Anthropic's applied AI team just dropped its best practices for deploying Claude Code across enterprise projects with millions of lines of code. The official take? Traditional RAG-based AI coding tools frequently cite stale code in large, active codebases because their indexes can't keep up. Claude Code's solution? Ditch RAG entirely and rely on agentic search over the local filesystem, paired with a layered extension framework to handle scale.
In massive C, C++, and Java projects, Claude Code navigates code like a human engineer — using grep and following references directly, without maintaining a centralized index. But this approach demands extremely high initial context quality. Anthropic calls the configuration system that determines model behavior the "extension framework." It has five layers: CLAUDE.md files provide directory-level coding conventions; Hooks trigger automated checks at specific nodes; Skills load specialized knowledge on demand (so security reviews only fire when needed, without polluting normal session context); Plugins bundle and distribute these configurations; and MCP servers connect to internal data sources.
On the deployment side, Anthropic heavily emphasizes the value of Language Server Protocol (LSP) integration. LSP gives the model symbol-level navigation precision, letting it distinguish between identically named functions across different files during searches — a key accuracy booster for multi-language, complex codebases. Teams must also proactively clean up outdated context rules as models evolve; rules written to compensate for older model limitations often end up hampering newer models.
The takeaway? Improving next-gen AI coding tools is no longer about maintaining giant vector databases of every line of code. Instead, it's about getting the local directory structure right, optimizing Language Server Protocol configurations, and building bespoke workflows.