the failure modes that silently degrade what the model sees, and apply targeted fixes — without rearchitecting the agent itself. You'll leave knowing where to look when your agent is wrong, what to measure, and what to fix first.