RAG settled the fine-tuning debate. Agentic RAG is the new argument.
A 2026 practitioner guide claims 80% of fine-tuning asks are really retrieval problems. ACM publishes a consolidating RAG survey. And 'agentic RAG' is now a vendor category. Here's what that means.
Three data points from the past week point to the same conclusion: RAG has settled into the default pattern for production LLM systems, and the interesting questions have moved up the stack to agentic orchestration. The "should we fine-tune?" debate that dominated 2024-2025 is effectively over — retrieval won, and the new argument is about how much agency to give the retrieval loop.
The clearest practitioner signal comes from hjLabs's 2026 fine-tuning guide, which puts the number at 80% — that's the share of "we need fine-tuning" requests they say are actually solved by better retrieval and prompt engineering. The guide is opinionated and worth reading: QLoRA as the fine-tuning default when you do need it, but RAG first, always. On the academic side, ACM Computing Surveys published a consolidating survey of retrieval-augmented text generation on May 15, 2026 — the kind of literature review that appears when a field has matured enough to taxonomize its sub-patterns. And on the vendor side, Progress Software announced that its Agentic RAG product won a 2026 AI Excellence Award. We'd file the award itself under industry-marketing rather than independent quality bar, but it does mark how thoroughly "agentic RAG" has crystallized as a named product category.
For builders shipping AI features, the practical move is procedural: when a stakeholder asks for fine-tuning, audit the retrieval layer first. Chunking strategy, embedding model choice, hybrid search, and reranking will move the needle further than a custom-trained model in most enterprise contexts — and they're cheaper to iterate on. For buyers evaluating vendors, the proliferation of "agentic RAG" products means asking sharper questions: what does agentic mean here — query planning, multi-step retrieval, tool use, self-correction, all of the above? The label is doing a lot of work right now, and definitions vary by vendor.
The interesting frontier isn't whether to use RAG. It's where agency lives in the retrieval pipeline — when the model should iterate on its own queries, when it should ask for more context, and when it should answer with what it has.
The bottom line: Default to retrieval, reserve fine-tuning for the 20% of cases where it actually moves the needle, and treat "agentic RAG" vendor labels as a starting point for due diligence rather than the answer.