Agentic RAG: Embedding Agents Inside the Retrieval Pipeline

A new ArXiv survey (2501.09136) and the Agent-UniRAG framework (2505.22571) document the shift from static RAG to agentic RAG, where autonomous agents dynamically manage retrieval strategies using reflection, planning, tool use, and multi-agent collaboration. The key insight from Phil Schmid (Hugging Face): most agent failures aren't model failures — they're context failures. Weak retrieval creates problems; weak prompts make systems unusable. LangChain formalizes four context strategies: write, select, compress, and isolate.

---

Sources: ArXiv Agentic RAG Survey · Agent-UniRAG · LangChain Docs