Claude researcher explains why agentic search beats RAG for code generation (officechai.com)

Anthropic engineer Boris Cherny explained why the team moved from Retrieval-Augmented Generation (RAG) to an “agentic search” approach for code-generation tasks: agentic search—where the model actively performs iterative search cycles with standard developer tools (e.g., glob, grep) against the live codebase—consistently outperformed their RAG setups. Beyond raw quality gains (Cherny said it beat other methods “by a lot”), Anthropic found agentic search avoids the operational and security headaches of maintaining pre-indexed RAG stores: indexes go stale, require synchronization, and create additional attack surface or third‑party exposure for sensitive code. Technically, the trade-offs are clear: agentic search uses more latency and tokens because the model performs repeated lookups, but it delivers fresher, more precise context and simpler auditability since it queries live sources rather than relying on replicated vectors. For the AI/ML community, this signals a shift toward tool-enabled agents that interact with live systems and APIs, especially for dynamic, sensitive domains like code. While long-context LLMs may reduce some retrieval needs, agentic approaches offer a compelling alternative to static RAG pipelines where timeliness, security, and developer experience matter most.