Same vector DB, same embedding model. Agent queries are more specific and less semantically aligned with the docs that hold the answer. Fix: hybrid BM25 + keyword boosting for entity terms (recall +23% on agent queries vs +6% on human queries).