Agentic search beats RAG for live codebases

Claude Code’s search evolution tells a clear story: the early version used RAG with a local vector DB, and the current version uses agentic search (glob/grep/read). The agent won. For live codebases, the problems with RAG are compounding: embeddings go stale as code changes, re-indexing needs a continuous background job, code must be sent to an embedding service (privacy concern), and retrieval quality depends on chunking choices that may not match the query’s needs.

Agentic search sidesteps all of this — the LLM thinks “authentication logic probably lives in /src/auth or files named auth,” runs glob and grep to verify, then reads the specific files. No index, no staleness, no privacy leakage. This is Similarity is not relevance — relevance requires reasoning applied to code: the agent reasons about codebase structure rather than relying on embedding similarity. It connects to The context window is the fundamental constraint — everything else follows because the agent’s search strategy is fundamentally about managing what enters the context window. Mintlify’s ChromaFs proves the pattern extends beyond code: Virtual filesystems replace sandboxes for agent navigation — intercept commands instead of provisioning infrastructure intercepts Unix commands into database queries, giving documentation assistants the same navigational search — 460x faster than sandbox provisioning. The trade-off is more LLM calls per search — but for Boring tech wins for AI-native startups — simpler stack means faster AI-assisted shipping, the simplicity of “just use bash tools” beats maintaining a complex indexing pipeline.