Agentic RAG is retrieval-augmented generation in which the model drives the retrieval process itself: reformulating queries, running multiple searches across so
In agentic RAG the model drives retrieval itself: reformulating queries, running multiple searches across sources, judging whether the evidence is sufficient, verifying and citing sources, and self-correcting before answering, instead of a single one-shot retrieval. It became the mature enterprise pattern by 2026.
Cost and latency. Each reasoning-retrieval loop spends tokens, so complex questions can trigger several searches before an answer. Production systems bound the number of iterations and cache aggressively to keep response times and bills under control.