Why Claude Code's harness matters more than the model in large codebases

Anthropic argues that the success of Claude Code in monorepos, legacy stacks, and multi-repo architectures hinges less on raw model benchmarks than on the surrounding harness. Instead of relying on RAG-style embedding indexes that go stale as engineers commit changes, Claude Code performs agentic search directly against the live filesystem, traversing files and grepping references like a human engineer. The tradeoff is that it needs decent starting context, which is why codebase setup determines how well navigation actually works at scale.

The recommended harness rests on five extension points layered in order: CLAUDE.md files supply persistent context, hooks enforce deterministic checks and feed continuous improvement back into the setup, skills provide on-demand expertise via progressive disclosure (optionally scoped to specific paths), plugins package skills and configs so good setups stop being tribal, and MCP servers connect Claude to internal tools and structured search. LSP integrations add symbol-level navigation that is especially valuable for C, C++, and other multi-language codebases where text pattern-matching alone misidentifies symbols.

The piece also highlights subagents as a way to separate exploration from editing — a read-only instance maps a subsystem and writes findings to disk, letting the main agent edit with full context without blowing its own window. The throughline: teams that invest in CLAUDE.md hygiene, hooks, scoped skills, LSP, and shared plugin distribution get markedly better results than those expecting model upgrades alone to carry them.